Merge branch 'main' into fix-bnb-test

enable 2 test cases on XPU (#11332 )
* enable 2 test cases on XPU Signed-off-by: YAO Matrix <matrix.yao@intel.com> * Apply style fixes --------- Signed-off-by: YAO Matrix <matrix.yao@intel.com> Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>
2025-04-18 08:55:15 +05:30 · 2025-04-17 13:27:41 -10:00 · 2025-04-17 10:05:01 -10:00 · 2025-04-17 18:00:25 +02:00 · 2025-04-17 19:51:49 +05:30 · 2025-04-17 01:17:44 -10:00
554 changed files with 55883 additions and 8715 deletions
@@ -38,6 +38,7 @@ jobs:
          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
          python -m uv pip install -e [quality,test]
          python -m uv pip install pandas peft
+          python -m uv pip uninstall transformers && python -m uv pip install transformers==4.48.0
      - name: Environment
        run: |
          python utils/print_env.py
@@ -414,10 +414,16 @@ jobs:
        config:
          - backend: "bitsandbytes"
            test_location: "bnb"
+            additional_deps: ["peft"]
          - backend: "gguf"
            test_location: "gguf"
+            additional_deps: ["peft"]
          - backend: "torchao"
            test_location: "torchao"
+            additional_deps: []
+          - backend: "optimum_quanto"
+            test_location: "quanto"
+            additional_deps: []
    runs-on:
      group: aws-g6e-xlarge-plus
    container:
@@ -435,6 +441,9 @@ jobs:
          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
          python -m uv pip install -e [quality,test]
          python -m uv pip install -U ${{ matrix.config.backend }}
+          if [ "${{ join(matrix.config.additional_deps, ' ') }}" != "" ]; then
+              python -m uv pip install ${{ join(matrix.config.additional_deps, ' ') }}
+          fi
          python -m uv pip install pytest-reportlog
      - name: Environment
        run: |
@@ -9,119 +9,9 @@ permissions:
  pull-requests: write

 jobs:
-  run-style-bot:
-    if: >
-      contains(github.event.comment.body, '@bot /style') &&
-      github.event.issue.pull_request != null
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Extract PR details
-        id: pr_info
-        uses: actions/github-script@v6
-        with:
-          script: |
-            const prNumber = context.payload.issue.number;
-            const { data: pr } = await github.rest.pulls.get({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              pull_number: prNumber
-            });
-            
-            // We capture both the branch ref and the "full_name" of the head repo
-            // so that we can check out the correct repository & branch (including forks).
-            core.setOutput("prNumber", prNumber);
-            core.setOutput("headRef", pr.head.ref);
-            core.setOutput("headRepoFullName", pr.head.repo.full_name);
-
-      - name: Check out PR branch
-        uses: actions/checkout@v3
-        env: 
-          HEADREPOFULLNAME: ${{ steps.pr_info.outputs.headRepoFullName }}
-          HEADREF: ${{ steps.pr_info.outputs.headRef }}
-        with:
-          # Instead of checking out the base repo, use the contributor's repo name
-          repository: ${{ env.HEADREPOFULLNAME }}
-          ref: ${{ env.HEADREF }}
-          # You may need fetch-depth: 0 for being able to push
-          fetch-depth: 0
-          token: ${{ secrets.GITHUB_TOKEN }}
-      
-      - name: Debug
-        env: 
-          HEADREPOFULLNAME: ${{ steps.pr_info.outputs.headRepoFullName }}
-          HEADREF: ${{ steps.pr_info.outputs.headRef }}
-          PRNUMBER: ${{ steps.pr_info.outputs.prNumber }}
-        run: |
-          echo "PR number: $PRNUMBER"
-          echo "Head Ref: $HEADREF"
-          echo "Head Repo Full Name: $HEADREPOFULLNAME"
-
-      - name: Set up Python
-        uses: actions/setup-python@v4
-
-      - name: Install dependencies
-        run: |
-          pip install .[quality]
-
-      - name: Download Makefile from main branch
-        run: |
-          curl -o main_Makefile https://raw.githubusercontent.com/huggingface/diffusers/main/Makefile
-        
-      - name: Compare Makefiles
-        run: |
-          if ! diff -q main_Makefile Makefile; then
-            echo "Error: The Makefile has changed. Please ensure it matches the main branch."
-            exit 1
-          fi
-          echo "No changes in Makefile. Proceeding..."
-          rm -rf main_Makefile
-
-      - name: Run make style and make quality
-        run: |
-          make style && make quality
-
-      - name: Commit and push changes
-        id: commit_and_push
-        env: 
-          HEADREPOFULLNAME: ${{ steps.pr_info.outputs.headRepoFullName }}
-          HEADREF: ${{ steps.pr_info.outputs.headRef }}
-          PRNUMBER: ${{ steps.pr_info.outputs.prNumber }}
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          echo "HEADREPOFULLNAME: $HEADREPOFULLNAME, HEADREF: $HEADREF"
-          # Configure git with the Actions bot user
-          git config user.name "github-actions[bot]"
-          git config user.email "github-actions[bot]@users.noreply.github.com"
-
-          # Make sure your 'origin' remote is set to the contributor's fork
-          git remote set-url origin "https://x-access-token:${GITHUB_TOKEN}@github.com/$HEADREPOFULLNAME.git"
-
-          # If there are changes after running style/quality, commit them
-          if [ -n "$(git status --porcelain)" ]; then
-            git add .
-            git commit -m "Apply style fixes"
-            # Push to the original contributor's forked branch
-            git push origin HEAD:$HEADREF
-            echo "changes_pushed=true" >> $GITHUB_OUTPUT
-          else
-            echo "No changes to commit."
-            echo "changes_pushed=false" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Comment on PR with workflow run link
-        if: steps.commit_and_push.outputs.changes_pushed == 'true'
-        uses: actions/github-script@v6
-        with:
-          script: |
-            const prNumber = parseInt(process.env.prNumber, 10);
-            const runUrl = `${process.env.GITHUB_SERVER_URL}/${process.env.GITHUB_REPOSITORY}/actions/runs/${process.env.GITHUB_RUN_ID}`
-
-            await github.rest.issues.createComment({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              issue_number: prNumber,
-              body: `Style fixes have been applied. [View the workflow run here](${runUrl}).`
-            });
-        env:
-          prNumber: ${{ steps.pr_info.outputs.prNumber }}
+  style:
+    uses: huggingface/huggingface_hub/.github/workflows/style-bot-action.yml@main
+    with:
+      python_quality_dependencies: "[quality]"
+    secrets:
+      bot_token: ${{ secrets.GITHUB_TOKEN }}
@@ -3,7 +3,6 @@ name: Fast tests for PRs
 on:
  pull_request:
    branches: [main]
-    types: [synchronize]
    paths:
      - "src/diffusers/**.py"
      - "benchmarks/**.py"
@@ -28,7 +28,51 @@ env:
  PIPELINE_USAGE_CUTOFF: 1000000000 # set high cutoff so that only always-test pipelines run

 jobs:
+  check_code_quality:
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[quality]
+      - name: Check quality
+        run: make quality
+      - name: Check if failure
+        if: ${{ failure() }}
+        run: |
+          echo "Quality check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make style && make quality'" >> $GITHUB_STEP_SUMMARY
+
+  check_repository_consistency:
+    needs: check_code_quality
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[quality]
+      - name: Check repo consistency
+        run: |
+          python utils/check_copies.py
+          python utils/check_dummies.py
+          python utils/check_support_list.py
+          make deps_table_check_updated
+      - name: Check if failure
+        if: ${{ failure() }}
+        run: |
+          echo "Repo consistency check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make fix-copies'" >> $GITHUB_STEP_SUMMARY
+  
  setup_torch_cuda_pipeline_matrix:
+    needs: [check_code_quality, check_repository_consistency]
    name: Setup Torch Pipelines CUDA Slow Tests Matrix
    runs-on:
      group: aws-general-8-plus
@@ -133,6 +177,7 @@ jobs:

  torch_cuda_tests:
    name: Torch CUDA Tests
+    needs: [check_code_quality, check_repository_consistency]
    runs-on:
      group: aws-g4dn-2xlarge
    container:
@@ -201,7 +246,7 @@ jobs:

  run_examples_tests:
    name: Examples PyTorch CUDA tests on Ubuntu
-        pip uninstall transformers -y && python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
+    needs: [check_code_quality, check_repository_consistency]
    runs-on:
      group: aws-g4dn-2xlarge

@@ -220,6 +265,7 @@ jobs:
    - name: Install dependencies
      run: |
        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        pip uninstall transformers -y && python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
        python -m uv pip install -e [quality,test,training]

    - name: Environment
@@ -28,9 +28,9 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
    python3 -m uv pip install --no-cache-dir \
-        torch==2.1.2 \
-        torchvision==0.16.2 \
-        torchaudio==2.1.2 \
+        torch \
+        torchvision \
+        torchaudio\
        onnxruntime \
        --extra-index-url https://download.pytorch.org/whl/cpu && \
    python3 -m uv pip install --no-cache-dir \
@@ -76,6 +76,16 @@
  - local: advanced_inference/outpaint
    title: Outpainting
  title: Advanced inference
+- sections:
+  - local: hybrid_inference/overview
+    title: Overview
+  - local: hybrid_inference/vae_decode
+    title: VAE Decode
+  - local: hybrid_inference/vae_encode
+    title: VAE Encode
+  - local: hybrid_inference/api_reference
+    title: API Reference
+  title: Hybrid Inference
 - sections:
  - local: using-diffusers/cogvideox
    title: CogVideoX
@@ -165,6 +175,8 @@
    title: gguf
  - local: quantization/torchao
    title: torchao
+  - local: quantization/quanto
+    title: quanto
  title: Quantization Methods
 - sections:
  - local: optimization/fp16
@@ -253,19 +265,23 @@
    sections:
    - local: api/models/overview
      title: Overview
+    - local: api/models/auto_model
+      title: AutoModel
    - sections:
      - local: api/models/controlnet
        title: ControlNetModel
+      - local: api/models/controlnet_union
+        title: ControlNetUnionModel
      - local: api/models/controlnet_flux
        title: FluxControlNetModel
      - local: api/models/controlnet_hunyuandit
        title: HunyuanDiT2DControlNetModel
+      - local: api/models/controlnet_sana
+        title: SanaControlNetModel
      - local: api/models/controlnet_sd3
        title: SD3ControlNetModel
      - local: api/models/controlnet_sparsectrl
        title: SparseControlNetModel
-      - local: api/models/controlnet_union
-        title: ControlNetUnionModel
      title: ControlNets
    - sections:
      - local: api/models/allegro_transformer3d
@@ -274,28 +290,32 @@
        title: AuraFlowTransformer2DModel
      - local: api/models/cogvideox_transformer3d
        title: CogVideoXTransformer3DModel
-      - local: api/models/consisid_transformer3d
-        title: ConsisIDTransformer3DModel
      - local: api/models/cogview3plus_transformer2d
        title: CogView3PlusTransformer2DModel
      - local: api/models/cogview4_transformer2d
        title: CogView4Transformer2DModel
+      - local: api/models/consisid_transformer3d
+        title: ConsisIDTransformer3DModel
      - local: api/models/dit_transformer2d
        title: DiTTransformer2DModel
+      - local: api/models/easyanimate_transformer3d
+        title: EasyAnimateTransformer3DModel
      - local: api/models/flux_transformer
        title: FluxTransformer2DModel
+      - local: api/models/hidream_image_transformer
+        title: HiDreamImageTransformer2DModel
      - local: api/models/hunyuan_transformer2d
        title: HunyuanDiT2DModel
      - local: api/models/hunyuan_video_transformer_3d
        title: HunyuanVideoTransformer3DModel
      - local: api/models/latte_transformer3d
        title: LatteTransformer3DModel
-      - local: api/models/lumina_nextdit2d
-        title: LuminaNextDiT2DModel
-      - local: api/models/lumina2_transformer2d
-        title: Lumina2Transformer2DModel
      - local: api/models/ltx_video_transformer3d
        title: LTXVideoTransformer3DModel
+      - local: api/models/lumina2_transformer2d
+        title: Lumina2Transformer2DModel
+      - local: api/models/lumina_nextdit2d
+        title: LuminaNextDiT2DModel
      - local: api/models/mochi_transformer3d
        title: MochiTransformer3DModel
      - local: api/models/omnigen_transformer
@@ -304,26 +324,28 @@
        title: PixArtTransformer2DModel
      - local: api/models/prior_transformer
        title: PriorTransformer
-      - local: api/models/sd3_transformer2d
-        title: SD3Transformer2DModel
      - local: api/models/sana_transformer2d
        title: SanaTransformer2DModel
+      - local: api/models/sd3_transformer2d
+        title: SD3Transformer2DModel
      - local: api/models/stable_audio_transformer
        title: StableAudioDiTModel
      - local: api/models/transformer2d
        title: Transformer2DModel
      - local: api/models/transformer_temporal
        title: TransformerTemporalModel
+      - local: api/models/wan_transformer_3d
+        title: WanTransformer3DModel
      title: Transformers
    - sections:
      - local: api/models/stable_cascade_unet
        title: StableCascadeUNet
      - local: api/models/unet
        title: UNet1DModel
-      - local: api/models/unet2d
-        title: UNet2DModel
      - local: api/models/unet2d-cond
        title: UNet2DConditionModel
+      - local: api/models/unet2d
+        title: UNet2DModel
      - local: api/models/unet3d-cond
        title: UNet3DConditionModel
      - local: api/models/unet-motion
@@ -332,6 +354,10 @@
        title: UViT2DModel
      title: UNets
    - sections:
+      - local: api/models/asymmetricautoencoderkl
+        title: AsymmetricAutoencoderKL
+      - local: api/models/autoencoder_dc
+        title: AutoencoderDC
      - local: api/models/autoencoderkl
        title: AutoencoderKL
      - local: api/models/autoencoderkl_allegro
@@ -342,12 +368,12 @@
        title: AutoencoderKLHunyuanVideo
      - local: api/models/autoencoderkl_ltx_video
        title: AutoencoderKLLTXVideo
+      - local: api/models/autoencoderkl_magvit
+        title: AutoencoderKLMagvit
      - local: api/models/autoencoderkl_mochi
        title: AutoencoderKLMochi
-      - local: api/models/asymmetricautoencoderkl
-        title: AsymmetricAutoencoderKL
-      - local: api/models/autoencoder_dc
-        title: AutoencoderDC
+      - local: api/models/autoencoder_kl_wan
+        title: AutoencoderKLWan
      - local: api/models/consistency_decoder_vae
        title: ConsistencyDecoderVAE
      - local: api/models/autoencoder_oobleck
@@ -400,6 +426,8 @@
      title: ControlNet with Stable Diffusion 3
    - local: api/pipelines/controlnet_sdxl
      title: ControlNet with Stable Diffusion XL
+    - local: api/pipelines/controlnet_sana
+      title: ControlNet-Sana
    - local: api/pipelines/controlnetxs
      title: ControlNet-XS
    - local: api/pipelines/controlnetxs_sdxl
@@ -418,10 +446,14 @@
      title: DiffEdit
    - local: api/pipelines/dit
      title: DiT
+    - local: api/pipelines/easyanimate
+      title: EasyAnimate
    - local: api/pipelines/flux
      title: Flux
    - local: api/pipelines/control_flux_inpaint
      title: FluxControlInpaint
+    - local: api/pipelines/hidream
+      title: HiDream-I1
    - local: api/pipelines/hunyuandit
      title: Hunyuan-DiT
    - local: api/pipelines/hunyuan_video
@@ -474,6 +506,8 @@
      title: PixArt-Σ
    - local: api/pipelines/sana
      title: Sana
+    - local: api/pipelines/sana_sprint
+      title: Sana Sprint
    - local: api/pipelines/self_attention_guidance
      title: Self-Attention Guidance
    - local: api/pipelines/semantic_stable_diffusion
@@ -487,40 +521,40 @@
    - sections:
      - local: api/pipelines/stable_diffusion/overview
        title: Overview
-      - local: api/pipelines/stable_diffusion/text2img
-        title: Text-to-image
+      - local: api/pipelines/stable_diffusion/depth2img
+        title: Depth-to-image
+      - local: api/pipelines/stable_diffusion/gligen
+        title: GLIGEN (Grounded Language-to-Image Generation)
+      - local: api/pipelines/stable_diffusion/image_variation
+        title: Image variation
      - local: api/pipelines/stable_diffusion/img2img
        title: Image-to-image
      - local: api/pipelines/stable_diffusion/svd
        title: Image-to-video
      - local: api/pipelines/stable_diffusion/inpaint
        title: Inpainting
-      - local: api/pipelines/stable_diffusion/depth2img
-        title: Depth-to-image
-      - local: api/pipelines/stable_diffusion/image_variation
-        title: Image variation
+      - local: api/pipelines/stable_diffusion/k_diffusion
+        title: K-Diffusion
+      - local: api/pipelines/stable_diffusion/latent_upscale
+        title: Latent upscaler
+      - local: api/pipelines/stable_diffusion/ldm3d_diffusion
+        title: LDM3D Text-to-(RGB, Depth), Text-to-(RGB-pano, Depth-pano), LDM3D Upscaler
      - local: api/pipelines/stable_diffusion/stable_diffusion_safe
        title: Safe Stable Diffusion
+      - local: api/pipelines/stable_diffusion/sdxl_turbo
+        title: SDXL Turbo
      - local: api/pipelines/stable_diffusion/stable_diffusion_2
        title: Stable Diffusion 2
      - local: api/pipelines/stable_diffusion/stable_diffusion_3
        title: Stable Diffusion 3
      - local: api/pipelines/stable_diffusion/stable_diffusion_xl
        title: Stable Diffusion XL
-      - local: api/pipelines/stable_diffusion/sdxl_turbo
-        title: SDXL Turbo
-      - local: api/pipelines/stable_diffusion/latent_upscale
-        title: Latent upscaler
      - local: api/pipelines/stable_diffusion/upscale
        title: Super-resolution
-      - local: api/pipelines/stable_diffusion/k_diffusion
-        title: K-Diffusion
-      - local: api/pipelines/stable_diffusion/ldm3d_diffusion
-        title: LDM3D Text-to-(RGB, Depth), Text-to-(RGB-pano, Depth-pano), LDM3D Upscaler
      - local: api/pipelines/stable_diffusion/adapter
        title: T2I-Adapter
-      - local: api/pipelines/stable_diffusion/gligen
-        title: GLIGEN (Grounded Language-to-Image Generation)
+      - local: api/pipelines/stable_diffusion/text2img
+        title: Text-to-image
      title: Stable Diffusion
    - local: api/pipelines/stable_unclip
      title: Stable unCLIP
@@ -534,6 +568,8 @@
      title: UniDiffuser
    - local: api/pipelines/value_guided_sampling
      title: Value-guided sampling
+    - local: api/pipelines/wan
+      title: Wan
    - local: api/pipelines/wuerstchen
      title: Wuerstchen
    title: Pipelines
@@ -38,6 +38,33 @@ config = PyramidAttentionBroadcastConfig(
 pipe.transformer.enable_cache(config)
 ```

+## Faster Cache
+
+[FasterCache](https://huggingface.co/papers/2410.19355) from Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, Kwan-Yee K. Wong.
+
+FasterCache is a method that speeds up inference in diffusion transformers by:
+- Reusing attention states between successive inference steps, due to high similarity between them
+- Skipping unconditional branch prediction used in classifier-free guidance by revealing redundancies between unconditional and conditional branch outputs for the same timestep, and therefore approximating the unconditional branch output using the conditional branch output
+
+```python
+import torch
+from diffusers import CogVideoXPipeline, FasterCacheConfig
+
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+config = FasterCacheConfig(
+    spatial_attention_block_skip_range=2,
+    spatial_attention_timestep_skip_range=(-1, 681),
+    current_timestep_callback=lambda: pipe.current_timestep,
+    attention_weight_callback=lambda _: 0.3,
+    unconditional_batch_skip_range=5,
+    unconditional_batch_timestep_skip_range=(-1, 781),
+    tensor_format="BFCHW",
+)
+pipe.transformer.enable_cache(config)
+```
+
 ### CacheMixin

 [[autodoc]] CacheMixin
@@ -47,3 +74,9 @@ pipe.transformer.enable_cache(config)
 [[autodoc]] PyramidAttentionBroadcastConfig

 [[autodoc]] apply_pyramid_attention_broadcast
+
+### FasterCacheConfig
+
+[[autodoc]] FasterCacheConfig
+
+[[autodoc]] apply_faster_cache
@@ -20,10 +20,13 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 - [`FluxLoraLoaderMixin`] provides similar functions for [Flux](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux).
 - [`CogVideoXLoraLoaderMixin`] provides similar functions for [CogVideoX](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox).
 - [`Mochi1LoraLoaderMixin`] provides similar functions for [Mochi](https://huggingface.co/docs/diffusers/main/en/api/pipelines/mochi).
+- [`AuraFlowLoraLoaderMixin`] provides similar functions for [AuraFlow](https://huggingface.co/fal/AuraFlow).
 - [`LTXVideoLoraLoaderMixin`] provides similar functions for [LTX-Video](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ltx_video).
 - [`SanaLoraLoaderMixin`] provides similar functions for [Sana](https://huggingface.co/docs/diffusers/main/en/api/pipelines/sana).
 - [`HunyuanVideoLoraLoaderMixin`] provides similar functions for [HunyuanVideo](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hunyuan_video).
 - [`Lumina2LoraLoaderMixin`] provides similar functions for [Lumina2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/lumina2).
+- [`WanLoraLoaderMixin`] provides similar functions for [Wan](https://huggingface.co/docs/diffusers/main/en/api/pipelines/wan).
+- [`CogView4LoraLoaderMixin`] provides similar functions for [CogView4](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogview4).
 - [`AmusedLoraLoaderMixin`] is for the [`AmusedPipeline`].
 - [`LoraBaseMixin`] provides a base class with several utility methods to fuse, unfuse, unload, LoRAs and more.

@@ -56,6 +59,9 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse
 ## Mochi1LoraLoaderMixin

 [[autodoc]] loaders.lora_pipeline.Mochi1LoraLoaderMixin
+## AuraFlowLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.AuraFlowLoraLoaderMixin

 ## LTXVideoLoraLoaderMixin

@@ -73,6 +79,14 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse

 [[autodoc]] loaders.lora_pipeline.Lumina2LoraLoaderMixin

+## CogView4LoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.CogView4LoraLoaderMixin
+
+## WanLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.WanLoraLoaderMixin
+
 ## AmusedLoraLoaderMixin

 [[autodoc]] loaders.lora_pipeline.AmusedLoraLoaderMixin
@@ -0,0 +1,29 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# AutoModel
+
+The `AutoModel` is designed to make it easy to load a checkpoint without needing to know the specific model class. `AutoModel` automatically retrieves the correct model class from the checkpoint `config.json` file.
+
+```python
+from diffusers import AutoModel, AutoPipelineForText2Image
+
+unet = AutoModel.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", subfolder="unet")
+pipe = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", unet=unet)
+```
+
+
+## AutoModel
+
+[[autodoc]] AutoModel
+	- all
+	- from_pretrained
@@ -0,0 +1,32 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# AutoencoderKLWan
+
+The 3D variational autoencoder (VAE) model with KL loss used in [Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLWan
+
+vae = AutoencoderKLWan.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
+```
+
+## AutoencoderKLWan
+
+[[autodoc]] AutoencoderKLWan
+  - decode
+  - all
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput
@@ -18,7 +18,7 @@ The model can be loaded with the following code snippet.
 ```python
 from diffusers import AutoencoderKLAllegro

-vae = AutoencoderKLCogVideoX.from_pretrained("rhymes-ai/Allegro", subfolder="vae", torch_dtype=torch.float32).to("cuda")
+vae = AutoencoderKLAllegro.from_pretrained("rhymes-ai/Allegro", subfolder="vae", torch_dtype=torch.float32).to("cuda")
 ```

 ## AutoencoderKLAllegro
@@ -0,0 +1,37 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# AutoencoderKLMagvit
+
+The 3D variational autoencoder (VAE) model with KL loss used in [EasyAnimate](https://github.com/aigc-apps/EasyAnimate) was introduced by Alibaba PAI.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLMagvit
+
+vae = AutoencoderKLMagvit.from_pretrained("alibaba-pai/EasyAnimateV5.1-12b-zh", subfolder="vae", torch_dtype=torch.float16).to("cuda")
+```
+
+## AutoencoderKLMagvit
+
+[[autodoc]] AutoencoderKLMagvit
+    - decode
+    - encode
+    - all
+
+## AutoencoderKLOutput
+
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput
@@ -0,0 +1,29 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# SanaControlNetModel
+
+The ControlNet model was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, Maneesh Agrawala. It provides a greater degree of control over text-to-image generation by conditioning the model on additional inputs such as edge maps, depth maps, segmentation maps, and keypoints for pose detection.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+This model was contributed by [ishan24](https://huggingface.co/ishan24). ❤️
+The original codebase can be found at [NVlabs/Sana](https://github.com/NVlabs/Sana), and you can find official ControlNet checkpoints on [Efficient-Large-Model's](https://huggingface.co/Efficient-Large-Model) Hub profile.
+
+## SanaControlNetModel
+[[autodoc]] SanaControlNetModel
+
+## SanaControlNetOutput
+[[autodoc]] models.controlnets.controlnet_sana.SanaControlNetOutput
+
@@ -0,0 +1,30 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# EasyAnimateTransformer3DModel
+
+A Diffusion Transformer model for 3D data from [EasyAnimate](https://github.com/aigc-apps/EasyAnimate) was introduced by Alibaba PAI.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import EasyAnimateTransformer3DModel
+
+transformer = EasyAnimateTransformer3DModel.from_pretrained("alibaba-pai/EasyAnimateV5.1-12b-zh", subfolder="transformer", torch_dtype=torch.float16).to("cuda")
+```
+
+## EasyAnimateTransformer3DModel
+
+[[autodoc]] EasyAnimateTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput
@@ -0,0 +1,30 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# HiDreamImageTransformer2DModel
+
+A Transformer model for image-like data from [HiDream-I1](https://huggingface.co/HiDream-ai).
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import HiDreamImageTransformer2DModel
+
+transformer = HiDreamImageTransformer2DModel.from_pretrained("HiDream-ai/HiDream-I1-Full", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## HiDreamImageTransformer2DModel
+
+[[autodoc]] HiDreamImageTransformer2DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput
@@ -0,0 +1,30 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# WanTransformer3DModel
+
+A Diffusion Transformer model for 3D video-like data was introduced in [Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import WanTransformer3DModel
+
+transformer = WanTransformer3DModel.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## WanTransformer3DModel
+
+[[autodoc]] WanTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput
@@ -89,6 +89,23 @@ image = pipeline(prompt).images[0]
 image.save("auraflow.png")
 ```

+## Support for `torch.compile()`
+
+AuraFlow can be compiled with `torch.compile()` to speed up inference latency even for different resolutions. First, install PyTorch nightly following the instructions from [here](https://pytorch.org/). The snippet below shows the changes needed to enable this:
+
+```diff
+ torch.fx.experimental._config.use_duck_shape = False
+ pipeline.transformer = torch.compile(
+    pipeline.transformer, fullgraph=True, dynamic=True
+)
+```
+
+Specifying `use_duck_shape` to be `False` instructs the compiler if it should use the same symbolic variable to represent input sizes that are the same. For more details, check out [this comment](https://github.com/huggingface/diffusers/pull/11327#discussion_r2047659790).
+
+This enables from 100% (on low resolutions) to a 30% (on 1536x1536 resolution) speed improvements.
+
+Thanks to [AstraliteHeart](https://github.com/huggingface/diffusers/pull/11297/) who helped us rewrite the [`AuraFlowTransformer2DModel`] class so that the above works for different resolutions ([PR](https://github.com/huggingface/diffusers/pull/11297/)).
+
 ## AuraFlowPipeline

 [[autodoc]] AuraFlowPipeline
@@ -0,0 +1,36 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# ControlNet
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
+ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.
+
+With a ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+This pipeline was contributed by [ishan24](https://huggingface.co/ishan24). ❤️
+The original codebase can be found at [NVlabs/Sana](https://github.com/NVlabs/Sana), and you can find official ControlNet checkpoints on [Efficient-Large-Model's](https://huggingface.co/Efficient-Large-Model) Hub profile.
+
+## SanaControlNetPipeline
+[[autodoc]] SanaControlNetPipeline
+	- all
+	- __call__
+
+## SanaPipelineOutput
+[[autodoc]] pipelines.sana.pipeline_output.SanaPipelineOutput
@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.

 <div class="flex flex-wrap space-x-1">
  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>

 ## Overview
@@ -0,0 +1,88 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+-->
+
+# EasyAnimate
+[EasyAnimate](https://github.com/aigc-apps/EasyAnimate) by Alibaba PAI.
+
+The description from it's GitHub page:
+*EasyAnimate is a pipeline based on the transformer architecture, designed for generating AI images and videos, and for training baseline models and Lora models for Diffusion Transformer. We support direct prediction from pre-trained EasyAnimate models, allowing for the generation of videos with various resolutions, approximately 6 seconds in length, at 8fps (EasyAnimateV5.1, 1 to 49 frames). Additionally, users can train their own baseline and Lora models for specific style transformations.*
+
+This pipeline was contributed by [bubbliiiing](https://github.com/bubbliiiing). The original codebase can be found [here](https://huggingface.co/alibaba-pai). The original weights can be found under [hf.co/alibaba-pai](https://huggingface.co/alibaba-pai).
+
+There are two official EasyAnimate checkpoints for text-to-video and video-to-video.
+
+| checkpoints | recommended inference dtype |
+|:---:|:---:|
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh) | torch.float16 |
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh-InP`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh-InP) | torch.float16 |
+
+There is one official EasyAnimate checkpoints available for image-to-video and video-to-video.
+
+| checkpoints | recommended inference dtype |
+|:---:|:---:|
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh-InP`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh-InP) | torch.float16 |
+
+There are two official EasyAnimate checkpoints available for control-to-video.
+
+| checkpoints | recommended inference dtype |
+|:---:|:---:|
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh-Control`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh-Control) | torch.float16 |
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh-Control-Camera`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh-Control-Camera) | torch.float16 |
+
+For the EasyAnimateV5.1 series:
+- Text-to-video (T2V) and Image-to-video (I2V) works for multiple resolutions. The width and height can vary from 256 to 1024.
+- Both T2V and I2V models support generation with 1~49 frames and work best at this value. Exporting videos at 8 FPS is recommended.
+
+## Quantization
+
+Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
+
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`EasyAnimatePipeline`] for inference with bitsandbytes.
+
+```py
+import torch
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, EasyAnimateTransformer3DModel, EasyAnimatePipeline
+from diffusers.utils import export_to_video
+
+quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
+transformer_8bit = EasyAnimateTransformer3DModel.from_pretrained(
+    "alibaba-pai/EasyAnimateV5.1-12b-zh",
+    subfolder="transformer",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+pipeline = EasyAnimatePipeline.from_pretrained(
+    "alibaba-pai/EasyAnimateV5.1-12b-zh",
+    transformer=transformer_8bit,
+    torch_dtype=torch.float16,
+    device_map="balanced",
+)
+
+prompt = "A cat walks on the grass, realistic style."
+negative_prompt = "bad detailed"
+video = pipeline(prompt=prompt, negative_prompt=negative_prompt, num_frames=49, num_inference_steps=30).frames[0]
+export_to_video(video, "cat.mp4", fps=8)
+```
+
+## EasyAnimatePipeline
+
+[[autodoc]] EasyAnimatePipeline
+  - all
+  - __call__
+
+## EasyAnimatePipelineOutput
+
+[[autodoc]] pipelines.easyanimate.pipeline_output.EasyAnimatePipelineOutput
@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.

 <div class="flex flex-wrap space-x-1">
  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>

 Flux is a series of text-to-image generation models based on diffusion transformers. To know more about Flux, check out the original [blog post](https://blackforestlabs.ai/announcing-black-forest-labs/) by the creators of Flux, Black Forest Labs.
@@ -0,0 +1,43 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+# HiDreamImage
+
+[HiDream-I1](https://huggingface.co/HiDream-ai) by HiDream.ai
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+## Available models
+
+The following models are available for the [`HiDreamImagePipeline`](text-to-image) pipeline:
+
+| Model name | Description |
+|:---|:---|
+| [`HiDream-ai/HiDream-I1-Full`](https://huggingface.co/HiDream-ai/HiDream-I1-Full) | - |
+| [`HiDream-ai/HiDream-I1-Dev`](https://huggingface.co/HiDream-ai/HiDream-I1-Dev) | - |
+| [`HiDream-ai/HiDream-I1-Fast`](https://huggingface.co/HiDream-ai/HiDream-I1-Fast) | - |
+
+## HiDreamImagePipeline
+
+[[autodoc]] HiDreamImagePipeline
+  - all
+  - __call__
+
+## HiDreamImagePipelineOutput
+
+[[autodoc]] pipelines.hidream_image.pipeline_output.HiDreamImagePipelineOutput
@@ -49,7 +49,9 @@ The following models are available for the image-to-video pipeline:

 | Model name | Description |
 |:---|:---|
-| [`https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-I2V`](https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-I2V) | Skywork's custom finetune of HunyuanVideo (de-distilled). Performs best with `97x544x960` resolution. Performs best at `97x544x960` resolution, `guidance_scale=1.0`, `true_cfg_scale=6.0` and a negative prompt. |
+| [`Skywork/SkyReels-V1-Hunyuan-I2V`](https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-I2V) | Skywork's custom finetune of HunyuanVideo (de-distilled). Performs best with `97x544x960` resolution. Performs best at `97x544x960` resolution, `guidance_scale=1.0`, `true_cfg_scale=6.0` and a negative prompt. |
+| [`hunyuanvideo-community/HunyuanVideo-I2V-33ch`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo 33-channel I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20). |
+| [`hunyuanvideo-community/HunyuanVideo-I2V`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo 16-channel I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20) |

 ## Quantization

@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.

 <div class="flex flex-wrap space-x-1">
  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>

 ![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/kolors/kolors_header_collage.png)
@@ -16,6 +16,7 @@

 <div class="flex flex-wrap space-x-1">
  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>

 [LTX Video](https://huggingface.co/Lightricks/LTX-Video) is the first DiT-based video generation model capable of generating high-quality videos in real-time. It produces 24 FPS videos at a 768x512 resolution faster than they can be watched. Trained on a large-scale dataset of diverse videos, the model generates high-resolution videos with realistic and varied content. We provide a model for both text-to-video as well as image + text-to-video usecases.
@@ -32,6 +33,7 @@ Available models:
 |:-------------:|:-----------------:|
 | [`LTX Video 0.9.0`](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.safetensors) | `torch.bfloat16` |
 | [`LTX Video 0.9.1`](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.1.safetensors) | `torch.bfloat16` |
+| [`LTX Video 0.9.5`](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.5.safetensors) | `torch.bfloat16` |

 Note: The recommended dtype is for the transformer component. The VAE and text encoders can be either `torch.float32`, `torch.bfloat16` or `torch.float16` but the recommended dtype is `torch.bfloat16` as used in the original repository.

@@ -196,6 +198,12 @@ export_to_video(video, "ship.mp4", fps=24)
  - all
  - __call__

+## LTXConditionPipeline
+
+[[autodoc]] LTXConditionPipeline
+  - all
+  - __call__
+
 ## LTXPipelineOutput

 [[autodoc]] pipelines.ltx.pipeline_output.LTXPipelineOutput
@@ -58,10 +58,10 @@ Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fa
 First, load the pipeline:

 ```python
-from diffusers import LuminaText2ImgPipeline
+from diffusers import LuminaPipeline
 import torch

-pipeline = LuminaText2ImgPipeline.from_pretrained(
+pipeline = LuminaPipeline.from_pretrained(
 	"Alpha-VLLM/Lumina-Next-SFT-diffusers", torch_dtype=torch.bfloat16
 ).to("cuda")
 ```
@@ -86,11 +86,11 @@ image = pipeline(prompt="Upper body of a young woman in a Victorian-era outfit w

 Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.

-Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`LuminaText2ImgPipeline`] for inference with bitsandbytes.
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`LuminaPipeline`] for inference with bitsandbytes.

 ```py
 import torch
-from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, Transformer2DModel, LuminaText2ImgPipeline
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, Transformer2DModel, LuminaPipeline
 from transformers import BitsAndBytesConfig as BitsAndBytesConfig, T5EncoderModel

 quant_config = BitsAndBytesConfig(load_in_8bit=True)
@@ -109,7 +109,7 @@ transformer_8bit = Transformer2DModel.from_pretrained(
    torch_dtype=torch.float16,
 )

-pipeline = LuminaText2ImgPipeline.from_pretrained(
+pipeline = LuminaPipeline.from_pretrained(
    "Alpha-VLLM/Lumina-Next-SFT-diffusers",
    text_encoder=text_encoder_8bit,
    transformer=transformer_8bit,
@@ -122,9 +122,9 @@ image = pipeline(prompt).images[0]
 image.save("lumina.png")
 ```

-## LuminaText2ImgPipeline
+## LuminaPipeline

-[[autodoc]] LuminaText2ImgPipeline
+[[autodoc]] LuminaPipeline
 	- all
 	- __call__

@@ -36,14 +36,14 @@ Single file loading for Lumina Image 2.0 is available for the `Lumina2Transforme

 ```python
 import torch
-from diffusers import Lumina2Transformer2DModel, Lumina2Text2ImgPipeline
+from diffusers import Lumina2Transformer2DModel, Lumina2Pipeline

 ckpt_path = "https://huggingface.co/Alpha-VLLM/Lumina-Image-2.0/blob/main/consolidated.00-of-01.pth"
 transformer = Lumina2Transformer2DModel.from_single_file(
    ckpt_path, torch_dtype=torch.bfloat16
 )

-pipe = Lumina2Text2ImgPipeline.from_pretrained(
+pipe = Lumina2Pipeline.from_pretrained(
    "Alpha-VLLM/Lumina-Image-2.0", transformer=transformer, torch_dtype=torch.bfloat16
 )
 pipe.enable_model_cpu_offload()
@@ -60,7 +60,7 @@ image.save("lumina-single-file.png")
 GGUF Quantized checkpoints for the `Lumina2Transformer2DModel` can be loaded via `from_single_file` with the `GGUFQuantizationConfig` 

 ```python
-from diffusers import Lumina2Transformer2DModel, Lumina2Text2ImgPipeline, GGUFQuantizationConfig 
+from diffusers import Lumina2Transformer2DModel, Lumina2Pipeline, GGUFQuantizationConfig 

 ckpt_path = "https://huggingface.co/calcuis/lumina-gguf/blob/main/lumina2-q4_0.gguf"
 transformer = Lumina2Transformer2DModel.from_single_file(
@@ -69,7 +69,7 @@ transformer = Lumina2Transformer2DModel.from_single_file(
    torch_dtype=torch.bfloat16,
 )

-pipe = Lumina2Text2ImgPipeline.from_pretrained(
+pipe = Lumina2Pipeline.from_pretrained(
    "Alpha-VLLM/Lumina-Image-2.0", transformer=transformer, torch_dtype=torch.bfloat16
 )
 pipe.enable_model_cpu_offload()
@@ -80,8 +80,8 @@ image = pipe(
 image.save("lumina-gguf.png")
 ```

-## Lumina2Text2ImgPipeline
+## Lumina2Pipeline

-[[autodoc]] Lumina2Text2ImgPipeline
+[[autodoc]] Lumina2Pipeline
  - all
  - __call__
@@ -16,6 +16,7 @@

 <div class="flex flex-wrap space-x-1">
  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>

 [SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.
@@ -0,0 +1,100 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+# SANA-Sprint
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
+[SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation](https://huggingface.co/papers/2503.09641) from NVIDIA, MIT HAN Lab, and Hugging Face by Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Enze Xie, Song Han
+
+The abstract from the paper is:
+
+*This paper presents SANA-Sprint, an efficient diffusion model for ultra-fast text-to-image (T2I) generation. SANA-Sprint is built on a pre-trained foundation model and augmented with hybrid distillation, dramatically reducing inference steps from 20 to 1-4. We introduce three key innovations: (1) We propose a training-free approach that transforms a pre-trained flow-matching model for continuous-time consistency distillation (sCM), eliminating costly training from scratch and achieving high training efficiency. Our hybrid distillation strategy combines sCM with latent adversarial distillation (LADD): sCM ensures alignment with the teacher model, while LADD enhances single-step generation fidelity. (2) SANA-Sprint is a unified step-adaptive model that achieves high-quality generation in 1-4 steps, eliminating step-specific training and improving efficiency. (3) We integrate ControlNet with SANA-Sprint for real-time interactive image generation, enabling instant visual feedback for user interaction. SANA-Sprint establishes a new Pareto frontier in speed-quality tradeoffs, achieving state-of-the-art performance with 7.59 FID and 0.74 GenEval in only 1 step — outperforming FLUX-schnell (7.94 FID / 0.71 GenEval) while being 10× faster (0.1s vs 1.1s on H100). It also achieves 0.1s (T2I) and 0.25s (ControlNet) latency for 1024×1024 images on H100, and 0.31s (T2I) on an RTX 4090, showcasing its exceptional efficiency and potential for AI-powered consumer applications (AIPC). Code and pre-trained models will be open-sourced.*
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+This pipeline was contributed by [lawrence-cj](https://github.com/lawrence-cj), [shuchen Xue](https://github.com/scxue) and [Enze Xie](https://github.com/xieenze). The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://huggingface.co/Efficient-Large-Model/).
+
+Available models:
+
+|                                                                    Model                                                                    | Recommended dtype |
+|:-------------------------------------------------------------------------------------------------------------------------------------------:|:-----------------:|
+| [`Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers) | `torch.bfloat16`  |
+| [`Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers) | `torch.bfloat16`  |
+
+Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-sprint-67d6810d65235085b3b17c76) collection for more information.
+
+Note: The recommended dtype mentioned is for the transformer weights. The text encoder must stay in `torch.bfloat16` and VAE weights must stay in `torch.bfloat16` or `torch.float32` for the model to work correctly. Please refer to the inference example below to see how to load the model with the recommended dtype. 
+
+
+## Quantization
+
+Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
+
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`SanaSprintPipeline`] for inference with bitsandbytes.
+
+```py
+import torch
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, SanaTransformer2DModel, SanaSprintPipeline
+from transformers import BitsAndBytesConfig as BitsAndBytesConfig, AutoModel
+
+quant_config = BitsAndBytesConfig(load_in_8bit=True)
+text_encoder_8bit = AutoModel.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
+    subfolder="text_encoder",
+    quantization_config=quant_config,
+    torch_dtype=torch.bfloat16,
+)
+
+quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
+transformer_8bit = SanaTransformer2DModel.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
+    subfolder="transformer",
+    quantization_config=quant_config,
+    torch_dtype=torch.bfloat16,
+)
+
+pipeline = SanaSprintPipeline.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
+    text_encoder=text_encoder_8bit,
+    transformer=transformer_8bit,
+    torch_dtype=torch.bfloat16,
+    device_map="balanced",
+)
+
+prompt = "a tiny astronaut hatching from an egg on the moon"
+image = pipeline(prompt).images[0]
+image.save("sana.png")
+```
+
+## Setting `max_timesteps`
+
+Users can tweak the `max_timesteps` value for experimenting with the visual quality of the generated outputs. The default `max_timesteps` value was obtained with an inference-time search process. For more details about it, check out the paper.
+
+## SanaSprintPipeline
+
+[[autodoc]] SanaSprintPipeline
+  - all
+  - __call__
+
+
+## SanaPipelineOutput
+
+[[autodoc]] pipelines.sana.pipeline_output.SanaPipelineOutput
@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.

 <div class="flex flex-wrap space-x-1">
  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>

 Stable Diffusion 3 (SD3) was proposed in [Scaling Rectified Flow Transformers for High-Resolution Image Synthesis](https://arxiv.org/pdf/2403.03206.pdf) by Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach.
@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.

 <div class="flex flex-wrap space-x-1">
  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>

 Stable Diffusion XL (SDXL) was proposed in [SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis](https://huggingface.co/papers/2307.01952) by Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach.
@@ -0,0 +1,465 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+# Wan
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
+[Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+
+<!-- TODO(aryan): update abstract once paper is out -->
+
+## Generating Videos with Wan 2.1
+
+We will first need to install some addtional dependencies.
+
+```shell
+pip install -u ftfy imageio-ffmpeg imageio
+```
+
+### Text to Video Generation
+
+The following example requires 11GB VRAM to run and uses the smaller `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` model. You can switch it out
+for the larger `Wan2.1-I2V-14B-720P-Diffusers` or `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` if you have at least 35GB VRAM available.
+
+```python
+from diffusers import WanPipeline
+from diffusers.utils import export_to_video
+
+# Available models: Wan-AI/Wan2.1-I2V-14B-720P-Diffusers or Wan-AI/Wan2.1-I2V-14B-480P-Diffusers
+model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+
+pipe = WanPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
+pipe.enable_model_cpu_offload()
+
+prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+num_frames = 33
+
+frames = pipe(prompt=prompt, negative_prompt=negative_prompt, num_frames=num_frames).frames[0]
+export_to_video(frames, "wan-t2v.mp4", fps=16)
+```
+
+<Tip>
+You can improve the quality of the generated video by running the decoding step in full precision.
+</Tip>
+
+```python
+from diffusers import WanPipeline, AutoencoderKLWan
+from diffusers.utils import export_to_video
+
+model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+
+# replace this with pipe.to("cuda") if you have sufficient VRAM
+pipe.enable_model_cpu_offload()
+
+prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+num_frames = 33
+
+frames = pipe(prompt=prompt, num_frames=num_frames).frames[0]
+export_to_video(frames, "wan-t2v.mp4", fps=16)
+```
+
+### Image to Video Generation
+
+The Image to Video pipeline requires loading the `AutoencoderKLWan` and the `CLIPVisionModel` components in full precision. The following example will need at least
+35GB of VRAM to run.
+
+```python
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, WanImageToVideoPipeline
+from diffusers.utils import export_to_video, load_image
+from transformers import CLIPVisionModel
+
+# Available models: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers, Wan-AI/Wan2.1-I2V-14B-720P-Diffusers
+model_id = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(
+    model_id, subfolder="image_encoder", torch_dtype=torch.float32
+)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id, vae=vae, image_encoder=image_encoder, torch_dtype=torch.bfloat16
+)
+
+# replace this with pipe.to("cuda") if you have sufficient VRAM
+pipe.enable_model_cpu_offload()
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+)
+
+max_area = 480 * 832
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+image = image.resize((width, height))
+
+prompt = (
+    "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in "
+    "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+)
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+num_frames = 33
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    num_frames=num_frames,
+    guidance_scale=5.0,
+).frames[0]
+export_to_video(output, "wan-i2v.mp4", fps=16)
+```
+
+### Video to Video Generation
+
+```python
+import torch
+from diffusers.utils import load_video, export_to_video
+from diffusers import AutoencoderKLWan, WanVideoToVideoPipeline, UniPCMultistepScheduler
+
+# Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(
+    model_id, subfolder="vae", torch_dtype=torch.float32
+)
+pipe = WanVideoToVideoPipeline.from_pretrained(
+    model_id, vae=vae, torch_dtype=torch.bfloat16
+)
+flow_shift = 3.0  # 5.0 for 720P, 3.0 for 480P
+pipe.scheduler = UniPCMultistepScheduler.from_config(
+    pipe.scheduler.config, flow_shift=flow_shift
+)
+# change to pipe.to("cuda") if you have sufficient VRAM
+pipe.enable_model_cpu_offload()
+
+prompt = "A robot standing on a mountain top. The sun is setting in the background"
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+video = load_video(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/hiker.mp4"
+)
+output = pipe(
+    video=video,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=480,
+    width=512,
+    guidance_scale=7.0,
+    strength=0.7,
+).frames[0]
+
+export_to_video(output, "wan-v2v.mp4", fps=16)
+```
+
+## Memory Optimizations for Wan 2.1
+
+Base inference with the large 14B Wan 2.1 models can take up to 35GB of VRAM when generating videos at 720p resolution. We'll outline a few memory optimizations we can apply to reduce the VRAM required to run the model.
+
+We'll use `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers` model in these examples to demonstrate the memory savings, but the techniques are applicable to all model checkpoints.
+
+### Group Offloading the Transformer and UMT5 Text Encoder
+
+Find more information about group offloading [here](../optimization/memory.md)
+
+#### Block Level Group Offloading
+
+We can reduce our VRAM requirements by applying group offloading to the larger model components of the pipeline; the `WanTransformer3DModel` and `UMT5EncoderModel`. Group offloading will break up the individual modules of a model and offload/onload them onto your GPU as needed during inference. In this example, we'll apply `block_level` offloading, which will group the modules in a model into blocks of size `num_blocks_per_group` and offload/onload them to GPU. Moving to between CPU and GPU does add latency to the inference process. You can trade off between latency and memory savings by increasing or decreasing the `num_blocks_per_group`.
+
+The following example will now only require 14GB of VRAM to run, but will take approximately 30 minutes to generate a video.
+
+```python
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, WanTransformer3DModel, WanImageToVideoPipeline
+from diffusers.hooks.group_offloading import apply_group_offloading
+from diffusers.utils import export_to_video, load_image
+from transformers import UMT5EncoderModel, CLIPVisionModel
+
+# Available models: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers, Wan-AI/Wan2.1-I2V-14B-720P-Diffusers
+model_id = "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(
+    model_id, subfolder="image_encoder", torch_dtype=torch.float32
+)
+
+text_encoder = UMT5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+transformer = WanTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+
+onload_device = torch.device("cuda")
+offload_device = torch.device("cpu")
+
+apply_group_offloading(text_encoder,
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="block_level",
+    num_blocks_per_group=4
+)
+
+transformer.enable_group_offload(
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="block_level",
+    num_blocks_per_group=4,
+)
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id,
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    image_encoder=image_encoder,
+    torch_dtype=torch.bfloat16
+)
+# Since we've offloaded the larger models alrady, we can move the rest of the model components to GPU
+pipe.to("cuda")
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+)
+
+max_area = 720 * 832
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+image = image.resize((width, height))
+
+prompt = (
+    "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in "
+    "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+)
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+num_frames = 33
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    num_frames=num_frames,
+    guidance_scale=5.0,
+).frames[0]
+
+export_to_video(output, "wan-i2v.mp4", fps=16)
+```
+
+#### Block Level Group Offloading with CUDA Streams
+
+We can speed up group offloading inference, by enabling the use of [CUDA streams](https://pytorch.org/docs/stable/generated/torch.cuda.Stream.html). However, using CUDA streams requires moving the model parameters into pinned memory. This allocation is handled by Pytorch under the hood, and can result in a significant spike in CPU RAM usage. Please consider this option if your CPU RAM is atleast 2X the size of the model you are group offloading.
+
+In the following example we will use CUDA streams when group offloading the `WanTransformer3DModel`. When testing on an A100, this example will require 14GB of VRAM, 52GB of CPU RAM, but will generate a video in approximately 9 minutes.
+
+```python
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, WanTransformer3DModel, WanImageToVideoPipeline
+from diffusers.hooks.group_offloading import apply_group_offloading
+from diffusers.utils import export_to_video, load_image
+from transformers import UMT5EncoderModel, CLIPVisionModel
+
+# Available models: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers, Wan-AI/Wan2.1-I2V-14B-720P-Diffusers
+model_id = "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(
+    model_id, subfolder="image_encoder", torch_dtype=torch.float32
+)
+
+text_encoder = UMT5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+transformer = WanTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+
+onload_device = torch.device("cuda")
+offload_device = torch.device("cpu")
+
+apply_group_offloading(text_encoder,
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="block_level",
+    num_blocks_per_group=4
+)
+
+transformer.enable_group_offload(
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="leaf_level",
+    use_stream=True
+)
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id,
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    image_encoder=image_encoder,
+    torch_dtype=torch.bfloat16
+)
+# Since we've offloaded the larger models alrady, we can move the rest of the model components to GPU
+pipe.to("cuda")
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+)
+
+max_area = 720 * 832
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+image = image.resize((width, height))
+
+prompt = (
+    "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in "
+    "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+)
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+num_frames = 33
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    num_frames=num_frames,
+    guidance_scale=5.0,
+).frames[0]
+
+export_to_video(output, "wan-i2v.mp4", fps=16)
+```
+
+### Applying Layerwise Casting to the Transformer
+
+Find more information about layerwise casting [here](../optimization/memory.md)
+
+In this example, we will model offloading with layerwise casting. Layerwise casting will downcast each layer's weights to `torch.float8_e4m3fn`, temporarily upcast to `torch.bfloat16` during the forward pass of the layer, then revert to `torch.float8_e4m3fn` afterward. This approach reduces memory requirements by approximately 50% while introducing a minor quality reduction in the generated video due to the precision trade-off.
+
+This example will require 20GB of VRAM.
+
+```python
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, WanTransformer3DModel, WanImageToVideoPipeline
+from diffusers.hooks.group_offloading import apply_group_offloading
+from diffusers.utils import export_to_video, load_image
+from transformers import UMT5EncoderModel, CLIPVisionModel
+
+model_id = "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(
+    model_id, subfolder="image_encoder", torch_dtype=torch.float32
+)
+text_encoder = UMT5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+
+transformer = WanTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+transformer.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
+
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id,
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    image_encoder=image_encoder,
+    torch_dtype=torch.bfloat16
+)
+pipe.enable_model_cpu_offload()
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg")
+
+max_area = 720 * 832
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+image = image.resize((width, height))
+prompt = (
+    "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in "
+    "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+)
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+num_frames = 33
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    num_frames=num_frames,
+    num_inference_steps=50,
+    guidance_scale=5.0,
+).frames[0]
+export_to_video(output, "wan-i2v.mp4", fps=16)
+```
+
+## Using a Custom Scheduler
+
+Wan can be used with many different schedulers, each with their own benefits regarding speed and generation quality. By default, Wan uses the `UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)` scheduler. You can use a different scheduler as follows:
+
+```python
+from diffusers import FlowMatchEulerDiscreteScheduler, UniPCMultistepScheduler, WanPipeline
+
+scheduler_a = FlowMatchEulerDiscreteScheduler(shift=5.0)
+scheduler_b = UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=4.0)
+
+pipe = WanPipeline.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", scheduler=<CUSTOM_SCHEDULER_HERE>)
+
+# or,
+pipe.scheduler = <CUSTOM_SCHEDULER_HERE>
+```
+
+## Using Single File Loading with Wan 2.1
+
+The `WanTransformer3DModel` and `AutoencoderKLWan` models support loading checkpoints in their original format via the `from_single_file` loading
+method.
+
+```python
+import torch
+from diffusers import WanPipeline, WanTransformer3DModel
+
+ckpt_path = "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/diffusion_models/wan2.1_t2v_1.3B_bf16.safetensors"
+transformer = WanTransformer3DModel.from_single_file(ckpt_path, torch_dtype=torch.bfloat16)
+
+pipe = WanPipeline.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", transformer=transformer)
+```
+
+## Recommendations for Inference
+- Keep `AutencoderKLWan` in `torch.float32` for better decoding quality.
+- `num_frames` should satisfy the following constraint: `(num_frames - 1) % 4 == 0`
+- For smaller resolution videos, try lower values of `shift` (between `2.0` to `5.0`) in the [Scheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete#diffusers.FlowMatchEulerDiscreteScheduler.shift). For larger resolution videos, try higher values (between `7.0` and `12.0`). The default value is `3.0` for Wan.
+
+## WanPipeline
+
+[[autodoc]] WanPipeline
+  - all
+  - __call__
+
+## WanImageToVideoPipeline
+
+[[autodoc]] WanImageToVideoPipeline
+  - all
+  - __call__
+
+## WanPipelineOutput
+
+[[autodoc]] pipelines.wan.pipeline_output.WanPipelineOutput
@@ -31,6 +31,11 @@ Learn how to quantize models in the [Quantization](../quantization/overview) gui
 ## GGUFQuantizationConfig

 [[autodoc]] GGUFQuantizationConfig
+
+## QuantoConfig
+
+[[autodoc]] QuantoConfig
+
 ## TorchAoConfig

 [[autodoc]] TorchAoConfig
@@ -83,4 +83,8 @@ Happy exploring, and thank you for being part of the Diffusers community!
    <td><a href="https://github.com/suzukimain/auto_diffusers"> Model Search </a></td>
    <td>Search models on Civitai and Hugging Face</td>
  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/beinsezii/skrample"> Skrample </a></td>
+    <td>Fully modular scheduler functions with 1st class diffusers integration.</td>
+  </tr>
 </table>
@@ -16,6 +16,11 @@ specific language governing permissions and limitations under the License.
    <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
 </a>

+> [!TIP]
+> This document has now grown outdated given the emergence of existing evaluation frameworks for diffusion models for image generation. Please check
+> out works like [HEIM](https://crfm.stanford.edu/helm/heim/latest/), [T2I-Compbench](https://arxiv.org/abs/2307.06350),
+> [GenEval](https://arxiv.org/abs/2310.11513).
+
 Evaluation of generative models like [Stable Diffusion](https://huggingface.co/docs/diffusers/stable_diffusion) is subjective in nature. But as practitioners and researchers, we often have to make careful choices amongst many different possibilities. So, when working with different generative models (like GANs, Diffusion, etc.), how do we choose one over the other?

 Qualitative evaluation of such models can be error-prone and might incorrectly influence a decision.
@@ -0,0 +1,9 @@
+# Hybrid Inference API Reference
+
+## Remote Decode
+
+[[autodoc]] utils.remote_utils.remote_decode
+
+## Remote Encode
+
+[[autodoc]] utils.remote_utils.remote_encode
@@ -0,0 +1,60 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Hybrid Inference
+
+**Empowering local AI builders with Hybrid Inference**
+
+
+> [!TIP]
+> Hybrid Inference is an [experimental feature](https://huggingface.co/blog/remote_vae).
+> Feedback can be provided [here](https://github.com/huggingface/diffusers/issues/new?template=remote-vae-pilot-feedback.yml).
+
+
+
+## Why use Hybrid Inference?
+
+Hybrid Inference offers a fast and simple way to offload local generation requirements.
+
+- 🚀 **Reduced Requirements:** Access powerful models without expensive hardware.
+- 💎 **Without Compromise:** Achieve the highest quality without sacrificing performance.
+- 💰 **Cost Effective:** It's free! 🤑
+- 🎯 **Diverse Use Cases:** Fully compatible with Diffusers 🧨 and the wider community.
+- 🔧 **Developer-Friendly:** Simple requests, fast responses.
+
+---
+
+## Available Models
+
+* **VAE Decode 🖼️:** Quickly decode latent representations into high-quality images without compromising performance or workflow speed.
+* **VAE Encode 🔢:** Efficiently encode images into latent representations for generation and training.
+* **Text Encoders 📃 (coming soon):** Compute text embeddings for your prompts quickly and accurately, ensuring a smooth and high-quality workflow.
+
+---
+
+## Integrations
+
+* **[SD.Next](https://github.com/vladmandic/sdnext):** All-in-one UI with direct supports Hybrid Inference.
+* **[ComfyUI-HFRemoteVae](https://github.com/kijai/ComfyUI-HFRemoteVae):** ComfyUI node for Hybrid Inference.
+
+## Changelog
+
+- March 10 2025: Added VAE encode
+- March 2 2025: Initial release with VAE decoding
+
+## Contents
+
+The documentation is organized into three sections:
+
+* **VAE Decode** Learn the basics of how to use VAE Decode with Hybrid Inference.
+* **VAE Encode** Learn the basics of how to use VAE Encode with Hybrid Inference.
+* **API Reference** Dive into task-specific settings and parameters.
@@ -0,0 +1,345 @@
+# Getting Started: VAE Decode with Hybrid Inference
+
+VAE decode is an essential component of diffusion models - turning latent representations into images or videos.
+
+## Memory
+
+These tables demonstrate the VRAM requirements for VAE decode with SD v1 and SD XL on different GPUs.
+
+For the majority of these GPUs the memory usage % dictates other models (text encoders, UNet/Transformer) must be offloaded, or tiled decoding has to be used which increases time taken and impacts quality.
+
+<details><summary>SD v1.5</summary>
+
+| GPU | Resolution | Time (seconds) | Memory (%) | Tiled Time (secs) | Tiled Memory (%) |
+| --- | --- | --- | --- | --- | --- |
+| NVIDIA GeForce RTX 4090 | 512x512 | 0.031 | 5.60% | 0.031 (0%) | 5.60% |
+| NVIDIA GeForce RTX 4090 | 1024x1024 | 0.148 | 20.00% | 0.301 (+103%) | 5.60% |
+| NVIDIA GeForce RTX 4080 | 512x512 | 0.05 | 8.40% | 0.050 (0%) | 8.40% |
+| NVIDIA GeForce RTX 4080 | 1024x1024 | 0.224 | 30.00% | 0.356 (+59%) | 8.40% |
+| NVIDIA GeForce RTX 4070 Ti | 512x512 | 0.066 | 11.30% | 0.066 (0%) | 11.30% |
+| NVIDIA GeForce RTX 4070 Ti | 1024x1024 | 0.284 | 40.50% | 0.454 (+60%) | 11.40% |
+| NVIDIA GeForce RTX 3090 | 512x512 | 0.062 | 5.20% | 0.062 (0%) | 5.20% |
+| NVIDIA GeForce RTX 3090 | 1024x1024 | 0.253 | 18.50% | 0.464 (+83%) | 5.20% |
+| NVIDIA GeForce RTX 3080 | 512x512 | 0.07 | 12.80% | 0.070 (0%) | 12.80% |
+| NVIDIA GeForce RTX 3080 | 1024x1024 | 0.286 | 45.30% | 0.466 (+63%) | 12.90% |
+| NVIDIA GeForce RTX 3070 | 512x512 | 0.102 | 15.90% | 0.102 (0%) | 15.90% |
+| NVIDIA GeForce RTX 3070 | 1024x1024 | 0.421 | 56.30% | 0.746 (+77%) | 16.00% |
+
+</details>
+
+<details><summary>SDXL</summary>
+
+| GPU | Resolution | Time (seconds) | Memory Consumed (%) | Tiled Time (seconds) | Tiled Memory (%) |
+| --- | --- | --- | --- | --- | --- |
+| NVIDIA GeForce RTX 4090 | 512x512 | 0.057 | 10.00% | 0.057 (0%) | 10.00% |
+| NVIDIA GeForce RTX 4090 | 1024x1024 | 0.256 | 35.50% | 0.257 (+0.4%) | 35.50% |
+| NVIDIA GeForce RTX 4080 | 512x512 | 0.092 | 15.00% | 0.092 (0%) | 15.00% |
+| NVIDIA GeForce RTX 4080 | 1024x1024 | 0.406 | 53.30% | 0.406 (0%) | 53.30% |
+| NVIDIA GeForce RTX 4070 Ti | 512x512 | 0.121 | 20.20% | 0.120 (-0.8%) | 20.20% |
+| NVIDIA GeForce RTX 4070 Ti | 1024x1024 | 0.519 | 72.00% | 0.519 (0%) | 72.00% |
+| NVIDIA GeForce RTX 3090 | 512x512 | 0.107 | 10.50% | 0.107 (0%) | 10.50% |
+| NVIDIA GeForce RTX 3090 | 1024x1024 | 0.459 | 38.00% | 0.460 (+0.2%) | 38.00% |
+| NVIDIA GeForce RTX 3080 | 512x512 | 0.121 | 25.60% | 0.121 (0%) | 25.60% |
+| NVIDIA GeForce RTX 3080 | 1024x1024 | 0.524 | 93.00% | 0.524 (0%) | 93.00% |
+| NVIDIA GeForce RTX 3070 | 512x512 | 0.183 | 31.80% | 0.183 (0%) | 31.80% |
+| NVIDIA GeForce RTX 3070 | 1024x1024 | 0.794 | 96.40% | 0.794 (0%) | 96.40% |
+
+</details>
+
+## Available VAEs
+
+|   | **Endpoint** | **Model** |
+|:-:|:-----------:|:--------:|
+| **Stable Diffusion v1** | [https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud](https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud) | [`stabilityai/sd-vae-ft-mse`](https://hf.co/stabilityai/sd-vae-ft-mse) |
+| **Stable Diffusion XL** | [https://x2dmsqunjd6k9prw.us-east-1.aws.endpoints.huggingface.cloud](https://x2dmsqunjd6k9prw.us-east-1.aws.endpoints.huggingface.cloud) | [`madebyollin/sdxl-vae-fp16-fix`](https://hf.co/madebyollin/sdxl-vae-fp16-fix) |
+| **Flux** | [https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud](https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud) | [`black-forest-labs/FLUX.1-schnell`](https://hf.co/black-forest-labs/FLUX.1-schnell) |
+| **HunyuanVideo** | [https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud](https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud) | [`hunyuanvideo-community/HunyuanVideo`](https://hf.co/hunyuanvideo-community/HunyuanVideo) |
+
+
+> [!TIP]
+> Model support can be requested [here](https://github.com/huggingface/diffusers/issues/new?template=remote-vae-pilot-feedback.yml).
+
+
+## Code
+
+> [!TIP]
+> Install `diffusers` from `main` to run the code: `pip install git+https://github.com/huggingface/diffusers@main`
+
+
+A helper method simplifies interacting with Hybrid Inference.
+
+```python
+from diffusers.utils.remote_utils import remote_decode
+```
+
+### Basic example
+
+Here, we show how to use the remote VAE on random tensors.
+
+<details><summary>Code</summary>
+
+```python
+image = remote_decode(
+    endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=torch.randn([1, 4, 64, 64], dtype=torch.float16),
+    scaling_factor=0.18215,
+)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/output.png"/>
+</figure>
+
+Usage for Flux is slightly different. Flux latents are packed so we need to send the `height` and `width`.
+
+<details><summary>Code</summary>
+
+```python
+image = remote_decode(
+    endpoint="https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=torch.randn([1, 4096, 64], dtype=torch.float16),
+    height=1024,
+    width=1024,
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/flux_random_latent.png"/>
+</figure>
+
+Finally, an example for HunyuanVideo.
+
+<details><summary>Code</summary>
+
+```python
+video = remote_decode(
+    endpoint="https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=torch.randn([1, 16, 3, 40, 64], dtype=torch.float16),
+    output_type="mp4",
+)
+with open("video.mp4", "wb") as f:
+    f.write(video)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+   <video
+      alt="queue.mp4"
+      autoplay loop autobuffer muted playsinline
+    >
+    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/video_1.mp4" type="video/mp4">
+  </video>
+</figure>
+
+
+### Generation
+
+But we want to use the VAE on an actual pipeline to get an actual image, not random noise. The example below shows how to do it with SD v1.5. 
+
+<details><summary>Code</summary>
+
+```python
+from diffusers import StableDiffusionPipeline
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    vae=None,
+).to("cuda")
+
+prompt = "Strawberry ice cream, in a stylish modern glass, coconut, splashing milk cream and honey, in a gradient purple background, fluid motion, dynamic movement, cinematic lighting, Mysterious"
+
+latent = pipe(
+    prompt=prompt,
+    output_type="latent",
+).images
+image = remote_decode(
+    endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    scaling_factor=0.18215,
+)
+image.save("test.jpg")
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/test.jpg"/>
+</figure>
+
+Here’s another example with Flux.
+
+<details><summary>Code</summary>
+
+```python
+from diffusers import FluxPipeline
+
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-schnell",
+    torch_dtype=torch.bfloat16,
+    vae=None,
+).to("cuda")
+
+prompt = "Strawberry ice cream, in a stylish modern glass, coconut, splashing milk cream and honey, in a gradient purple background, fluid motion, dynamic movement, cinematic lighting, Mysterious"
+
+latent = pipe(
+    prompt=prompt,
+    guidance_scale=0.0,
+    num_inference_steps=4,
+    output_type="latent",
+).images
+image = remote_decode(
+    endpoint="https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    height=1024,
+    width=1024,
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+image.save("test.jpg")
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/test_1.jpg"/>
+</figure>
+
+Here’s an example with HunyuanVideo.
+
+<details><summary>Code</summary>
+
+```python
+from diffusers import HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
+
+model_id = "hunyuanvideo-community/HunyuanVideo"
+transformer = HunyuanVideoTransformer3DModel.from_pretrained(
+    model_id, subfolder="transformer", torch_dtype=torch.bfloat16
+)
+pipe = HunyuanVideoPipeline.from_pretrained(
+    model_id, transformer=transformer, vae=None, torch_dtype=torch.float16
+).to("cuda")
+
+latent = pipe(
+    prompt="A cat walks on the grass, realistic",
+    height=320,
+    width=512,
+    num_frames=61,
+    num_inference_steps=30,
+    output_type="latent",
+).frames
+
+video = remote_decode(
+    endpoint="https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    output_type="mp4",
+)
+
+if isinstance(video, bytes):
+    with open("video.mp4", "wb") as f:
+        f.write(video)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+   <video
+      alt="queue.mp4"
+      autoplay loop autobuffer muted playsinline
+    >
+    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/video.mp4" type="video/mp4">
+  </video>
+</figure>
+
+
+### Queueing
+
+One of the great benefits of using a remote VAE is that we can queue multiple generation requests. While the current latent is being processed for decoding, we can already queue another one. This helps improve concurrency. 
+
+
+<details><summary>Code</summary>
+
+```python
+import queue
+import threading
+from IPython.display import display
+from diffusers import StableDiffusionPipeline
+
+def decode_worker(q: queue.Queue):
+    while True:
+        item = q.get()
+        if item is None:
+            break
+        image = remote_decode(
+            endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
+            tensor=item,
+            scaling_factor=0.18215,
+        )
+        display(image)
+        q.task_done()
+
+q = queue.Queue()
+thread = threading.Thread(target=decode_worker, args=(q,), daemon=True)
+thread.start()
+
+def decode(latent: torch.Tensor):
+    q.put(latent)
+
+prompts = [
+    "Blueberry ice cream, in a stylish modern glass , ice cubes, nuts, mint leaves, splashing milk cream, in a gradient purple background, fluid motion, dynamic movement, cinematic lighting, Mysterious",
+    "Lemonade in a glass, mint leaves, in an aqua and white background, flowers, ice cubes, halo, fluid motion, dynamic movement, soft lighting, digital painting, rule of thirds composition, Art by Greg rutkowski, Coby whitmore",
+    "Comic book art, beautiful, vintage, pastel neon colors, extremely detailed pupils, delicate features, light on face, slight smile, Artgerm, Mary Blair, Edmund Dulac, long dark locks, bangs, glowing, fashionable style, fairytale ambience, hot pink.",
+    "Masterpiece, vanilla cone ice cream garnished with chocolate syrup, crushed nuts, choco flakes, in a brown background, gold, cinematic lighting, Art by WLOP",
+    "A bowl of milk, falling cornflakes, berries, blueberries, in a white background, soft lighting, intricate details, rule of thirds, octane render, volumetric lighting",
+    "Cold Coffee with cream, crushed almonds, in a glass, choco flakes, ice cubes, wet, in a wooden background, cinematic lighting, hyper realistic painting, art by Carne Griffiths, octane render, volumetric lighting, fluid motion, dynamic movement, muted colors,",
+]
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "Lykon/dreamshaper-8",
+    torch_dtype=torch.float16,
+    vae=None,
+).to("cuda")
+
+pipe.unet = pipe.unet.to(memory_format=torch.channels_last)
+pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+
+_ = pipe(
+    prompt=prompts[0],
+    output_type="latent",
+)
+
+for prompt in prompts:
+    latent = pipe(
+        prompt=prompt,
+        output_type="latent",
+    ).images
+    decode(latent)
+
+q.put(None)
+thread.join()
+```
+
+</details>
+
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+   <video
+      alt="queue.mp4"
+      autoplay loop autobuffer muted playsinline
+    >
+    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/queue.mp4" type="video/mp4">
+  </video>
+</figure>
+
+## Integrations
+
+* **[SD.Next](https://github.com/vladmandic/sdnext):** All-in-one UI with direct supports Hybrid Inference.
+* **[ComfyUI-HFRemoteVae](https://github.com/kijai/ComfyUI-HFRemoteVae):** ComfyUI node for Hybrid Inference.
@@ -0,0 +1,183 @@
+# Getting Started: VAE Encode with Hybrid Inference
+
+VAE encode is used for training, image-to-image and image-to-video - turning into images or videos into latent representations.
+
+## Memory
+
+These tables demonstrate the VRAM requirements for VAE encode with SD v1 and SD XL on different GPUs.
+
+For the majority of these GPUs the memory usage % dictates other models (text encoders, UNet/Transformer) must be offloaded, or tiled encoding has to be used which increases time taken and impacts quality.
+
+<details><summary>SD v1.5</summary>
+
+| GPU                           | Resolution   |   Time (seconds) |   Memory (%) |   Tiled Time (secs) |   Tiled Memory (%) |
+|:------------------------------|:-------------|-----------------:|-------------:|--------------------:|-------------------:|
+| NVIDIA GeForce RTX 4090       | 512x512      |            0.015 |      3.51901 |               0.015 |            3.51901 |
+| NVIDIA GeForce RTX 4090       | 256x256      |            0.004 |      1.3154  |               0.005 |            1.3154  |
+| NVIDIA GeForce RTX 4090       | 2048x2048    |            0.402 |     47.1852  |               0.496 |            3.51901 |
+| NVIDIA GeForce RTX 4090       | 1024x1024    |            0.078 |     12.2658  |               0.094 |            3.51901 |
+| NVIDIA GeForce RTX 4080 SUPER | 512x512      |            0.023 |      5.30105 |               0.023 |            5.30105 |
+| NVIDIA GeForce RTX 4080 SUPER | 256x256      |            0.006 |      1.98152 |               0.006 |            1.98152 |
+| NVIDIA GeForce RTX 4080 SUPER | 2048x2048    |            0.574 |     71.08    |               0.656 |            5.30105 |
+| NVIDIA GeForce RTX 4080 SUPER | 1024x1024    |            0.111 |     18.4772  |               0.14  |            5.30105 |
+| NVIDIA GeForce RTX 3090       | 512x512      |            0.032 |      3.52782 |               0.032 |            3.52782 |
+| NVIDIA GeForce RTX 3090       | 256x256      |            0.01  |      1.31869 |               0.009 |            1.31869 |
+| NVIDIA GeForce RTX 3090       | 2048x2048    |            0.742 |     47.3033  |               0.954 |            3.52782 |
+| NVIDIA GeForce RTX 3090       | 1024x1024    |            0.136 |     12.2965  |               0.207 |            3.52782 |
+| NVIDIA GeForce RTX 3080       | 512x512      |            0.036 |      8.51761 |               0.036 |            8.51761 |
+| NVIDIA GeForce RTX 3080       | 256x256      |            0.01  |      3.18387 |               0.01  |            3.18387 |
+| NVIDIA GeForce RTX 3080       | 2048x2048    |            0.863 |     86.7424  |               1.191 |            8.51761 |
+| NVIDIA GeForce RTX 3080       | 1024x1024    |            0.157 |     29.6888  |               0.227 |            8.51761 |
+| NVIDIA GeForce RTX 3070       | 512x512      |            0.051 |     10.6941  |               0.051 |           10.6941  |
+| NVIDIA GeForce RTX 3070       | 256x256      |            0.015 |      3.99743 |               0.015 |            3.99743 |
+| NVIDIA GeForce RTX 3070       | 2048x2048    |            1.217 |     96.054   |               1.482 |           10.6941  |
+| NVIDIA GeForce RTX 3070       | 1024x1024    |            0.223 |     37.2751  |               0.327 |           10.6941  |
+
+
+</details>
+
+<details><summary>SDXL</summary>
+
+| GPU                           | Resolution   |   Time (seconds) |   Memory Consumed (%) |   Tiled Time (seconds) |   Tiled Memory (%) |
+|:------------------------------|:-------------|-----------------:|----------------------:|-----------------------:|-------------------:|
+| NVIDIA GeForce RTX 4090       | 512x512      |            0.029 |               4.95707 |                  0.029 |            4.95707 |
+| NVIDIA GeForce RTX 4090       | 256x256      |            0.007 |               2.29666 |                  0.007 |            2.29666 |
+| NVIDIA GeForce RTX 4090       | 2048x2048    |            0.873 |              66.3452  |                  0.863 |           15.5649  |
+| NVIDIA GeForce RTX 4090       | 1024x1024    |            0.142 |              15.5479  |                  0.143 |           15.5479  |
+| NVIDIA GeForce RTX 4080 SUPER | 512x512      |            0.044 |               7.46735 |                  0.044 |            7.46735 |
+| NVIDIA GeForce RTX 4080 SUPER | 256x256      |            0.01  |               3.4597  |                  0.01  |            3.4597  |
+| NVIDIA GeForce RTX 4080 SUPER | 2048x2048    |            1.317 |              87.1615  |                  1.291 |           23.447   |
+| NVIDIA GeForce RTX 4080 SUPER | 1024x1024    |            0.213 |              23.4215  |                  0.214 |           23.4215  |
+| NVIDIA GeForce RTX 3090       | 512x512      |            0.058 |               5.65638 |                  0.058 |            5.65638 |
+| NVIDIA GeForce RTX 3090       | 256x256      |            0.016 |               2.45081 |                  0.016 |            2.45081 |
+| NVIDIA GeForce RTX 3090       | 2048x2048    |            1.755 |              77.8239  |                  1.614 |           18.4193  |
+| NVIDIA GeForce RTX 3090       | 1024x1024    |            0.265 |              18.4023  |                  0.265 |           18.4023  |
+| NVIDIA GeForce RTX 3080       | 512x512      |            0.064 |              13.6568  |                  0.064 |           13.6568  |
+| NVIDIA GeForce RTX 3080       | 256x256      |            0.018 |               5.91728 |                  0.018 |            5.91728 |
+| NVIDIA GeForce RTX 3080       | 2048x2048    |          OOM     |             OOM       |                  1.866 |           44.4717  |
+| NVIDIA GeForce RTX 3080       | 1024x1024    |            0.302 |              44.4308  |                  0.302 |           44.4308  |
+| NVIDIA GeForce RTX 3070       | 512x512      |            0.093 |              17.1465  |                  0.093 |           17.1465  |
+| NVIDIA GeForce RTX 3070       | 256x256      |            0.025 |               7.42931 |                  0.026 |            7.42931 |
+| NVIDIA GeForce RTX 3070       | 2048x2048    |          OOM     |             OOM       |                  2.674 |           55.8355  |
+| NVIDIA GeForce RTX 3070       | 1024x1024    |            0.443 |              55.7841  |                  0.443 |           55.7841  |
+
+</details>
+
+## Available VAEs
+
+|   | **Endpoint** | **Model** |
+|:-:|:-----------:|:--------:|
+| **Stable Diffusion v1** | [https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud](https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud) | [`stabilityai/sd-vae-ft-mse`](https://hf.co/stabilityai/sd-vae-ft-mse) |
+| **Stable Diffusion XL** | [https://xjqqhmyn62rog84g.us-east-1.aws.endpoints.huggingface.cloud](https://xjqqhmyn62rog84g.us-east-1.aws.endpoints.huggingface.cloud) | [`madebyollin/sdxl-vae-fp16-fix`](https://hf.co/madebyollin/sdxl-vae-fp16-fix) |
+| **Flux** | [https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud](https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud) | [`black-forest-labs/FLUX.1-schnell`](https://hf.co/black-forest-labs/FLUX.1-schnell) |
+
+
+> [!TIP]
+> Model support can be requested [here](https://github.com/huggingface/diffusers/issues/new?template=remote-vae-pilot-feedback.yml).
+
+
+## Code
+
+> [!TIP]
+> Install `diffusers` from `main` to run the code: `pip install git+https://github.com/huggingface/diffusers@main`
+
+
+A helper method simplifies interacting with Hybrid Inference.
+
+```python
+from diffusers.utils.remote_utils import remote_encode
+```
+
+### Basic example
+
+Let's encode an image, then decode it to demonstrate.
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"/>
+</figure>
+
+<details><summary>Code</summary>
+
+```python
+from diffusers.utils import load_image
+from diffusers.utils.remote_utils import remote_decode
+
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg?download=true")
+
+latent = remote_encode(
+    endpoint="https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud/",
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+
+decoded = remote_decode(
+    endpoint="https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/decoded.png"/>
+</figure>
+
+
+### Generation
+
+Now let's look at a generation example, we'll encode the image, generate then remotely decode too!
+
+<details><summary>Code</summary>
+
+```python
+import torch
+from diffusers import StableDiffusionImg2ImgPipeline
+from diffusers.utils import load_image
+from diffusers.utils.remote_utils import remote_decode, remote_encode
+
+pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    vae=None,
+).to("cuda")
+
+init_image = load_image(
+    "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+)
+init_image = init_image.resize((768, 512))
+
+init_latent = remote_encode(
+    endpoint="https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud/",
+    image=init_image,
+    scaling_factor=0.18215,
+)
+
+prompt = "A fantasy landscape, trending on artstation"
+latent = pipe(
+    prompt=prompt,
+    image=init_latent,
+    strength=0.75,
+    output_type="latent",
+).images
+
+image = remote_decode(
+    endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    scaling_factor=0.18215,
+)
+image.save("fantasy_landscape.jpg")
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/fantasy_landscape.png"/>
+</figure>
+
+## Integrations
+
+* **[SD.Next](https://github.com/vladmandic/sdnext):** All-in-one UI with direct supports Hybrid Inference.
+* **[ComfyUI-HFRemoteVae](https://github.com/kijai/ComfyUI-HFRemoteVae):** ComfyUI node for Hybrid Inference.
@@ -161,10 +161,10 @@ Your Python environment will find the `main` version of 🤗 Diffusers on the ne

 Model weights and files are downloaded from the Hub to a cache which is usually your home directory. You can change the cache location by specifying the `HF_HOME` or `HUGGINFACE_HUB_CACHE` environment variables or configuring the `cache_dir` parameter in methods like [`~DiffusionPipeline.from_pretrained`].

-Cached files allow you to run 🤗 Diffusers offline. To prevent 🤗 Diffusers from connecting to the internet, set the `HF_HUB_OFFLINE` environment variable to `True` and 🤗 Diffusers will only load previously downloaded files in the cache.
+Cached files allow you to run 🤗 Diffusers offline. To prevent 🤗 Diffusers from connecting to the internet, set the `HF_HUB_OFFLINE` environment variable to `1` and 🤗 Diffusers will only load previously downloaded files in the cache.

 ```shell
-export HF_HUB_OFFLINE=True
+export HF_HUB_OFFLINE=1
 ```

 For more details about managing and cleaning the cache, take a look at the [caching](https://huggingface.co/docs/huggingface_hub/guides/manage-cache) guide.
@@ -179,14 +179,16 @@ Telemetry is only sent when loading models and pipelines from the Hub,
 and it is not collected if you're loading local files.

 We understand that not everyone wants to share additional information,and we respect your privacy.
-You can disable telemetry collection by setting the `DISABLE_TELEMETRY` environment variable from your terminal:
+You can disable telemetry collection by setting the `HF_HUB_DISABLE_TELEMETRY` environment variable from your terminal:

 On Linux/MacOS:
+
 ```bash
-export DISABLE_TELEMETRY=YES
+export HF_HUB_DISABLE_TELEMETRY=1
 ```

 On Windows:
+
 ```bash
-set DISABLE_TELEMETRY=YES
+set HF_HUB_DISABLE_TELEMETRY=1
 ```
@@ -178,6 +178,9 @@ pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch
 # We can utilize the enable_group_offload method for Diffusers model implementations
 pipe.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True)

+# Uncomment the following to also allow recording the current streams.
+# pipe.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True, record_stream=True)
+
 # For any other model implementations, the apply_group_offloading function can be used
 apply_group_offloading(pipe.text_encoder, onload_device=onload_device, offload_type="block_level", num_blocks_per_group=2)
 apply_group_offloading(pipe.vae, onload_device=onload_device, offload_type="leaf_level")
@@ -198,6 +201,19 @@ export_to_video(video, "output.mp4", fps=8)

 Group offloading (for CUDA devices with support for asynchronous data transfer streams) overlaps data transfer and computation to reduce the overall execution time compared to sequential offloading. This is enabled using layer prefetching with CUDA streams. The next layer to be executed is loaded onto the accelerator device while the current layer is being executed - this increases the memory requirements slightly. Group offloading also supports leaf-level offloading (equivalent to sequential CPU offloading) but can be made much faster when using streams.

+<Tip>
+
+- Group offloading may not work with all models out-of-the-box. If the forward implementations of the model contain weight-dependent device-casting of inputs, it may clash with the offloading mechanism's handling of device-casting.
+- The `offload_type` parameter can be set to either `block_level` or `leaf_level`. `block_level` offloads groups of `torch::nn::ModuleList` or `torch::nn:Sequential` modules based on a configurable attribute `num_blocks_per_group`. For example, if you set `num_blocks_per_group=2` on a standard transformer model containing 40 layers, it will onload/offload 2 layers at a time for a total of 20 onload/offloads. This drastically reduces the VRAM requirements. `leaf_level` offloads individual layers at the lowest level, which is equivalent to sequential offloading. However, unlike sequential offloading, group offloading can be made much faster when using streams, with minimal compromise to end-to-end generation time.
+- The `use_stream` parameter can be used with CUDA devices to enable prefetching layers for onload. It defaults to `False`. Layer prefetching allows overlapping computation and data transfer of model weights, which drastically reduces the overall execution time compared to other offloading methods. However, it can increase the CPU RAM usage significantly. Ensure that available CPU RAM that is at least twice the size of the model when setting `use_stream=True`. You can find more information about CUDA streams [here](https://pytorch.org/docs/stable/generated/torch.cuda.Stream.html)
+- If specifying `use_stream=True` on VAEs with tiling enabled, make sure to do a dummy forward pass (possibly with dummy inputs) before the actual inference to avoid device-mismatch errors. This may not work on all implementations. Please open an issue if you encounter any problems.
+- The parameter `low_cpu_mem_usage` can be set to `True` to reduce CPU memory usage when using streams for group offloading. This is useful when the CPU memory is the bottleneck, but it may counteract the benefits of using streams and increase the overall execution time. The CPU memory savings come from creating pinned-tensors on-the-fly instead of pre-pinning them. This parameter is better suited for using `leaf_level` offloading.
+- When using `use_stream=True`, users can additionally specify `record_stream=True` to get better speedups at the expense of slightly increased memory usage. Refer to the [official PyTorch docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html) to know more about this.
+
+For more information about available parameters and an explanation of how group offloading works, refer to [`~hooks.group_offloading.apply_group_offloading`].
+
+</Tip>
+
 ## FP8 layerwise weight-casting

 PyTorch supports `torch.float8_e4m3fn` and `torch.float8_e5m2` as weight storage dtypes, but they can't be used for computation in many different tensor operations due to unimplemented kernel support. However, you can use these dtypes to store model weights in fp8 precision and upcast them on-the-fly when the layers are used in the forward pass. This is known as layerwise weight-casting.
@@ -235,6 +251,14 @@ In the above example, layerwise casting is enabled on the transformer component

 However, you gain more control and flexibility by directly utilizing the [`~hooks.layerwise_casting.apply_layerwise_casting`] function instead of [`~ModelMixin.enable_layerwise_casting`].

+<Tip>
+
+- Layerwise casting may not work with all models out-of-the-box. Sometimes, the forward implementations of the model might contain internal typecasting of weight values. Such implementations are not supported due to the currently simplistic implementation of layerwise casting, which assumes that the forward pass is independent of the weight precision and that the input dtypes are always in `compute_dtype`. An example of an incompatible implementation can be found [here](https://github.com/huggingface/transformers/blob/7f5077e53682ca855afc826162b204ebf809f1f9/src/transformers/models/t5/modeling_t5.py#L294-L299).
+- Layerwise casting may fail on custom modeling implementations that make use of [PEFT](https://github.com/huggingface/peft) layers. Some minimal checks to handle this case is implemented but is not extensively tested or guaranteed to work in all cases.
+- It can be also be applied partially to specific layers of a model. Partially applying layerwise casting can either be done manually by calling the `apply_layerwise_casting` function on specific internal modules, or by specifying the `skip_modules_pattern` and `skip_modules_classes` parameters for a root module. These parameters are particularly useful for layers such as normalization and modulation.
+
+</Tip>
+
 ## Channels-last memory format

 The channels-last memory format is an alternative way of ordering NCHW tensors in memory to preserve dimension ordering. Channels-last tensors are ordered in such a way that the channels become the densest dimension (storing images pixel-per-pixel). Since not all operators currently support the channels-last format, it may result in worst performance but you should still try and see if it works for your model.
@@ -12,6 +12,9 @@ specific language governing permissions and limitations under the License.

 # Metal Performance Shaders (MPS)

+> [!TIP]
+> Pipelines with a <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22"> badge indicate a model can take advantage of the MPS backend on Apple silicon devices for faster inference. Feel free to open a [Pull Request](https://github.com/huggingface/diffusers/compare) to add this badge to pipelines that are missing it.
+
 🤗 Diffusers is compatible with Apple silicon (M1/M2 chips) using the PyTorch [`mps`](https://pytorch.org/docs/stable/notes/mps.html) device, which uses the Metal framework to leverage the GPU on MacOS devices. You'll need to have:

 - macOS computer with Apple silicon (M1/M2) hardware
@@ -37,7 +40,7 @@ image

 <Tip warning={true}>

-Generating multiple prompts in a batch can [crash](https://github.com/huggingface/diffusers/issues/363) or fail to work reliably. We believe this is related to the [`mps`](https://github.com/pytorch/pytorch/issues/84039) backend in PyTorch. While this is being investigated, you should iterate instead of batching.
+The PyTorch [mps](https://pytorch.org/docs/stable/notes/mps.html) backend does not support NDArray sizes greater than `2**32`. Please open an [Issue](https://github.com/huggingface/diffusers/issues/new/choose) if you encounter this problem so we can investigate.

 </Tip>

@@ -59,6 +62,10 @@ If you're using **PyTorch 1.13**, you need to "prime" the pipeline with an addit

 ## Troubleshoot

+This section lists some common issues with using the `mps` backend and how to solve them.
+
+### Attention slicing
+
 M1/M2 performance is very sensitive to memory pressure. When this occurs, the system automatically swaps if it needs to which significantly degrades performance.

 To prevent this from happening, we recommend *attention slicing* to reduce memory pressure during inference and prevent swapping. This is especially relevant if your computer has less than 64GB of system RAM, or if you generate images at non-standard resolutions larger than 512×512 pixels. Call the [`~DiffusionPipeline.enable_attention_slicing`] function on your pipeline:
@@ -72,3 +79,7 @@ pipeline.enable_attention_slicing()
 ```

 Attention slicing performs the costly attention operation in multiple steps instead of all at once. It usually improves performance by ~20% in computers without universal memory, but we've observed *better performance* in most Apple silicon computers unless you have 64GB of RAM or more.
+
+### Batch inference
+
+Generating multiple prompts in a batch can crash or fail to work reliably. If this is the case, try iterating instead of batching.
@@ -49,7 +49,7 @@ For Ada and higher-series GPUs. we recommend changing `torch_dtype` to `torch.bf
 from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
 from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig

-from diffusers import FluxTransformer2DModel
+from diffusers import AutoModel
 from transformers import T5EncoderModel

 quant_config = TransformersBitsAndBytesConfig(load_in_8bit=True,)
@@ -63,7 +63,7 @@ text_encoder_2_8bit = T5EncoderModel.from_pretrained(

 quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True,)

-transformer_8bit = FluxTransformer2DModel.from_pretrained(
+transformer_8bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
@@ -74,7 +74,7 @@ transformer_8bit = FluxTransformer2DModel.from_pretrained(
 By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter.

 ```diff
-transformer_8bit = FluxTransformer2DModel.from_pretrained(
+transformer_8bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
@@ -133,7 +133,7 @@ For Ada and higher-series GPUs. we recommend changing `torch_dtype` to `torch.bf
 from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
 from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig

-from diffusers import FluxTransformer2DModel
+from diffusers import AutoModel
 from transformers import T5EncoderModel

 quant_config = TransformersBitsAndBytesConfig(load_in_4bit=True,)
@@ -147,7 +147,7 @@ text_encoder_2_4bit = T5EncoderModel.from_pretrained(

 quant_config = DiffusersBitsAndBytesConfig(load_in_4bit=True,)

-transformer_4bit = FluxTransformer2DModel.from_pretrained(
+transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
@@ -158,7 +158,7 @@ transformer_4bit = FluxTransformer2DModel.from_pretrained(
 By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter.

 ```diff
-transformer_4bit = FluxTransformer2DModel.from_pretrained(
+transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
@@ -217,11 +217,11 @@ print(model.get_memory_footprint())
 Quantized models can be loaded from the [`~ModelMixin.from_pretrained`] method without needing to specify the `quantization_config` parameters:

 ```py
-from diffusers import FluxTransformer2DModel, BitsAndBytesConfig
+from diffusers import AutoModel, BitsAndBytesConfig

 quantization_config = BitsAndBytesConfig(load_in_4bit=True)

-model_4bit = FluxTransformer2DModel.from_pretrained(
+model_4bit = AutoModel.from_pretrained(
    "hf-internal-testing/flux.1-dev-nf4-pkg", subfolder="transformer"
 )
 ```
@@ -243,13 +243,13 @@ An "outlier" is a hidden state value greater than a certain threshold, and these
 To find the best threshold for your model, we recommend experimenting with the `llm_int8_threshold` parameter in [`BitsAndBytesConfig`]:

 ```py
-from diffusers import FluxTransformer2DModel, BitsAndBytesConfig
+from diffusers import AutoModel, BitsAndBytesConfig

 quantization_config = BitsAndBytesConfig(
    load_in_8bit=True, llm_int8_threshold=10,
 )

-model_8bit = FluxTransformer2DModel.from_pretrained(
+model_8bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quantization_config,
@@ -305,7 +305,7 @@ NF4 is a 4-bit data type from the [QLoRA](https://hf.co/papers/2305.14314) paper
 from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
 from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig

-from diffusers import FluxTransformer2DModel
+from diffusers import AutoModel
 from transformers import T5EncoderModel

 quant_config = TransformersBitsAndBytesConfig(
@@ -325,7 +325,7 @@ quant_config = DiffusersBitsAndBytesConfig(
    bnb_4bit_quant_type="nf4",
 )

-transformer_4bit = FluxTransformer2DModel.from_pretrained(
+transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
@@ -343,7 +343,7 @@ Nested quantization is a technique that can save additional memory at no additio
 from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
 from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig

-from diffusers import FluxTransformer2DModel
+from diffusers import AutoModel
 from transformers import T5EncoderModel

 quant_config = TransformersBitsAndBytesConfig(
@@ -363,7 +363,7 @@ quant_config = DiffusersBitsAndBytesConfig(
    bnb_4bit_use_double_quant=True,
 )

-transformer_4bit = FluxTransformer2DModel.from_pretrained(
+transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
@@ -379,7 +379,7 @@ Once quantized, you can dequantize a model to its original precision, but this m
 from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
 from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig

-from diffusers import FluxTransformer2DModel
+from diffusers import AutoModel
 from transformers import T5EncoderModel

 quant_config = TransformersBitsAndBytesConfig(
@@ -399,7 +399,7 @@ quant_config = DiffusersBitsAndBytesConfig(
    bnb_4bit_use_double_quant=True,
 )

-transformer_4bit = FluxTransformer2DModel.from_pretrained(
+transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
@@ -36,5 +36,6 @@ Diffusers currently supports the following quantization methods.
 - [BitsandBytes](./bitsandbytes)
 - [TorchAO](./torchao)
 - [GGUF](./gguf)
+- [Quanto](./quanto.md)

 [This resource](https://huggingface.co/docs/transformers/main/en/quantization/overview#when-to-use-what) provides a good overview of the pros and cons of different quantization techniques.
@@ -0,0 +1,148 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+-->
+
+# Quanto
+
+[Quanto](https://github.com/huggingface/optimum-quanto) is a PyTorch quantization backend for [Optimum](https://huggingface.co/docs/optimum/en/index). It has been designed with versatility and simplicity in mind:
+
+- All features are available in eager mode (works with non-traceable models)
+- Supports quantization aware training
+- Quantized models are compatible with `torch.compile`
+- Quantized models are Device agnostic (e.g CUDA,XPU,MPS,CPU)
+
+In order to use the Quanto backend, you will first need to install `optimum-quanto>=0.2.6` and `accelerate`
+
+```shell
+pip install optimum-quanto accelerate
+```
+
+Now you can quantize a model by passing the `QuantoConfig` object to the `from_pretrained()` method. Although the Quanto library does allow quantizing `nn.Conv2d` and `nn.LayerNorm` modules, currently, Diffusers only supports quantizing the weights in the `nn.Linear` layers of a model. The following snippet demonstrates how to apply `float8` quantization with Quanto.   
+
+```python
+import torch
+from diffusers import FluxTransformer2DModel, QuantoConfig
+
+model_id = "black-forest-labs/FLUX.1-dev"
+quantization_config = QuantoConfig(weights_dtype="float8")
+transformer = FluxTransformer2DModel.from_pretrained(
+      model_id,
+      subfolder="transformer",
+      quantization_config=quantization_config,
+      torch_dtype=torch.bfloat16,
+)
+
+pipe = FluxPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch_dtype)
+pipe.to("cuda")
+
+prompt = "A cat holding a sign that says hello world"
+image = pipe(
+    prompt, num_inference_steps=50, guidance_scale=4.5, max_sequence_length=512
+).images[0]
+image.save("output.png")
+```
+
+## Skipping Quantization on specific modules
+
+It is possible to skip applying quantization on certain modules using the `modules_to_not_convert` argument in the `QuantoConfig`. Please ensure that the modules passed in to this argument match the keys of the modules in the `state_dict`  
+
+```python
+import torch
+from diffusers import FluxTransformer2DModel, QuantoConfig
+
+model_id = "black-forest-labs/FLUX.1-dev"
+quantization_config = QuantoConfig(weights_dtype="float8", modules_to_not_convert=["proj_out"])
+transformer = FluxTransformer2DModel.from_pretrained(
+      model_id,
+      subfolder="transformer",
+      quantization_config=quantization_config,
+      torch_dtype=torch.bfloat16,
+)
+```
+
+## Using `from_single_file` with the Quanto Backend
+
+`QuantoConfig` is compatible with `~FromOriginalModelMixin.from_single_file`. 
+
+```python
+import torch
+from diffusers import FluxTransformer2DModel, QuantoConfig
+
+ckpt_path = "https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors"
+quantization_config = QuantoConfig(weights_dtype="float8")
+transformer = FluxTransformer2DModel.from_single_file(ckpt_path, quantization_config=quantization_config, torch_dtype=torch.bfloat16)
+```
+
+## Saving Quantized models
+
+Diffusers supports serializing Quanto models using the `~ModelMixin.save_pretrained` method.
+
+The serialization and loading requirements are different for models quantized directly with the Quanto library and models quantized
+with Diffusers using Quanto as the backend. It is currently not possible to load models quantized directly with Quanto into Diffusers using `~ModelMixin.from_pretrained`
+
+```python
+import torch
+from diffusers import FluxTransformer2DModel, QuantoConfig
+
+model_id = "black-forest-labs/FLUX.1-dev"
+quantization_config = QuantoConfig(weights_dtype="float8")
+transformer = FluxTransformer2DModel.from_pretrained(
+      model_id,
+      subfolder="transformer",
+      quantization_config=quantization_config,
+      torch_dtype=torch.bfloat16,
+)
+# save quantized model to reuse
+transformer.save_pretrained("<your quantized model save path>")
+
+# you can reload your quantized model with
+model = FluxTransformer2DModel.from_pretrained("<your quantized model save path>")
+```
+
+## Using `torch.compile` with Quanto
+
+Currently the Quanto backend supports `torch.compile` for the following quantization types:
+
+- `int8` weights 
+
+```python
+import torch
+from diffusers import FluxPipeline, FluxTransformer2DModel, QuantoConfig
+
+model_id = "black-forest-labs/FLUX.1-dev"
+quantization_config = QuantoConfig(weights_dtype="int8")
+transformer = FluxTransformer2DModel.from_pretrained(
+    model_id,
+    subfolder="transformer",
+    quantization_config=quantization_config,
+    torch_dtype=torch.bfloat16,
+)
+transformer = torch.compile(transformer, mode="max-autotune", fullgraph=True)
+
+pipe = FluxPipeline.from_pretrained(
+    model_id, transformer=transformer, torch_dtype=torch_dtype
+)
+pipe.to("cuda")
+images = pipe("A cat holding a sign that says hello").images[0]
+images.save("flux-quanto-compile.png")
+```
+
+## Supported Quantization Types
+
+### Weights
+
+- float8
+- int8
+- int4
+- int2
+
+
@@ -26,13 +26,13 @@ The example below only quantizes the weights to int8.

 ```python
 import torch
-from diffusers import FluxPipeline, FluxTransformer2DModel, TorchAoConfig
+from diffusers import FluxPipeline, AutoModel, TorchAoConfig

 model_id = "black-forest-labs/FLUX.1-dev"
 dtype = torch.bfloat16

 quantization_config = TorchAoConfig("int8wo")
-transformer = FluxTransformer2DModel.from_pretrained(
+transformer = AutoModel.from_pretrained(
    model_id,
    subfolder="transformer",
    quantization_config=quantization_config,
@@ -99,10 +99,10 @@ To serialize a quantized model in a given dtype, first load the model with the d

 ```python
 import torch
-from diffusers import FluxTransformer2DModel, TorchAoConfig
+from diffusers import AutoModel, TorchAoConfig

 quantization_config = TorchAoConfig("int8wo")
-transformer = FluxTransformer2DModel.from_pretrained(
+transformer = AutoModel.from_pretrained(
    "black-forest-labs/Flux.1-Dev",
    subfolder="transformer",
    quantization_config=quantization_config,
@@ -115,9 +115,9 @@ To load a serialized quantized model, use the [`~ModelMixin.from_pretrained`] me

 ```python
 import torch
-from diffusers import FluxPipeline, FluxTransformer2DModel
+from diffusers import FluxPipeline, AutoModel

-transformer = FluxTransformer2DModel.from_pretrained("/path/to/flux_int8wo", torch_dtype=torch.bfloat16, use_safetensors=False)
+transformer = AutoModel.from_pretrained("/path/to/flux_int8wo", torch_dtype=torch.bfloat16, use_safetensors=False)
 pipe = FluxPipeline.from_pretrained("black-forest-labs/Flux.1-Dev", transformer=transformer, torch_dtype=torch.bfloat16)
 pipe.to("cuda")

@@ -126,15 +126,15 @@ image = pipe(prompt, num_inference_steps=30, guidance_scale=7.0).images[0]
 image.save("output.png")
 ```

-Some quantization methods, such as `uint4wo`, cannot be loaded directly and may result in an `UnpicklingError` when trying to load the models, but work as expected when saving them. In order to work around this, one can load the state dict manually into the model. Note, however, that this requires using `weights_only=False` in `torch.load`, so it should be run only if the weights were obtained from a trustable source.
+If you are using `torch<=2.6.0`, some quantization methods, such as `uint4wo`, cannot be loaded directly and may result in an `UnpicklingError` when trying to load the models, but work as expected when saving them. In order to work around this, one can load the state dict manually into the model. Note, however, that this requires using `weights_only=False` in `torch.load`, so it should be run only if the weights were obtained from a trustable source.

 ```python
 import torch
 from accelerate import init_empty_weights
-from diffusers import FluxPipeline, FluxTransformer2DModel, TorchAoConfig
+from diffusers import FluxPipeline, AutoModel, TorchAoConfig

 # Serialize the model
-transformer = FluxTransformer2DModel.from_pretrained(
+transformer = AutoModel.from_pretrained(
    "black-forest-labs/Flux.1-Dev",
    subfolder="transformer",
    quantization_config=TorchAoConfig("uint4wo"),
@@ -146,10 +146,13 @@ transformer.save_pretrained("/path/to/flux_uint4wo", safe_serialization=False, m
 # Load the model
 state_dict = torch.load("/path/to/flux_uint4wo/diffusion_pytorch_model.bin", weights_only=False, map_location="cpu")
 with init_empty_weights():
-    transformer = FluxTransformer2DModel.from_config("/path/to/flux_uint4wo/config.json")
+    transformer = AutoModel.from_config("/path/to/flux_uint4wo/config.json")
 transformer.load_state_dict(state_dict, strict=True, assign=True)
 ```

+> [!TIP]
+> The [`AutoModel`] API is supported for PyTorch >= 2.6 as shown in the examples below.
+
 ## Resources

 - [TorchAO Quantization API](https://github.com/pytorch/ao/blob/main/torchao/quantization/README.md)
@@ -163,6 +163,9 @@ Models are initiated with the [`~ModelMixin.from_pretrained`] method which also
 >>> model = UNet2DModel.from_pretrained(repo_id, use_safetensors=True)
 ```

+> [!TIP]
+> Use the [`AutoModel`] API to automatically select a model class if you're unsure of which one to use.
+
 To access the model parameters, call `model.config`:

 ```py
@@ -31,10 +31,10 @@ To adapt your text-to-image model for inpainting, you'll need to change the numb
 Initialize a [`UNet2DConditionModel`] with the pretrained text-to-image model weights, and change `in_channels` to 9. Changing the number of `in_channels` means you need to set `ignore_mismatched_sizes=True` and `low_cpu_mem_usage=False` to avoid a size mismatch error because the shape is different now.

 ```py
-from diffusers import UNet2DConditionModel
+from diffusers import AutoModel

 model_id = "stable-diffusion-v1-5/stable-diffusion-v1-5"
-unet = UNet2DConditionModel.from_pretrained(
+unet = AutoModel.from_pretrained(
    model_id,
    subfolder="unet",
    in_channels=9,
@@ -165,10 +165,10 @@ flush()
 Load the diffusion transformer next which has 12.5B parameters. This time, set `device_map="auto"` to automatically distribute the model across two 16GB GPUs. The `auto` strategy is backed by [Accelerate](https://hf.co/docs/accelerate/index) and available as a part of the [Big Model Inference](https://hf.co/docs/accelerate/concept_guides/big_model_inference) feature. It starts by distributing a model across the fastest device first (GPU) before moving to slower devices like the CPU and hard drive if needed. The trade-off of storing model parameters on slower devices is slower inference latency.

 ```py
-from diffusers import FluxTransformer2DModel
+from diffusers import AutoModel
 import torch 

-transformer = FluxTransformer2DModel.from_pretrained(
+transformer = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev", 
    subfolder="transformer",
    device_map="auto",
@@ -32,9 +32,9 @@ The denoiser checkpoint can also have multiple shards and supports inference tha
 For example, let's save a sharded checkpoint for the [SDXL UNet](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main/unet):

 ```python
-from diffusers import UNet2DConditionModel
+from diffusers import AutoModel

-unet = UNet2DConditionModel.from_pretrained(
+unet = AutoModel.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet"
 )
 unet.save_pretrained("sdxl-unet-sharded", max_shard_size="5GB")
@@ -43,10 +43,10 @@ unet.save_pretrained("sdxl-unet-sharded", max_shard_size="5GB")
 The size of the fp32 variant of the SDXL UNet checkpoint is ~10.4GB. Set the `max_shard_size` parameter to 5GB to create 3 shards. After saving, you can load them in [`StableDiffusionXLPipeline`]:

 ```python
-from diffusers import UNet2DConditionModel, StableDiffusionXLPipeline
+from diffusers import AutoModel, StableDiffusionXLPipeline
 import torch

-unet = UNet2DConditionModel.from_pretrained(
+unet = AutoModel.from_pretrained(
    "sayakpaul/sdxl-unet-sharded", torch_dtype=torch.float16
 )
 pipeline = StableDiffusionXLPipeline.from_pretrained(
@@ -157,6 +157,84 @@ pipeline(
 )
 ```

+## IP Adapter Cutoff
+
+IP Adapter is an image prompt adapter that can be used for diffusion models without any changes to the underlying model. We can use the IP Adapter Cutoff Callback to disable the IP Adapter after a certain number of steps. To set up the callback, you need to specify the number of denoising steps after which the callback comes into effect. You can do so by using either one of these two arguments:
+
+- `cutoff_step_ratio`: Float number with the ratio of the steps.
+- `cutoff_step_index`: Integer number with the exact number of the step.
+
+We need to download the diffusion model and load the ip_adapter for it as follows:
+
+```py
+from diffusers import AutoPipelineForText2Image
+from diffusers.utils import load_image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name="ip-adapter_sdxl.bin")
+pipeline.set_ip_adapter_scale(0.6)
+```
+The setup for the callback should look something like this:
+
+```py
+
+from diffusers import AutoPipelineForText2Image
+from diffusers.callbacks import IPAdapterScaleCutoffCallback
+from diffusers.utils import load_image
+import torch
+ 
+
+pipeline = AutoPipelineForText2Image.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", 
+    torch_dtype=torch.float16
+).to("cuda")
+
+
+pipeline.load_ip_adapter(
+    "h94/IP-Adapter", 
+    subfolder="sdxl_models", 
+    weight_name="ip-adapter_sdxl.bin"
+)
+
+pipeline.set_ip_adapter_scale(0.6)
+
+
+callback = IPAdapterScaleCutoffCallback(
+    cutoff_step_ratio=None, 
+    cutoff_step_index=5
+)
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/ip_adapter_diner.png"
+)
+
+generator = torch.Generator(device="cuda").manual_seed(2628670641)
+
+images = pipeline(
+    prompt="a tiger sitting in a chair drinking orange juice",
+    ip_adapter_image=image,
+    negative_prompt="deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
+    generator=generator,
+    num_inference_steps=50,
+    callback_on_step_end=callback,
+).images
+
+images[0].save("custom_callback_img.png")
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/without_callback.png" alt="generated image of a tiger sitting in a chair drinking orange juice" />
+    <figcaption class="mt-2 text-center text-sm text-gray-500">without IPAdapterScaleCutoffCallback</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/with_callback2.png" alt="generated image of a tiger sitting in a chair drinking orange juice with ip adapter callback" />
+    <figcaption class="mt-2 text-center text-sm text-gray-500">with IPAdapterScaleCutoffCallback</figcaption>
+  </div>
+</div>
+
+
 ## Display image after each generation step

 > [!TIP]
@@ -95,6 +95,23 @@ Use the Space below to gauge a pipeline's memory requirements before you downloa
    ></iframe>
 </div>

+### Specifying Component-Specific Data Types
+
+You can customize the data types for individual sub-models by passing a dictionary to the `torch_dtype` parameter. This allows you to load different components of a pipeline in different floating point precisions. For instance, if you want to load the transformer with `torch.bfloat16` and all other components with `torch.float16`, you can pass a dictionary mapping:
+
+```python
+from diffusers import HunyuanVideoPipeline
+import torch
+
+pipe = HunyuanVideoPipeline.from_pretrained(
+    "hunyuanvideo-community/HunyuanVideo",
+    torch_dtype={"transformer": torch.bfloat16, "default": torch.float16},
+)
+print(pipe.transformer.dtype, pipe.vae.dtype)  # (torch.bfloat16, torch.float16)
+```
+
+If a component is not explicitly specified in the dictionary and no `default` is provided, it will be loaded with `torch.float32`.
+
 ### Local pipeline

 To load a pipeline locally, use [git-lfs](https://git-lfs.github.com/) to manually download a checkpoint to your local disk.
@@ -134,7 +134,7 @@ The [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] method loads L
 - the LoRA weights don't have separate identifiers for the UNet and text encoder
 - the LoRA weights have separate identifiers for the UNet and text encoder

-To directly load (and save) a LoRA adapter at the *model-level*, use [`~PeftAdapterMixin.load_lora_adapter`], which builds and prepares the necessary model configuration for the adapter. Like [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`], [`PeftAdapterMixin.load_lora_adapter`] can load LoRAs for both the UNet and text encoder. For example, if you're loading a LoRA for the UNet, [`PeftAdapterMixin.load_lora_adapter`] ignores the keys for the text encoder.
+To directly load (and save) a LoRA adapter at the *model-level*, use [`~loaders.PeftAdapterMixin.load_lora_adapter`], which builds and prepares the necessary model configuration for the adapter. Like [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`], [`~loaders.PeftAdapterMixin.load_lora_adapter`] can load LoRAs for both the UNet and text encoder. For example, if you're loading a LoRA for the UNet, [`~loaders.PeftAdapterMixin.load_lora_adapter`] ignores the keys for the text encoder.

 Use the `weight_name` parameter to specify the specific weight file and the `prefix` parameter to filter for the appropriate state dicts (`"unet"` in this case) to load.

@@ -155,7 +155,7 @@ image
    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_attn_proc.png" />
 </div>

-Save an adapter with [`~PeftAdapterMixin.save_lora_adapter`].
+Save an adapter with [`~loaders.PeftAdapterMixin.save_lora_adapter`].

 To unload the LoRA weights, use the [`~loaders.StableDiffusionLoraLoaderMixin.unload_lora_weights`] method to discard the LoRA weights and restore the model to its original weights:

@@ -194,6 +194,59 @@ Currently, [`~loaders.StableDiffusionLoraLoaderMixin.set_adapters`] only support

 </Tip>

+### Hotswapping LoRA adapters
+
+A common use case when serving multiple adapters is to load one adapter first, generate images, load another adapter, generate more images, load another adapter, etc. This workflow normally requires calling [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`], [`~loaders.StableDiffusionLoraLoaderMixin.set_adapters`], and possibly [`~loaders.peft.PeftAdapterMixin.delete_adapters`] to save memory. Moreover, if the model is compiled using `torch.compile`, performing these steps requires recompilation, which takes time.
+
+To better support this common workflow, you can "hotswap" a LoRA adapter, to avoid accumulating memory and in some cases, recompilation. It requires an adapter to already be loaded, and the new adapter weights are swapped in-place for the existing adapter.
+
+Pass `hotswap=True` when loading a LoRA adapter to enable this feature. It is important to indicate the name of the existing adapter, (`default_0` is the default adapter name), to be swapped. If you loaded the first adapter with a different name, use that name instead.
+
+```python
+pipe = ...
+# load adapter 1 as normal
+pipeline.load_lora_weights(file_name_adapter_1)
+# generate some images with adapter 1
+...
+# now hot swap the 2nd adapter
+pipeline.load_lora_weights(file_name_adapter_2, hotswap=True, adapter_name="default_0")
+# generate images with adapter 2
+```
+
+
+<Tip warning={true}>
+
+Hotswapping is not currently supported for LoRA adapters that target the text encoder.
+
+</Tip>
+
+For compiled models, it is often (though not always if the second adapter targets identical LoRA ranks and scales) necessary to call [`~loaders.lora_base.LoraBaseMixin.enable_lora_hotswap`] to avoid recompilation. Use [`~loaders.lora_base.LoraBaseMixin.enable_lora_hotswap`] _before_ loading the first adapter, and `torch.compile` should be called _after_ loading the first adapter.
+
+```python
+pipe = ...
+# call this extra method
+pipe.enable_lora_hotswap(target_rank=max_rank)
+# now load adapter 1
+pipe.load_lora_weights(file_name_adapter_1)
+# now compile the unet of the pipeline
+pipe.unet = torch.compile(pipeline.unet, ...)
+# generate some images with adapter 1
+...
+# now hot swap adapter 2
+pipeline.load_lora_weights(file_name_adapter_2, hotswap=True, adapter_name="default_0")
+# generate images with adapter 2
+```
+
+The `target_rank=max_rank` argument is important for setting the maximum rank among all LoRA adapters that will be loaded. If you have one adapter with rank 8 and another with rank 16, pass `target_rank=16`. You should use a higher value if in doubt. By default, this value is 128.
+
+However, there can be situations where recompilation is unavoidable. For example, if the hotswapped adapter targets more layers than the initial adapter, then recompilation is triggered. Try to load the adapter that targets the most layers first. Refer to the PEFT docs on [hotswapping](https://huggingface.co/docs/peft/main/en/package_reference/hotswap#peft.utils.hotswap.hotswap_adapter) for more details about the limitations of this feature.
+
+<Tip>
+
+Move your code inside the `with torch._dynamo.config.patch(error_on_recompile=True)` context manager to detect if a model was recompiled. If you detect recompilation despite following all the steps above, please open an issue with [Diffusers](https://github.com/huggingface/diffusers/issues) with a reproducible example.
+
+</Tip>
+
 ### Kohya and TheLastBen

 Other popular LoRA trainers from the community include those by [Kohya](https://github.com/kohya-ss/sd-scripts/) and [TheLastBen](https://github.com/TheLastBen/fast-stable-diffusion). These trainers create different LoRA checkpoints than those trained by 🤗 Diffusers, but they can still be loaded in the same way.
@@ -66,10 +66,10 @@ Let's dive deeper into what these steps entail.
 1. Load a UNet that corresponds to the UNet in the LoRA checkpoint. In this case, both LoRAs use the SDXL UNet as their base model.

 ```python
-from diffusers import UNet2DConditionModel
+from diffusers import AutoModel
 import torch

-unet = UNet2DConditionModel.from_pretrained(
+unet = AutoModel.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    use_safetensors=True,
@@ -136,7 +136,7 @@ feng_peft_model.load_state_dict(original_state_dict, strict=True)
 ```python
 from peft import PeftModel

-base_unet = UNet2DConditionModel.from_pretrained(
+base_unet = AutoModel.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    use_safetensors=True,
@@ -66,12 +66,6 @@ from accelerate.utils import write_basic_config
 write_basic_config()
 ```

-## 원을 채우는 데이터셋
-
-원본 데이터셋은 ControlNet [repo](https://huggingface.co/lllyasviel/ControlNet/blob/main/training/fill50k.zip)에 올라와있지만, 우리는 [여기](https://huggingface.co/datasets/fusing/fill50k)에 새롭게 다시 올려서 🤗 Datasets 과 호환가능합니다. 그래서 학습 스크립트 상에서 데이터 불러오기를 다룰 수 있습니다.
-
-우리의 학습 예시는 원래 ControlNet의 학습에 쓰였던 [`stable-diffusion-v1-5/stable-diffusion-v1-5`](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5)을 사용합니다. 그렇지만 ControlNet은 대응되는 어느 Stable Diffusion 모델([`CompVis/stable-diffusion-v1-4`](https://huggingface.co/CompVis/stable-diffusion-v1-4)) 혹은 [`stabilityai/stable-diffusion-2-1`](https://huggingface.co/stabilityai/stable-diffusion-2-1)의 증가를 위해 학습될 수 있습니다.
-
 자체 데이터셋을 사용하기 위해서는 [학습을 위한 데이터셋 생성하기](create_dataset) 가이드를 확인하세요.

 ## 학습
@@ -79,13 +79,13 @@ This command will prompt you for a token. Copy-paste yours from your [settings/t
 ### Target Modules
 When LoRA was first adapted from language models to diffusion models, it was applied to the cross-attention layers in the Unet that relate the image representations with the prompts that describe them. 
 More recently, SOTA text-to-image diffusion models replaced the Unet with a diffusion Transformer(DiT). With this change, we may also want to explore 
-applying LoRA training onto different types of layers and blocks. To allow more flexibility and control over the targeted modules we added `--lora_layers`- in which you can specify in a comma seperated string
+applying LoRA training onto different types of layers and blocks. To allow more flexibility and control over the targeted modules we added `--lora_layers`- in which you can specify in a comma separated string
 the exact modules for LoRA training. Here are some examples of target modules you can provide: 
 - for attention only layers: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0"`
 - to train the same modules as in the fal trainer: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2"`
 - to train the same modules as in ostris ai-toolkit / replicate trainer: `--lora_blocks="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2,norm1_context.linear, norm1.linear,norm.linear,proj_mlp,proj_out"`
 > [!NOTE]
-> `--lora_layers` can also be used to specify which **blocks** to apply LoRA training to. To do so, simply add a block prefix to each layer in the comma seperated string:
+> `--lora_layers` can also be used to specify which **blocks** to apply LoRA training to. To do so, simply add a block prefix to each layer in the comma separated string:
 > **single DiT blocks**: to target the ith single transformer block, add the prefix `single_transformer_blocks.i`, e.g. - `single_transformer_blocks.i.attn.to_k`
 > **MMDiT blocks**: to target the ith MMDiT block, add the prefix `transformer_blocks.i`, e.g. - `transformer_blocks.i.attn.to_k` 
 > [!NOTE]
@@ -1,7 +1,8 @@
-accelerate>=0.16.0
+accelerate>=0.31.0
 torchvision
-transformers>=4.25.1
+transformers>=4.41.2
 ftfy
 tensorboard
 Jinja2
-peft==0.7.0
+peft>=0.11.1
+sentencepiece
@@ -24,7 +24,7 @@ import re
 import shutil
 from contextlib import nullcontext
 from pathlib import Path
-from typing import List, Optional, Union
+from typing import List, Optional

 import numpy as np
 import torch
@@ -74,7 +74,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -227,11 +227,21 @@ def log_validation(
    pipeline.set_progress_bar_config(disable=True)

    # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
-    autocast_ctx = nullcontext()
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
+    autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()

-    with autocast_ctx:
-        images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
+    # pre-calculate  prompt embeds, pooled prompt embeds, text ids because t5 does not support autocast
+    with torch.no_grad():
+        prompt_embeds, pooled_prompt_embeds, text_ids = pipeline.encode_prompt(
+            pipeline_args["prompt"], prompt_2=pipeline_args["prompt"]
+        )
+    images = []
+    for _ in range(args.num_validation_images):
+        with autocast_ctx:
+            image = pipeline(
+                prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds, generator=generator
+            ).images[0]
+            images.append(image)

    for tracker in accelerator.trackers:
        phase_name = "test" if is_final_validation else "validation"
@@ -378,7 +388,7 @@ def parse_args(input_args=None):
        default=None,
        help="the concept to use to initialize the new inserted tokens when training with "
        "--train_text_encoder_ti = True. By default, new tokens (<si><si+1>) are initialized with random value. "
-        "Alternatively, you could specify a different word/words whos value will be used as the starting point for the new inserted tokens. "
+        "Alternatively, you could specify a different word/words whose value will be used as the starting point for the new inserted tokens. "
        "--num_new_tokens_per_abstraction is ignored when initializer_concept is provided",
    )
    parser.add_argument(
@@ -657,15 +667,17 @@ def parse_args(input_args=None):
    parser.add_argument(
        "--adam_weight_decay_text_encoder", type=float, default=1e-03, help="Weight decay to use for text_encoder"
    )
+
    parser.add_argument(
        "--lora_layers",
        type=str,
        default=None,
        help=(
-            "The transformer modules to apply LoRA training on. Please specify the layers in a comma seperated. "
+            "The transformer modules to apply LoRA training on. Please specify the layers in a comma separated. "
            'E.g. - "to_k,to_q,to_v,to_out.0" will result in lora training of attention layers only. For more examples refer to https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/README_flux.md'
        ),
    )
+
    parser.add_argument(
        "--adam_epsilon",
        type=float,
@@ -738,6 +750,15 @@ def parse_args(input_args=None):
            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
        ),
    )
+    parser.add_argument(
+        "--upcast_before_saving",
+        action="store_true",
+        default=False,
+        help=(
+            "Whether to upcast the trained transformer layers to float32 before saving (at the end of training). "
+            "Defaults to precision dtype used for training to save memory"
+        ),
+    )
    parser.add_argument(
        "--prior_generation_precision",
        type=str,
@@ -818,9 +839,9 @@ class TokenEmbeddingsHandler:
        idx = 0
        for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
            assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )

            self.inserting_toks = inserting_toks
            special_tokens_dict = {"additional_special_tokens": self.inserting_toks}
@@ -880,9 +901,7 @@ class TokenEmbeddingsHandler:
        idx_to_text_encoder_name = {0: "clip_l", 1: "t5"}
        for idx, text_encoder in enumerate(self.text_encoders):
            train_ids = self.train_ids if idx == 0 else self.train_ids_t5
-            embeds = (
-                text_encoder.text_model.embeddings.token_embedding if idx == 0 else text_encoder.encoder.embed_tokens
-            )
+            embeds = text_encoder.text_model.embeddings.token_embedding if idx == 0 else text_encoder.shared
            assert embeds.weight.data.shape[0] == len(self.tokenizers[idx]), "Tokenizers should be the same."
            new_token_embeddings = embeds.weight.data[train_ids]

@@ -904,9 +923,7 @@ class TokenEmbeddingsHandler:
    @torch.no_grad()
    def retract_embeddings(self):
        for idx, text_encoder in enumerate(self.text_encoders):
-            embeds = (
-                text_encoder.text_model.embeddings.token_embedding if idx == 0 else text_encoder.encoder.embed_tokens
-            )
+            embeds = text_encoder.text_model.embeddings.token_embedding if idx == 0 else text_encoder.shared
            index_no_updates = self.embeddings_settings[f"index_no_updates_{idx}"]
            embeds.weight.data[index_no_updates] = (
                self.embeddings_settings[f"original_embeddings_{idx}"][index_no_updates]
@@ -1151,7 +1168,7 @@ def tokenize_prompt(tokenizer, prompt, max_sequence_length, add_special_tokens=F
    return text_input_ids


-def _get_t5_prompt_embeds(
+def _encode_prompt_with_t5(
    text_encoder,
    tokenizer,
    max_sequence_length=512,
@@ -1180,7 +1197,10 @@ def _get_t5_prompt_embeds(

    prompt_embeds = text_encoder(text_input_ids.to(device))[0]

-    dtype = text_encoder.dtype
+    if hasattr(text_encoder, "module"):
+        dtype = text_encoder.module.dtype
+    else:
+        dtype = text_encoder.dtype
    prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)

    _, seq_len, _ = prompt_embeds.shape
@@ -1192,7 +1212,7 @@ def _get_t5_prompt_embeds(
    return prompt_embeds


-def _get_clip_prompt_embeds(
+def _encode_prompt_with_clip(
    text_encoder,
    tokenizer,
    prompt: str,
@@ -1221,9 +1241,13 @@ def _get_clip_prompt_embeds(

    prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=False)

+    if hasattr(text_encoder, "module"):
+        dtype = text_encoder.module.dtype
+    else:
+        dtype = text_encoder.dtype
    # Use pooled output of CLIPTextModel
    prompt_embeds = prompt_embeds.pooler_output
-    prompt_embeds = prompt_embeds.to(dtype=text_encoder.dtype, device=device)
+    prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)

    # duplicate text embeddings for each generation per prompt, using mps friendly method
    prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
@@ -1242,136 +1266,35 @@ def encode_prompt(
    text_input_ids_list=None,
 ):
    prompt = [prompt] if isinstance(prompt, str) else prompt
-    batch_size = len(prompt)
-    dtype = text_encoders[0].dtype
+    if hasattr(text_encoders[0], "module"):
+        dtype = text_encoders[0].module.dtype
+    else:
+        dtype = text_encoders[0].dtype

-    pooled_prompt_embeds = _get_clip_prompt_embeds(
+    pooled_prompt_embeds = _encode_prompt_with_clip(
        text_encoder=text_encoders[0],
        tokenizer=tokenizers[0],
        prompt=prompt,
        device=device if device is not None else text_encoders[0].device,
        num_images_per_prompt=num_images_per_prompt,
-        text_input_ids=text_input_ids_list[0] if text_input_ids_list is not None else None,
+        text_input_ids=text_input_ids_list[0] if text_input_ids_list else None,
    )

-    prompt_embeds = _get_t5_prompt_embeds(
+    prompt_embeds = _encode_prompt_with_t5(
        text_encoder=text_encoders[1],
        tokenizer=tokenizers[1],
        max_sequence_length=max_sequence_length,
        prompt=prompt,
        num_images_per_prompt=num_images_per_prompt,
        device=device if device is not None else text_encoders[1].device,
-        text_input_ids=text_input_ids_list[1] if text_input_ids_list is not None else None,
+        text_input_ids=text_input_ids_list[1] if text_input_ids_list else None,
    )

-    text_ids = torch.zeros(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
-    text_ids = text_ids.repeat(num_images_per_prompt, 1, 1)
+    text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)

    return prompt_embeds, pooled_prompt_embeds, text_ids


-# CustomFlowMatchEulerDiscreteScheduler was taken from ostris ai-toolkit trainer:
-# https://github.com/ostris/ai-toolkit/blob/9ee1ef2a0a2a9a02b92d114a95f21312e5906e54/toolkit/samplers/custom_flowmatch_sampler.py#L95
-class CustomFlowMatchEulerDiscreteScheduler(FlowMatchEulerDiscreteScheduler):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        with torch.no_grad():
-            # create weights for timesteps
-            num_timesteps = 1000
-
-            # generate the multiplier based on cosmap loss weighing
-            # this is only used on linear timesteps for now
-
-            # cosine map weighing is higher in the middle and lower at the ends
-            # bot = 1 - 2 * self.sigmas + 2 * self.sigmas ** 2
-            # cosmap_weighing = 2 / (math.pi * bot)
-
-            # sigma sqrt weighing is significantly higher at the end and lower at the beginning
-            sigma_sqrt_weighing = (self.sigmas**-2.0).float()
-            # clip at 1e4 (1e6 is too high)
-            sigma_sqrt_weighing = torch.clamp(sigma_sqrt_weighing, max=1e4)
-            # bring to a mean of 1
-            sigma_sqrt_weighing = sigma_sqrt_weighing / sigma_sqrt_weighing.mean()
-
-            # Create linear timesteps from 1000 to 0
-            timesteps = torch.linspace(1000, 0, num_timesteps, device="cpu")
-
-            self.linear_timesteps = timesteps
-            # self.linear_timesteps_weights = cosmap_weighing
-            self.linear_timesteps_weights = sigma_sqrt_weighing
-
-            # self.sigmas = self.get_sigmas(timesteps, n_dim=1, dtype=torch.float32, device='cpu')
-            pass
-
-    def get_weights_for_timesteps(self, timesteps: torch.Tensor) -> torch.Tensor:
-        # Get the indices of the timesteps
-        step_indices = [(self.timesteps == t).nonzero().item() for t in timesteps]
-
-        # Get the weights for the timesteps
-        weights = self.linear_timesteps_weights[step_indices].flatten()
-
-        return weights
-
-    def get_sigmas(self, timesteps: torch.Tensor, n_dim, dtype, device) -> torch.Tensor:
-        sigmas = self.sigmas.to(device=device, dtype=dtype)
-        schedule_timesteps = self.timesteps.to(device)
-        timesteps = timesteps.to(device)
-        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
-
-        sigma = sigmas[step_indices].flatten()
-        while len(sigma.shape) < n_dim:
-            sigma = sigma.unsqueeze(-1)
-
-        return sigma
-
-    def add_noise(
-        self,
-        original_samples: torch.Tensor,
-        noise: torch.Tensor,
-        timesteps: torch.Tensor,
-    ) -> torch.Tensor:
-        ## ref https://github.com/huggingface/diffusers/blob/fbe29c62984c33c6cf9cf7ad120a992fe6d20854/examples/dreambooth/train_dreambooth_sd3.py#L1578
-        ## Add noise according to flow matching.
-        ## zt = (1 - texp) * x + texp * z1
-
-        # sigmas = get_sigmas(timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
-        # noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise
-
-        # timestep needs to be in [0, 1], we store them in [0, 1000]
-        # noisy_sample = (1 - timestep) * latent + timestep * noise
-        t_01 = (timesteps / 1000).to(original_samples.device)
-        noisy_model_input = (1 - t_01) * original_samples + t_01 * noise
-
-        # n_dim = original_samples.ndim
-        # sigmas = self.get_sigmas(timesteps, n_dim, original_samples.dtype, original_samples.device)
-        # noisy_model_input = (1.0 - sigmas) * original_samples + sigmas * noise
-        return noisy_model_input
-
-    def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
-        return sample
-
-    def set_train_timesteps(self, num_timesteps, device, linear=False):
-        if linear:
-            timesteps = torch.linspace(1000, 0, num_timesteps, device=device)
-            self.timesteps = timesteps
-            return timesteps
-        else:
-            # distribute them closer to center. Inference distributes them as a bias toward first
-            # Generate values from 0 to 1
-            t = torch.sigmoid(torch.randn((num_timesteps,), device=device))
-
-            # Scale and reverse the values to go from 1000 to 0
-            timesteps = (1 - t) * 1000
-
-            # Sort the timesteps in descending order
-            timesteps, _ = torch.sort(timesteps, descending=True)
-
-            self.timesteps = timesteps.to(device=device)
-
-            return timesteps
-
-
 def main(args):
    if args.report_to == "wandb" and args.hub_token is not None:
        raise ValueError(
@@ -1503,7 +1426,7 @@ def main(args):
    )

    # Load scheduler and models
-    noise_scheduler = CustomFlowMatchEulerDiscreteScheduler.from_pretrained(
+    noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="scheduler"
    )
    noise_scheduler_copy = copy.deepcopy(noise_scheduler)
@@ -1623,7 +1546,6 @@ def main(args):
        target_modules=target_modules,
    )
    transformer.add_adapter(transformer_lora_config)
-
    if args.train_text_encoder:
        text_lora_config = LoraConfig(
            r=args.rank,
@@ -1683,7 +1605,7 @@ def main(args):
        lora_state_dict = FluxPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1731,7 +1653,6 @@ def main(args):
        cast_training_params(models, dtype=torch.float32)

    transformer_lora_parameters = list(filter(lambda p: p.requires_grad, transformer.parameters()))
-
    if args.train_text_encoder:
        text_lora_parameters_one = list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
    # if we use textual inversion, we freeze all parameters except for the token embeddings
@@ -1741,7 +1662,8 @@ def main(args):
        for name, param in text_encoder_one.named_parameters():
            if "token_embedding" in name:
                # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-                param.data = param.to(dtype=torch.float32)
+                if args.mixed_precision == "fp16":
+                    param.data = param.to(dtype=torch.float32)
                param.requires_grad = True
                text_lora_parameters_one.append(param)
            else:
@@ -1749,9 +1671,10 @@ def main(args):
        if args.enable_t5_ti:  # whether to do pivotal tuning/textual inversion for T5 as well
            text_lora_parameters_two = []
            for name, param in text_encoder_two.named_parameters():
-                if "token_embedding" in name:
+                if "shared" in name:
                    # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-                    param.data = param.to(dtype=torch.float32)
+                    if args.mixed_precision == "fp16":
+                        param.data = param.to(dtype=torch.float32)
                    param.requires_grad = True
                    text_lora_parameters_two.append(param)
                else:
@@ -1832,6 +1755,7 @@ def main(args):
            optimizer_class = bnb.optim.AdamW8bit
        else:
            optimizer_class = torch.optim.AdamW
+
        optimizer = optimizer_class(
            params_to_optimize,
            betas=(args.adam_beta1, args.adam_beta2),
@@ -2025,6 +1949,7 @@ def main(args):
                lr_scheduler,
            )
        else:
+            print("I SHOULD BE HERE")
            transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
                transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler
            )
@@ -2129,7 +2054,7 @@ def main(args):
                if args.train_text_encoder:
                    text_encoder_one.train()
                    # set top parameter requires_grad = True for gradient checkpointing works
-                    accelerator.unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
+                    unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
                elif args.train_text_encoder_ti:  # textual inversion / pivotal tuning
                    text_encoder_one.train()
                if args.enable_t5_ti:
@@ -2141,6 +2066,11 @@ def main(args):
                pivoted_tr = True

        for step, batch in enumerate(train_dataloader):
+            models_to_accumulate = [transformer]
+            if not freeze_text_encoder:
+                models_to_accumulate.extend([text_encoder_one])
+                if args.enable_t5_ti:
+                    models_to_accumulate.extend([text_encoder_two])
            if pivoted_te:
                # stopping optimization of text_encoder params
                optimizer.param_groups[te_idx]["lr"] = 0.0
@@ -2149,7 +2079,7 @@ def main(args):
                logger.info(f"PIVOT TRANSFORMER {epoch}")
                optimizer.param_groups[0]["lr"] = 0.0

-            with accelerator.accumulate(transformer):
+            with accelerator.accumulate(models_to_accumulate):
                prompts = batch["prompts"]

                # encode batch prompts when custom prompts are provided for each image -
@@ -2193,7 +2123,7 @@ def main(args):
                model_input = (model_input - vae_config_shift_factor) * vae_config_scaling_factor
                model_input = model_input.to(dtype=weight_dtype)

-                vae_scale_factor = 2 ** (len(vae_config_block_out_channels))
+                vae_scale_factor = 2 ** (len(vae_config_block_out_channels) - 1)

                latent_image_ids = FluxPipeline._prepare_latent_image_ids(
                    model_input.shape[0],
@@ -2232,7 +2162,7 @@ def main(args):
                )

                # handle guidance
-                if transformer.config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                    guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                    guidance = guidance.expand(model_input.shape[0])
                else:
@@ -2292,16 +2222,26 @@ def main(args):
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    if not freeze_text_encoder:
-                        if args.train_text_encoder:
+                        if args.train_text_encoder:  # text encoder tuning
                            params_to_clip = itertools.chain(transformer.parameters(), text_encoder_one.parameters())
                        elif pure_textual_inversion:
-                            params_to_clip = itertools.chain(
-                                text_encoder_one.parameters(), text_encoder_two.parameters()
-                            )
+                            if args.enable_t5_ti:
+                                params_to_clip = itertools.chain(
+                                    text_encoder_one.parameters(), text_encoder_two.parameters()
+                                )
+                            else:
+                                params_to_clip = itertools.chain(text_encoder_one.parameters())
                        else:
-                            params_to_clip = itertools.chain(
-                                transformer.parameters(), text_encoder_one.parameters(), text_encoder_two.parameters()
-                            )
+                            if args.enable_t5_ti:
+                                params_to_clip = itertools.chain(
+                                    transformer.parameters(),
+                                    text_encoder_one.parameters(),
+                                    text_encoder_two.parameters(),
+                                )
+                            else:
+                                params_to_clip = itertools.chain(
+                                    transformer.parameters(), text_encoder_one.parameters()
+                                )
                    else:
                        params_to_clip = itertools.chain(transformer.parameters())
                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
@@ -2343,6 +2283,10 @@ def main(args):

                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                        accelerator.save_state(save_path)
+                        if args.train_text_encoder_ti:
+                            embedding_handler.save_embeddings(
+                                f"{args.output_dir}/{Path(args.output_dir).name}_emb_checkpoint_{global_step}.safetensors"
+                            )
                        logger.info(f"Saved state to {save_path}")

            logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
@@ -2355,14 +2299,16 @@ def main(args):
        if accelerator.is_main_process:
            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                # create pipeline
-                if freeze_text_encoder:
+                if freeze_text_encoder:  # no text encoder one, two optimizations
                    text_encoder_one, text_encoder_two = load_text_encoders(text_encoder_cls_one, text_encoder_cls_two)
+                    text_encoder_one.to(weight_dtype)
+                    text_encoder_two.to(weight_dtype)
                pipeline = FluxPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
-                    transformer=accelerator.unwrap_model(transformer),
+                    text_encoder=unwrap_model(text_encoder_one),
+                    text_encoder_2=unwrap_model(text_encoder_two),
+                    transformer=unwrap_model(transformer),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
@@ -2376,21 +2322,21 @@ def main(args):
                    epoch=epoch,
                    torch_dtype=weight_dtype,
                )
-                images = None
-                del pipeline
-
                if freeze_text_encoder:
                    del text_encoder_one, text_encoder_two
                    free_memory()
-                elif args.train_text_encoder:
-                    del text_encoder_two
-                    free_memory()
+
+                images = None
+                del pipeline

    # Save the lora layers
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        transformer = unwrap_model(transformer)
-        transformer = transformer.to(weight_dtype)
+        if args.upcast_before_saving:
+            transformer.to(torch.float32)
+        else:
+            transformer = transformer.to(weight_dtype)
        transformer_lora_layers = get_peft_model_state_dict(transformer)

        if args.train_text_encoder:
@@ -2432,8 +2378,8 @@ def main(args):
                accelerator=accelerator,
                pipeline_args=pipeline_args,
                epoch=epoch,
-                torch_dtype=weight_dtype,
                is_final_validation=True,
+                torch_dtype=weight_dtype,
            )

        save_model_card(
@@ -2456,6 +2402,7 @@ def main(args):
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )
+
        images = None
        del pipeline

@@ -73,7 +73,7 @@ from diffusers.utils.import_utils import is_xformers_available


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -200,7 +200,8 @@ Special VAE used for training: {vae_path}.
        "diffusers",
        "diffusers-training",
        lora,
-        "template:sd-lora" "stable-diffusion",
+        "template:sd-lora",
+        "stable-diffusion",
        "stable-diffusion-diffusers",
    ]
    model_card = populate_model_card(model_card, tags=tags)
@@ -662,7 +663,7 @@ def parse_args(input_args=None):
        action="store_true",
        default=False,
        help=(
-            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
+            "Whether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
            "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
        ),
    )
@@ -724,9 +725,9 @@ class TokenEmbeddingsHandler:
        idx = 0
        for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
            assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )

            self.inserting_toks = inserting_toks
            special_tokens_dict = {"additional_special_tokens": self.inserting_toks}
@@ -746,9 +747,9 @@ class TokenEmbeddingsHandler:
                .to(dtype=self.dtype)
                * std_token_embedding
            )
-            self.embeddings_settings[
-                f"original_embeddings_{idx}"
-            ] = text_encoder.text_model.embeddings.token_embedding.weight.data.clone()
+            self.embeddings_settings[f"original_embeddings_{idx}"] = (
+                text_encoder.text_model.embeddings.token_embedding.weight.data.clone()
+            )
            self.embeddings_settings[f"std_token_embedding_{idx}"] = std_token_embedding

            inu = torch.ones((len(tokenizer),), dtype=torch.bool)
@@ -1322,7 +1323,7 @@ def main(args):

        lora_state_dict, network_alphas = StableDiffusionPipeline.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
        if incompatible_keys is not None:
@@ -1883,7 +1884,11 @@ def main(args):
                pipeline.set_progress_bar_config(disable=True)

                # run inference
-                generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+                generator = (
+                    torch.Generator(device=accelerator.device).manual_seed(args.seed)
+                    if args.seed is not None
+                    else None
+                )
                pipeline_args = {"prompt": args.validation_prompt}

                if torch.backends.mps.is_available():
@@ -1987,7 +1992,9 @@ def main(args):
                )
            # run inference
            pipeline = pipeline.to(accelerator.device)
-            generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+            generator = (
+                torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
+            )
            images = [
                pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
                for _ in range(args.num_validation_images)
@@ -71,6 +71,7 @@ from diffusers.utils import (
    convert_unet_state_dict_to_peft,
    is_wandb_available,
 )
+from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module

@@ -79,7 +80,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -101,7 +102,7 @@ def determine_scheduler_type(pretrained_model_name_or_path, revision):
 def save_model_card(
    repo_id: str,
    use_dora: bool,
-    images=None,
+    images: list = None,
    base_model: str = None,
    train_text_encoder=False,
    train_text_encoder_ti=False,
@@ -111,20 +112,17 @@ def save_model_card(
    repo_folder=None,
    vae_path=None,
 ):
-    img_str = "widget:\n"
    lora = "lora" if not use_dora else "dora"
-    for i, image in enumerate(images):
-        image.save(os.path.join(repo_folder, f"image_{i}.png"))
-        img_str += f"""
-        - text: '{validation_prompt if validation_prompt else ' ' }'
-          output:
-            url:
-                "image_{i}.png"
-        """
-    if not images:
-        img_str += f"""
-        - text: '{instance_prompt}'
-        """
+
+    widget_dict = []
+    if images is not None:
+        for i, image in enumerate(images):
+            image.save(os.path.join(repo_folder, f"image_{i}.png"))
+            widget_dict.append(
+                {"text": validation_prompt if validation_prompt else " ", "output": {"url": f"image_{i}.png"}}
+            )
+    else:
+        widget_dict.append({"text": instance_prompt})
    embeddings_filename = f"{repo_folder}_emb"
    instance_prompt_webui = re.sub(r"<s\d+>", "", re.sub(r"<s\d+>", embeddings_filename, instance_prompt, count=1))
    ti_keys = ", ".join(f'"{match}"' for match in re.findall(r"<s\d+>", instance_prompt))
@@ -169,23 +167,7 @@ pipeline.load_textual_inversion(state_dict["clip_g"], token=[{ti_keys}], text_en
 to trigger concept `{key}` → use `{tokens}` in your prompt \n
 """

-    yaml = f"""---
-tags:
- stable-diffusion-xl
- stable-diffusion-xl-diffusers
- diffusers-training
- text-to-image
- diffusers
- {lora}
- template:sd-lora
-{img_str}
-base_model: {base_model}
-instance_prompt: {instance_prompt}
-license: openrail++
---
-"""
-
-    model_card = f"""
+    model_description = f"""
 # SDXL LoRA DreamBooth - {repo_id}

 <Gallery />
@@ -234,8 +216,25 @@ Special VAE used for training: {vae_path}.

 {license}
 """
-    with open(os.path.join(repo_folder, "README.md"), "w") as f:
-        f.write(yaml + model_card)
+    model_card = load_or_create_model_card(
+        repo_id_or_path=repo_id,
+        from_training=True,
+        license="openrail++",
+        base_model=base_model,
+        prompt=instance_prompt,
+        model_description=model_description,
+        widget=widget_dict,
+    )
+    tags = [
+        "text-to-image",
+        "stable-diffusion-xl",
+        "stable-diffusion-xl-diffusers",
+        "text-to-image",
+        "diffusers",
+        lora,
+        "template:sd-lora",
+    ]
+    model_card = populate_model_card(model_card, tags=tags)


 def log_validation(
@@ -269,7 +268,7 @@ def log_validation(
    pipeline.set_progress_bar_config(disable=True)

    # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
    # Currently the context determination is a bit hand-wavy. We can improve it in the future if there's a better
    # way to condition it. Reference: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
    if torch.backends.mps.is_available() or "playground" in args.pretrained_model_name_or_path:
@@ -773,7 +772,7 @@ def parse_args(input_args=None):
        action="store_true",
        default=False,
        help=(
-            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
+            "Whether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
            "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
        ),
    )
@@ -891,9 +890,9 @@ class TokenEmbeddingsHandler:
        idx = 0
        for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
            assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )

            self.inserting_toks = inserting_toks
            special_tokens_dict = {"additional_special_tokens": self.inserting_toks}
@@ -913,9 +912,9 @@ class TokenEmbeddingsHandler:
                .to(dtype=self.dtype)
                * std_token_embedding
            )
-            self.embeddings_settings[
-                f"original_embeddings_{idx}"
-            ] = text_encoder.text_model.embeddings.token_embedding.weight.data.clone()
+            self.embeddings_settings[f"original_embeddings_{idx}"] = (
+                text_encoder.text_model.embeddings.token_embedding.weight.data.clone()
+            )
            self.embeddings_settings[f"std_token_embedding_{idx}"] = std_token_embedding

            inu = torch.ones((len(tokenizer),), dtype=torch.bool)
@@ -1648,7 +1647,7 @@ def main(args):

        lora_state_dict, network_alphas = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
        if incompatible_keys is not None:
@@ -1875,7 +1874,7 @@ def main(args):
    # pack the statically computed variables appropriately here. This is so that we don't
    # have to pass them to the dataloader.

-    # if --train_text_encoder_ti we need add_special_tokens to be True fo textual inversion
+    # if --train_text_encoder_ti we need add_special_tokens to be True for textual inversion
    add_special_tokens = True if args.train_text_encoder_ti else False

    if not train_dataset.custom_instance_prompts:
@@ -720,7 +720,7 @@ def main(args):
    # Train!
    logger.info("***** Running training *****")
    logger.info(f"  Num training steps = {args.max_train_steps}")
-    logger.info(f"  Instantaneous batch size per device = { args.train_batch_size}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")

@@ -61,7 +61,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -722,7 +722,7 @@ def log_validation(
    # pipe.set_progress_bar_config(disable=True)

    # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None

    videos = []
    for _ in range(args.num_validation_videos):
@@ -1138,7 +1138,7 @@ def main(args):
        lora_state_dict = CogVideoXImageToVideoPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -52,7 +52,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -739,7 +739,7 @@ def log_validation(
    # pipe.set_progress_bar_config(disable=True)

    # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None

    videos = []
    for _ in range(args.num_validation_videos):
@@ -1159,7 +1159,7 @@ def main(args):
        lora_state_dict = CogVideoXPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -0,0 +1,201 @@
+# Training CogView4 Control
+
+This (experimental) example shows how to train Control LoRAs with [CogView4](https://huggingface.co/THUDM/CogView4-6B) by conditioning it with additional structural controls (like depth maps, poses, etc.). We provide a script for full fine-tuning, too, refer to [this section](#full-fine-tuning). To know more about CogView4 Control family, refer to the following resources:
+
+To incorporate additional condition latents, we expand the input features of CogView-4 from 64 to 128. The first 64 channels correspond to the original input latents to be denoised, while the latter 64 channels correspond to control latents. This expansion happens on the `patch_embed` layer, where the combined latents are projected to the expected feature dimension of rest of the network. Inference is performed using the `CogView4ControlPipeline`.
+
+> [!NOTE]
+> **Gated model**
+>
+> As the model is gated, before using it with diffusers you first need to go to the [CogView4 Hugging Face page](https://huggingface.co/THUDM/CogView4-6B), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
+
+```bash
+huggingface-cli login
+```
+
+The example command below shows how to launch fine-tuning for pose conditions. The dataset ([`raulc0399/open_pose_controlnet`](https://huggingface.co/datasets/raulc0399/open_pose_controlnet)) being used here already has the pose conditions of the original images, so we don't have to compute them.
+
+```bash
+accelerate launch train_control_lora_cogview4.py \
+  --pretrained_model_name_or_path="THUDM/CogView4-6B" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control-lora" \
+  --mixed_precision="bf16" \
+  --train_batch_size=1 \
+  --rank=64 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=5000 \
+  --validation_image="openpose.png" \
+  --validation_prompt="A couple, 4k photo, highly detailed" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+`openpose.png` comes from [here](https://huggingface.co/Adapter/t2iadapter/resolve/main/openpose.png).
+
+You need to install `diffusers` from the branch of [this PR](https://github.com/huggingface/diffusers/pull/9999). When it's merged, you should install `diffusers` from the `main`.
+
+The training script exposes additional CLI args that might be useful to experiment with:
+
+* `use_lora_bias`: When set, additionally trains the biases of the `lora_B` layer. 
+* `train_norm_layers`: When set, additionally trains the normalization scales. Takes care of saving and loading.
+* `lora_layers`: Specify the layers you want to apply LoRA to. If you specify "all-linear", all the linear layers will be LoRA-attached.
+
+### Training with DeepSpeed
+
+It's possible to train with [DeepSpeed](https://github.com/microsoft/DeepSpeed), specifically leveraging the Zero2 system optimization. To use it, save the following config to an YAML file (feel free to modify as needed):
+
+```yaml
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  gradient_clipping: 1.0
+  offload_optimizer_device: cpu
+  offload_param_device: cpu
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+And then while launching training, pass the config file:
+
+```bash
+accelerate launch --config_file=CONFIG_FILE.yaml ...
+```
+
+### Inference
+
+The pose images in our dataset were computed using the [`controlnet_aux`](https://github.com/huggingface/controlnet_aux) library. Let's install it first:
+
+```bash
+pip install controlnet_aux
+```
+
+And then we are ready:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import CogView4ControlPipeline
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+pipe = CogView4ControlPipeline.from_pretrained("THUDM/CogView4-6B", torch_dtype=torch.bfloat16).to("cuda")
+pipe.load_lora_weights("...") # change this.
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  control_image=image,
+  num_inference_steps=50,
+  joint_attention_kwargs={"scale": 0.9},
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Full fine-tuning
+
+We provide a non-LoRA version of the training script `train_control_cogview4.py`. Here is an example command:
+
+```bash
+accelerate launch --config_file=accelerate_ds2.yaml train_control_cogview4.py \
+  --pretrained_model_name_or_path="THUDM/CogView4-6B" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control" \
+  --mixed_precision="bf16" \
+  --train_batch_size=2 \
+  --dataloader_num_workers=4 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --proportion_empty_prompts=0.2 \
+  --learning_rate=5e-5 \
+  --adam_weight_decay=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="cosine" \
+  --lr_warmup_steps=1000 \
+  --checkpointing_steps=1000 \
+  --max_train_steps=10000 \
+  --validation_steps=200 \
+  --validation_image "2_pose_1024.jpg" "3_pose_1024.jpg" \
+  --validation_prompt "two friends sitting by each other enjoying a day at the park, full hd, cinematic" "person enjoying a day at the park, full hd, cinematic" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+Change the `validation_image` and `validation_prompt` as needed.
+
+For inference, this time, we will run:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import CogView4ControlPipeline, CogView4Transformer2DModel
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+transformer = CogView4Transformer2DModel.from_pretrained("...") # change this.
+pipe = CogView4ControlPipeline.from_pretrained(
+  "THUDM/CogView4-6B",  transformer=transformer, torch_dtype=torch.bfloat16
+).to("cuda")
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  control_image=image,
+  num_inference_steps=50,
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Things to note
+
+* The scripts provided in this directory are experimental and educational. This means we may have to tweak things around to get good results on a given condition. We believe this is best done with the community 🤗
+* The scripts are not memory-optimized but we offload the VAE and the text encoders to CPU when they are not used if `--offload` is specified. 
+* We can extract LoRAs from the fully fine-tuned model. While we currently don't provide any utilities for that, users are welcome to refer to [this script](https://github.com/Stability-AI/stability-ComfyUI-nodes/blob/master/control_lora_create.py) that provides a similar functionality. 
@@ -0,0 +1,6 @@
+transformers==4.47.0
+wandb
+torch
+torchvision
+accelerate==1.2.0
+peft>=0.14.0
@@ -10,6 +10,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif

 | Example                                                                                                                               | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                                              | Colab                                                                                                                                                                                                              |                                                        Author |
 |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
+|Spatiotemporal Skip Guidance (STG)|[Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664) (CVPR 2025) enhances video diffusion models by generating a weaker model through layer skipping and using it as guidance, improving fidelity in models like HunyuanVideo, LTXVideo, and Mochi.|[Spatiotemporal Skip Guidance](#spatiotemporal-skip-guidance)|-|[Junha Hyung](https://junhahyung.github.io/), [Kinam Kim](https://kinam0252.github.io/), and [Ednaordinary](https://github.com/Ednaordinary)|
 |Adaptive Mask Inpainting|Adaptive Mask Inpainting algorithm from [Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models](https://github.com/snuvclab/coma) (ECCV '24, Oral) provides a way to insert human inside the scene image without altering the background, by inpainting with adapting mask.|[Adaptive Mask Inpainting](#adaptive-mask-inpainting)|-|[Hyeonwoo Kim](https://sshowbiz.xyz),[Sookwan Han](https://jellyheadandrew.github.io)|
 |Flux with CFG|[Flux with CFG](https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md) provides an implementation of using CFG in [Flux](https://blackforestlabs.ai/announcing-black-forest-labs/).|[Flux with CFG](#flux-with-cfg)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/flux_with_cfg.ipynb)|[Linoy Tsaban](https://github.com/linoytsaban), [Apolinário](https://github.com/apolinario), and [Sayak Paul](https://github.com/sayakpaul)|
 |Differential Diffusion|[Differential Diffusion](https://github.com/exx8/differential-diffusion) modifies an image according to a text prompt, and according to a map that specifies the amount of change in each region.|[Differential Diffusion](#differential-diffusion)|[![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/exx8/differential-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/exx8/differential-diffusion/blob/main/examples/SD2.ipynb)|[Eran Levin](https://github.com/exx8) and [Ohad Fried](https://www.ohadf.com/)|
@@ -23,12 +24,12 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | Long Prompt Weighting Stable Diffusion                                                                                                | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt.                                                                                                                                                                                                                                                                                                                                                                                                  | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion)         | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/long_prompt_weighting_stable_diffusion.ipynb)                                                                                        |                           [SkyTNT](https://github.com/SkyTNT) |
 | Speech to Image                                                                                                                       | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images                                                                                                                                                                                                                                                                                                                                                                                                            | [Speech to Image](#speech-to-image)                                                       |[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/speech_to_image.ipynb)                                                                                                                                                                                                   |             [Mikail Duzenli](https://github.com/MikailINTech)
 | Wild Card Stable Diffusion                                                                                                            | Stable Diffusion Pipeline that supports prompts that contain wildcard terms (indicated by surrounding double underscores), with values instantiated randomly from a corresponding txt file or a dictionary of possible values                                                                                                                                                                                                                                                                            | [Wildcard Stable Diffusion](#wildcard-stable-diffusion)                                   | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/wildcard_stable_diffusion.ipynb)                                                                                                                                                                                 |              [Shyam Sudhakaran](https://github.com/shyamsn97) |
-| [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "&#124;" in prompts (as an AND condition) and weights (separated by "&#124;" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                            | [Composable Stable Diffusion](#composable-stable-diffusion)                               | -                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
+| [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "&#124;" in prompts (as an AND condition) and weights (separated by "&#124;" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                            | [Composable Stable Diffusion](#composable-stable-diffusion)                               | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/composable_stable_diffusion.ipynb)                                                                                                                                                                                                                 |                      [Mark Rich](https://github.com/MarkRich) |
 | Seed Resizing Stable Diffusion                                                                                                        | Stable Diffusion Pipeline that supports resizing an image and retaining the concepts of the 512 by 512 generation.                                                                                                                                                                                                                                                                                                                                                                                       | [Seed Resizing](#seed-resizing)                                                           | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/seed_resizing.ipynb)                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
 | Imagic Stable Diffusion                                                                                                               | Stable Diffusion Pipeline that enables writing a text prompt to edit an existing image                                                                                                                                                                                                                                                                                                                                                                                                                   | [Imagic Stable Diffusion](#imagic-stable-diffusion)                                       | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/imagic_stable_diffusion.ipynb)                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
 | Multilingual Stable Diffusion                                                                                                         | Stable Diffusion Pipeline that supports prompts in 50 different languages.                                                                                                                                                                                                                                                                                                                                                                                                                               | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline)                  | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/multilingual_stable_diffusion.ipynb)                                                                                                                                                                             |          [Juan Carlos Piñeros](https://github.com/juancopi81) |
 | GlueGen Stable Diffusion                                                                                                         | Stable Diffusion Pipeline that supports prompts in different languages using GlueGen adapter.                                                                                                                                                                                                                                                                                                                                                                                                                               | [GlueGen Stable Diffusion](#gluegen-stable-diffusion-pipeline)                  | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/gluegen_stable_diffusion.ipynb)                                                                                                                                                                                                                  |          [Phạm Hồng Vinh](https://github.com/rootonchair) |
-| Image to Image Inpainting Stable Diffusion                                                                                            | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting                                                                                                                                                                                                                                                                                                                                                                                                            | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | -                                                                                                                                                                                                                  |                    [Alex McKinney](https://github.com/vvvm23) |
+| Image to Image Inpainting Stable Diffusion                                                                                            | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting                                                                                                                                                                                                                                                                                                                                                                                                            | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/image_to_image_inpainting_stable_diffusion.ipynb)                                                                                                                                                                                                                  |                    [Alex McKinney](https://github.com/vvvm23) |
 | Text Based Inpainting Stable Diffusion                                                                                                | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting                                                                                                                                                                                                                                                                                                                                                                                              | [Text Based Inpainting Stable Diffusion](#text-based-inpainting-stable-diffusion)     | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/text_based_inpainting_stable_dffusion.ipynb)                                                                                                                                                                                                    |                   [Dhruv Karan](https://github.com/unography) |
 | Bit Diffusion                                                                                                                         | Diffusion on discrete data                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | [Bit Diffusion](#bit-diffusion)                                                           | -  |                       [Stuti R.](https://github.com/kingstut) |
 | K-Diffusion Stable Diffusion                                                                                                          | Run Stable Diffusion with any of [K-Diffusion's samplers](https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py)                                                                                                                                                                                                                                                                                                                                                                  | [Stable Diffusion with K Diffusion](#stable-diffusion-with-k-diffusion)                   | -  |    [Patrick von Platen](https://github.com/patrickvonplaten/) |
@@ -40,7 +41,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | UnCLIP Image Interpolation Pipeline                                                                                                   | Diffusion Pipeline that allows passing two images/image_embeddings and produces images while interpolating between their image-embeddings                                                                                                                                                                                                                                                                                                                                                                | [UnCLIP Image Interpolation Pipeline](#unclip-image-interpolation-pipeline)               | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/unclip_image_interpolation.ipynb)| [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
 | DDIM Noise Comparative Analysis Pipeline                                                                                              | Investigating how the diffusion models learn visual concepts from each noise level (which is a contribution of [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227))                                                                                                                                                                                                                                                                                                                             | [DDIM Noise Comparative Analysis Pipeline](#ddim-noise-comparative-analysis-pipeline)     | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/ddim_noise_comparative_analysis.ipynb)|              [Aengus (Duc-Anh)](https://github.com/aengusng8) |
 | CLIP Guided Img2Img Stable Diffusion Pipeline                                                                                         | Doing CLIP guidance for image to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                  | [CLIP Guided Img2Img Stable Diffusion](#clip-guided-img2img-stable-diffusion)             | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/clip_guided_img2img_stable_diffusion.ipynb) |               [Nipun Jindal](https://github.com/nipunjindal/) |
-| TensorRT Stable Diffusion Text to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
+| TensorRT Stable Diffusion Text to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline)      | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/tensorrt_text2image_stable_diffusion_pipeline.ipynb) |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
 | EDICT Image Editing Pipeline                                                                                                          | Diffusion pipeline for text-guided image editing                                                                                                                                                                                                                                                                                                                                                                                                                                                         | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline)                             | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/edict_image_pipeline.ipynb) |                    [Joqsan Azocar](https://github.com/Joqsan) |
 | Stable Diffusion RePaint                                                                                                              | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.09865) for inpainting.                                                                                                                                                                                                                                                                                                                                                                                                               | [Stable Diffusion RePaint](#stable-diffusion-repaint )|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_repaint.ipynb)|                  [Markus Pobitzer](https://github.com/Markus-Pobitzer) |
 | TensorRT Stable Diffusion Image to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Image2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Image to Image Pipeline](#tensorrt-image2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
@@ -53,10 +54,11 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | Stable Diffusion Mixture Tiling Pipeline SD 1.5 | A pipeline generates cohesive images by integrating multiple diffusion processes, each focused on a specific image region and considering boundary effects for smooth blending | [Stable Diffusion Mixture Tiling Pipeline SD 1.5](#stable-diffusion-mixture-tiling-pipeline-sd-15) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/albarji/mixture-of-diffusers) | [Álvaro B Jiménez](https://github.com/albarji/) |
 | Stable Diffusion Mixture Canvas Pipeline SD 1.5 | A pipeline generates cohesive images by integrating multiple diffusion processes, each focused on a specific image region and considering boundary effects for smooth blending. Works by defining a list of Text2Image region objects that detail the region of influence of each diffuser. | [Stable Diffusion Mixture Canvas Pipeline SD 1.5](#stable-diffusion-mixture-canvas-pipeline-sd-15) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/albarji/mixture-of-diffusers) | [Álvaro B Jiménez](https://github.com/albarji/) |
 | Stable Diffusion Mixture Tiling Pipeline SDXL | A pipeline generates cohesive images by integrating multiple diffusion processes, each focused on a specific image region and considering boundary effects for smooth blending | [Stable Diffusion Mixture Tiling Pipeline SDXL](#stable-diffusion-mixture-tiling-pipeline-sdxl) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/elismasilva/mixture-of-diffusers-sdxl-tiling) | [Eliseu Silva](https://github.com/DEVAIEXP/) |
+| Stable Diffusion MoD ControlNet Tile SR Pipeline SDXL | This is an advanced pipeline that leverages ControlNet Tile and Mixture-of-Diffusers techniques, integrating tile diffusion directly into the latent space denoising process. Designed to overcome the limitations of conventional pixel-space tile processing, this pipeline delivers Super Resolution (SR) upscaling for higher-quality images, reduced processing time, and greater adaptability. | [Stable Diffusion MoD ControlNet Tile SR Pipeline SDXL](#stable-diffusion-mod-controlnet-tile-sr-pipeline-sdxl) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/elismasilva/mod-control-tile-upscaler-sdxl) | [Eliseu Silva](https://github.com/DEVAIEXP/) |
 | FABRIC - Stable Diffusion with feedback Pipeline | pipeline supports feedback from liked and disliked images | [Stable Diffusion Fabric Pipeline](#stable-diffusion-fabric-pipeline) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_fabric.ipynb)| [Shauray Singh](https://shauray8.github.io/about_shauray/) |
 | sketch inpaint - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion Pipeline](#stable-diffusion-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
 | sketch inpaint xl - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion XL Pipeline](#stable-diffusion-xl-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
-| prompt-to-prompt | change parts of a prompt and retain image structure (see [paper page](https://prompt-to-prompt.github.io/)) | [Prompt2Prompt Pipeline](#prompt2prompt-pipeline) | - | [Umer H. Adil](https://twitter.com/UmerHAdil) |
+| prompt-to-prompt | change parts of a prompt and retain image structure (see [paper page](https://prompt-to-prompt.github.io/)) | [Prompt2Prompt Pipeline](#prompt2prompt-pipeline) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/prompt_2_prompt_pipeline.ipynb) | [Umer H. Adil](https://twitter.com/UmerHAdil) |
 |   Latent Consistency Pipeline                                                                                                    | Implementation of [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://arxiv.org/abs/2310.04378)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Latent Consistency Pipeline](#latent-consistency-pipeline)      | - |              [Simian Luo](https://github.com/luosiallen) |
 |   Latent Consistency Img2img Pipeline                                                                                                    | Img2img pipeline for Latent Consistency Models                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Img2Img Pipeline](#latent-consistency-img2img-pipeline)      | - |              [Logan Zoellner](https://github.com/nagolinc) |
 |   Latent Consistency Interpolation Pipeline                                                                                                    | Interpolate the latent space of Latent Consistency Models with multiple prompts                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Interpolation Pipeline](#latent-consistency-interpolation-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1pK3NrLWJSiJsBynLns1K1-IDTW9zbPvl?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
@@ -82,7 +84,8 @@ PIXART-α Controlnet pipeline | Implementation of the controlnet model for pixar
 | [🪆Matryoshka Diffusion Models](https://huggingface.co/papers/2310.15111) | A diffusion process that denoises inputs at multiple resolutions jointly and uses a NestedUNet architecture where features and parameters for small scale inputs are nested within those of the large scales. See [original codebase](https://github.com/apple/ml-mdm). | [🪆Matryoshka Diffusion Models](#matryoshka-diffusion-models) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/pcuenq/mdm) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/gist/tolgacangoz/1f54875fc7aeaabcf284ebde64820966/matryoshka_hf.ipynb) | [M. Tolga Cangöz](https://github.com/tolgacangoz) |
 | Stable Diffusion XL Attentive Eraser Pipeline |[[AAAI2025 Oral] Attentive Eraser](https://github.com/Anonym0u3/AttentiveEraser) is a novel tuning-free method that enhances object removal capabilities in pre-trained diffusion models.|[Stable Diffusion XL Attentive Eraser Pipeline](#stable-diffusion-xl-attentive-eraser-pipeline)|-|[Wenhao Sun](https://github.com/Anonym0u3) and [Benlei Cui](https://github.com/Benny079)|
 | Perturbed-Attention Guidance |StableDiffusionPAGPipeline is a modification of StableDiffusionPipeline to support Perturbed-Attention Guidance (PAG).|[Perturbed-Attention Guidance](#perturbed-attention-guidance)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/perturbed_attention_guidance.ipynb)|[Hyoungwon Cho](https://github.com/HyoungwonCho)|
-
+| CogVideoX DDIM Inversion Pipeline | Implementation of DDIM inversion and guided attention-based editing denoising process on CogVideoX. | [CogVideoX DDIM Inversion Pipeline](#cogvideox-ddim-inversion-pipeline) | - | [LittleNyima](https://github.com/LittleNyima) |
+| FaithDiff Stable Diffusion XL Pipeline | Implementation of [(CVPR 2025) FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolutionUnleashing Diffusion Priors for Faithful Image Super-resolution](https://arxiv.org/abs/2411.18824) - FaithDiff is a faithful image super-resolution method that leverages latent diffusion models by actively adapting the diffusion prior and jointly fine-tuning its components (encoder and diffusion model) with an alignment module to ensure high fidelity and structural consistency. | [FaithDiff Stable Diffusion XL Pipeline](#faithdiff-stable-diffusion-xl-pipeline) | [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/jychen9811/FaithDiff) | [Junyang Chen, Jinshan Pan, Jiangxin Dong, IMAG Lab, (Adapted by Eliseu Silva)](https://github.com/JyChen9811/FaithDiff) |
 To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.

 ```py
@@ -91,6 +94,54 @@ pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion

 ## Example usages

+### Spatiotemporal Skip Guidance
+
+**Junha Hyung\*, Kinam Kim\*, Susung Hong, Min-Jung Kim, Jaegul Choo**
+
+**KAIST AI, University of Washington**
+
+[*Spatiotemporal Skip Guidance (STG) for Enhanced Video Diffusion Sampling*](https://arxiv.org/abs/2411.18664) (CVPR 2025) is a simple training-free sampling guidance method for enhancing transformer-based video diffusion models. STG employs an implicit weak model via self-perturbation, avoiding the need for external models or additional training. By selectively skipping spatiotemporal layers, STG produces an aligned, degraded version of the original model to boost sample quality without compromising diversity or dynamic degree.
+
+Following is the example video of STG applied to Mochi.
+
+
+https://github.com/user-attachments/assets/148adb59-da61-4c50-9dfa-425dcb5c23b3
+
+More examples and information can be found on the [GitHub repository](https://github.com/junhahyung/STGuidance) and the [Project website](https://junhahyung.github.io/STGuidance/).
+
+#### Usage example
+```python
+import torch
+from pipeline_stg_mochi import MochiSTGPipeline
+from diffusers.utils import export_to_video
+
+# Load the pipeline
+pipe = MochiSTGPipeline.from_pretrained("genmo/mochi-1-preview", variant="bf16", torch_dtype=torch.bfloat16)
+
+# Enable memory savings
+pipe = pipe.to("cuda")
+
+#--------Option--------#
+prompt = "A close-up of a beautiful woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues, realistic style."
+stg_applied_layers_idx = [34]
+stg_scale = 1.0 # 0.0 for CFG
+#----------------------#
+
+# Generate video frames
+frames = pipe(
+    prompt, 
+    height=480,
+    width=480,
+    num_frames=81,
+    stg_applied_layers_idx=stg_applied_layers_idx,
+    stg_scale=stg_scale,
+    generator = torch.Generator().manual_seed(42),
+    do_rescaling=do_rescaling,
+).frames[0]
+
+export_to_video(frames, "output.mp4", fps=30)
+```
+
 ### Adaptive Mask Inpainting

 **Hyeonwoo Kim\*, Sookwan Han\*, Patrick Kwon, Hanbyul Joo**
@@ -902,6 +953,7 @@ for i in range(args.num_images):
    images.append(th.from_numpy(np.array(image)).permute(2, 0, 1) / 255.)
 grid = tvu.make_grid(th.stack(images, dim=0), nrow=4, padding=0)
 tvu.save_image(grid, f'{prompt}_{args.weights}' + '.png')
+print("Image saved successfully!")
 ```

 ### Imagic Stable Diffusion
@@ -1217,28 +1269,39 @@ The aim is to overlay two images, then mask out the boundary between `image` and
 For example, this could be used to place a logo on a shirt and make it blend seamlessly.

 ```python
-import PIL
 import torch
-
+import requests
+from PIL import Image
+from io import BytesIO
 from diffusers import DiffusionPipeline

-image_path = "./path-to-image.png"
-inner_image_path = "./path-to-inner-image.png"
-mask_path = "./path-to-mask.png"
+image_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+inner_image_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"

-init_image = PIL.Image.open(image_path).convert("RGB").resize((512, 512))
-inner_image = PIL.Image.open(inner_image_path).convert("RGBA").resize((512, 512))
-mask_image = PIL.Image.open(mask_path).convert("RGB").resize((512, 512))
+def load_image(url, mode="RGB"):
+    response = requests.get(url)
+    if response.status_code == 200:
+        return Image.open(BytesIO(response.content)).convert(mode).resize((512, 512))
+    else:
+        raise FileNotFoundError(f"Could not retrieve image from {url}")
+
+
+init_image = load_image(image_url, mode="RGB")
+inner_image = load_image(inner_image_url, mode="RGBA")
+mask_image = load_image(mask_url, mode="RGB")

 pipe = DiffusionPipeline.from_pretrained(
-    "runwayml/stable-diffusion-inpainting",
+    "stable-diffusion-v1-5/stable-diffusion-inpainting",
    custom_pipeline="img2img_inpainting",
    torch_dtype=torch.float16
 )
 pipe = pipe.to("cuda")

-prompt = "Your prompt here!"
+prompt = "a mecha robot sitting on a bench"
 image = pipe(prompt=prompt, image=init_image, inner_image=inner_image, mask_image=mask_image).images[0]
+
+image.save("output.png")
 ```

 ![2 by 2 grid demonstrating image to image inpainting.](https://user-images.githubusercontent.com/44398246/203506577-ec303be4-887e-4ebd-a773-c83fcb3dd01a.png)
@@ -2630,6 +2693,103 @@ image = pipe(

 ![mixture_tiling_results](https://huggingface.co/datasets/elismasilva/results/resolve/main/mixture_of_diffusers_sdxl_1.png)

+### Stable Diffusion MoD ControlNet Tile SR Pipeline SDXL
+
+This pipeline implements the [MoD (Mixture-of-Diffusers)]("https://arxiv.org/pdf/2408.06072") tiled diffusion technique and combines it with SDXL's ControlNet Tile process to generate SR images.
+
+This works better with 4x scales, but you can try adjusts parameters to higher scales.
+
+````python
+import torch
+from diffusers import DiffusionPipeline, ControlNetUnionModel, AutoencoderKL, UniPCMultistepScheduler, UNet2DConditionModel
+from diffusers.utils import load_image
+from PIL import Image
+
+device = "cuda"
+
+# Initialize the models and pipeline
+controlnet = ControlNetUnionModel.from_pretrained(
+    "brad-twinkl/controlnet-union-sdxl-1.0-promax", torch_dtype=torch.float16
+).to(device=device)
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to(device=device)
+
+model_id = "SG161222/RealVisXL_V5.0"
+pipe = DiffusionPipeline.from_pretrained(
+    model_id,
+    torch_dtype=torch.float16,
+    vae=vae,
+    controlnet=controlnet,
+    custom_pipeline="mod_controlnet_tile_sr_sdxl",    
+    use_safetensors=True,
+    variant="fp16",
+).to(device)
+
+unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet", variant="fp16", use_safetensors=True)
+
+#pipe.enable_model_cpu_offload()  # << Enable this if you have limited VRAM
+pipe.enable_vae_tiling() # << Enable this if you have limited VRAM
+pipe.enable_vae_slicing() # << Enable this if you have limited VRAM
+
+# Set selected scheduler
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+
+# Load image
+control_image = load_image("https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/1.jpg")
+original_height = control_image.height
+original_width = control_image.width
+print(f"Current resolution: H:{original_height} x W:{original_width}")
+
+# Pre-upscale image for tiling
+resolution = 4096
+tile_gaussian_sigma = 0.3
+max_tile_size = 1024 # or 1280
+
+current_size = max(control_image.size)
+scale_factor = max(2, resolution / current_size)
+new_size = (int(control_image.width * scale_factor), int(control_image.height * scale_factor))
+image = control_image.resize(new_size, Image.LANCZOS)
+
+# Update target height and width
+target_height = image.height
+target_width = image.width
+print(f"Target resolution: H:{target_height} x W:{target_width}")
+
+# Calculate overlap size
+normal_tile_overlap, border_tile_overlap = pipe.calculate_overlap(target_width, target_height)
+
+# Set other params
+tile_weighting_method = pipe.TileWeightingMethod.COSINE.value
+guidance_scale = 4
+num_inference_steps = 35
+denoising_strenght = 0.65
+controlnet_strength = 1.0
+prompt = "high-quality, noise-free edges, high quality, 4k, hd, 8k"
+negative_prompt = "blurry, pixelated, noisy, low resolution, artifacts, poor details"
+
+# Image generation
+generated_image = pipe(
+    image=image,
+    control_image=control_image,
+    control_mode=[6],
+    controlnet_conditioning_scale=float(controlnet_strength),
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    normal_tile_overlap=normal_tile_overlap,
+    border_tile_overlap=border_tile_overlap,
+    height=target_height,
+    width=target_width,
+    original_size=(original_width, original_height),
+    target_size=(target_width, target_height),
+    guidance_scale=guidance_scale,        
+    strength=float(denoising_strenght),
+    tile_weighting_method=tile_weighting_method,
+    max_tile_size=max_tile_size,
+    tile_gaussian_sigma=float(tile_gaussian_sigma),
+    num_inference_steps=num_inference_steps,
+)["images"][0]
+````
+![Upscaled](https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/1_input_4x.png)
+
 ### TensorRT Inpainting Stable Diffusion Pipeline

 The TensorRT Pipeline can be used to accelerate the Inpainting Stable Diffusion Inference run.
@@ -3103,14 +3263,19 @@ Here's a full example for `ReplaceEdit``:

 ```python
 import torch
-import numpy as np
-import matplotlib.pyplot as plt
 from diffusers import DiffusionPipeline
+import numpy as np
+from PIL import Image

-pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="pipeline_prompt2prompt").to("cuda")
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4", 
+    custom_pipeline="pipeline_prompt2prompt"
+).to("cuda")

-prompts = ["A turtle playing with a ball",
-           "A monkey playing with a ball"]
+prompts = [
+    "A turtle playing with a ball",
+    "A monkey playing with a ball"
+]

 cross_attention_kwargs = {
    "edit_type": "replace",
@@ -3118,7 +3283,15 @@ cross_attention_kwargs = {
    "self_replace_steps": 0.4
 }

-outputs = pipe(prompt=prompts, height=512, width=512, num_inference_steps=50, cross_attention_kwargs=cross_attention_kwargs)
+outputs = pipe(
+    prompt=prompts,
+    height=512,
+    width=512,
+    num_inference_steps=50,
+    cross_attention_kwargs=cross_attention_kwargs
+)
+
+outputs.images[0].save("output_image_0.png")
 ```

 And abbreviated examples for the other edits:
@@ -5124,3 +5297,139 @@ with torch.no_grad():

 In the folder examples/pixart there is also a script that can be used to train new models.
 Please check the script `train_controlnet_hf_diffusers.sh` on how to start the training.
+
+# CogVideoX DDIM Inversion Pipeline
+
+This implementation performs DDIM inversion on the video based on CogVideoX and uses guided attention to reconstruct or edit the inversion latents.
+
+## Example Usage
+
+```python
+import torch
+
+from examples.community.cogvideox_ddim_inversion import CogVideoXPipelineForDDIMInversion
+
+
+# Load pretrained pipeline
+pipeline = CogVideoXPipelineForDDIMInversion.from_pretrained(
+    "THUDM/CogVideoX1.5-5B",
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+
+# Run DDIM inversion, and the videos will be generated in the output_path
+output = pipeline_for_inversion(
+    prompt="prompt that describes the edited video",
+    video_path="path/to/input.mp4",
+    guidance_scale=6.0,
+    num_inference_steps=50,
+    skip_frames_start=0,
+    skip_frames_end=0,
+    frame_sample_step=None,
+    max_num_frames=81,
+    width=720,
+    height=480,
+    seed=42,
+)
+pipeline.export_latents_to_video(output.inverse_latents[-1], "path/to/inverse_video.mp4", fps=8)
+pipeline.export_latents_to_video(output.recon_latents[-1], "path/to/recon_video.mp4", fps=8)
+```
+# FaithDiff Stable Diffusion XL Pipeline
+
+[Project](https://jychen9811.github.io/FaithDiff_page/) / [GitHub](https://github.com/JyChen9811/FaithDiff/)
+
+This the implementation of the FaithDiff pipeline for SDXL, adapted to use the HuggingFace Diffusers.
+
+For more details see the project links above.
+
+## Example Usage
+
+This example upscale and restores a low-quality image. The input image has a resolution of 512x512 and will be upscaled at a scale of 2x, to a final resolution of 1024x1024. It is possible to upscale to a larger scale, but it is recommended that the input image be at least 1024x1024 in these cases. To upscale this image by 4x, for example, it would be recommended to re-input the result into a new 2x processing, thus performing progressive scaling.
+
+````py
+import random
+import numpy as np
+import torch
+from diffusers import DiffusionPipeline, AutoencoderKL, UniPCMultistepScheduler
+from huggingface_hub import hf_hub_download
+from diffusers.utils import load_image
+from PIL import Image
+
+device = "cuda"
+dtype = torch.float16
+MAX_SEED = np.iinfo(np.int32).max
+
+# Download weights for additional unet layers
+model_file = hf_hub_download(
+    "jychen9811/FaithDiff",
+    filename="FaithDiff.bin", local_dir="./proc_data/faithdiff", local_dir_use_symlinks=False
+)
+
+# Initialize the models and pipeline
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=dtype)
+
+model_id = "SG161222/RealVisXL_V4.0"
+pipe = DiffusionPipeline.from_pretrained(
+    model_id,
+    torch_dtype=dtype,
+    vae=vae,
+    unet=None, #<- Do not load with original model.
+    custom_pipeline="pipeline_faithdiff_stable_diffusion_xl",    
+    use_safetensors=True,
+    variant="fp16",
+).to(device)
+
+# Here we need use pipeline internal unet model
+pipe.unet = pipe.unet_model.from_pretrained(model_id, subfolder="unet", variant="fp16", use_safetensors=True)
+
+# Load aditional layers to the model
+pipe.unet.load_additional_layers(weight_path="proc_data/faithdiff/FaithDiff.bin", dtype=dtype)
+
+# Enable vae tiling
+pipe.set_encoder_tile_settings()
+pipe.enable_vae_tiling()
+
+# Optimization
+pipe.enable_model_cpu_offload()
+
+# Set selected scheduler
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+
+#input params
+prompt = "The image features a woman in her 55s with blonde hair and a white shirt, smiling at the camera. She appears to be in a good mood and is wearing a white scarf around her neck. "
+upscale = 2 # scale here
+start_point = "lr" # or "noise"
+latent_tiled_overlap = 0.5
+latent_tiled_size = 1024
+
+# Load image
+lq_image = load_image("https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/woman.png")
+original_height = lq_image.height
+original_width = lq_image.width
+print(f"Current resolution: H:{original_height} x W:{original_width}")
+
+width = original_width * int(upscale)
+height = original_height * int(upscale)
+print(f"Final resolution: H:{height} x W:{width}")
+
+# Restoration
+image = lq_image.resize((width, height), Image.LANCZOS)
+input_image, width_init, height_init, width_now, height_now = pipe.check_image_size(image)
+
+generator = torch.Generator(device=device).manual_seed(random.randint(0, MAX_SEED))
+gen_image = pipe(lr_img=input_image, 
+                 prompt = prompt,                  
+                 num_inference_steps=20, 
+                 guidance_scale=5, 
+                 generator=generator, 
+                 start_point=start_point, 
+                 height = height_now, 
+                 width=width_now, 
+                 overlap=latent_tiled_overlap, 
+                 target_size=(latent_tiled_size, latent_tiled_size)
+                ).images[0]
+
+cropped_image = gen_image.crop((0, 0, width_init, height_init))
+cropped_image.save("data/result.png")
+````
+### Result
+[<img src="https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/faithdiff_restored.PNG" width="512px" height="512px"/>](https://imgsli.com/MzY1NzE2)
@@ -1103,7 +1103,7 @@ class AdaptiveMaskInpaintPipeline(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `default_mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -0,0 +1,645 @@
+"""
+This script performs DDIM inversion for video frames using a pre-trained model and generates
+a video reconstruction based on a provided prompt. It utilizes the CogVideoX pipeline to
+process video frames, apply the DDIM inverse scheduler, and produce an output video.
+
+**Please notice that this script is based on the CogVideoX 5B model, and would not generate
+a good result for 2B variants.**
+
+Usage:
+    python cogvideox_ddim_inversion.py
+        --model-path /path/to/model
+        --prompt "a prompt"
+        --video-path /path/to/video.mp4
+        --output-path /path/to/output
+
+For more details about the cli arguments, please run `python cogvideox_ddim_inversion.py --help`.
+
+Author:
+    LittleNyima <littlenyima[at]163[dot]com>
+"""
+
+import argparse
+import math
+import os
+from typing import Any, Dict, List, Optional, Tuple, TypedDict, Union, cast
+
+import torch
+import torch.nn.functional as F
+import torchvision.transforms as T
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers.models.attention_processor import Attention, CogVideoXAttnProcessor2_0
+from diffusers.models.autoencoders import AutoencoderKLCogVideoX
+from diffusers.models.embeddings import apply_rotary_emb
+from diffusers.models.transformers.cogvideox_transformer_3d import CogVideoXBlock, CogVideoXTransformer3DModel
+from diffusers.pipelines.cogvideo.pipeline_cogvideox import CogVideoXPipeline, retrieve_timesteps
+from diffusers.schedulers import CogVideoXDDIMScheduler, DDIMInverseScheduler
+from diffusers.utils import export_to_video
+
+
+# Must import after torch because this can sometimes lead to a nasty segmentation fault, or stack smashing error.
+# Very few bug reports but it happens. Look in decord Github issues for more relevant information.
+import decord  # isort: skip
+
+
+class DDIMInversionArguments(TypedDict):
+    model_path: str
+    prompt: str
+    video_path: str
+    output_path: str
+    guidance_scale: float
+    num_inference_steps: int
+    skip_frames_start: int
+    skip_frames_end: int
+    frame_sample_step: Optional[int]
+    max_num_frames: int
+    width: int
+    height: int
+    fps: int
+    dtype: torch.dtype
+    seed: int
+    device: torch.device
+
+
+def get_args() -> DDIMInversionArguments:
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--model_path", type=str, required=True, help="Path of the pretrained model")
+    parser.add_argument("--prompt", type=str, required=True, help="Prompt for the direct sample procedure")
+    parser.add_argument("--video_path", type=str, required=True, help="Path of the video for inversion")
+    parser.add_argument("--output_path", type=str, default="output", help="Path of the output videos")
+    parser.add_argument("--guidance_scale", type=float, default=6.0, help="Classifier-free guidance scale")
+    parser.add_argument("--num_inference_steps", type=int, default=50, help="Number of inference steps")
+    parser.add_argument("--skip_frames_start", type=int, default=0, help="Number of skipped frames from the start")
+    parser.add_argument("--skip_frames_end", type=int, default=0, help="Number of skipped frames from the end")
+    parser.add_argument("--frame_sample_step", type=int, default=None, help="Temporal stride of the sampled frames")
+    parser.add_argument("--max_num_frames", type=int, default=81, help="Max number of sampled frames")
+    parser.add_argument("--width", type=int, default=720, help="Resized width of the video frames")
+    parser.add_argument("--height", type=int, default=480, help="Resized height of the video frames")
+    parser.add_argument("--fps", type=int, default=8, help="Frame rate of the output videos")
+    parser.add_argument("--dtype", type=str, default="bf16", choices=["bf16", "fp16"], help="Dtype of the model")
+    parser.add_argument("--seed", type=int, default=42, help="Seed for the random number generator")
+    parser.add_argument("--device", type=str, default="cuda", choices=["cuda", "cpu"], help="Device for inference")
+
+    args = parser.parse_args()
+    args.dtype = torch.bfloat16 if args.dtype == "bf16" else torch.float16
+    args.device = torch.device(args.device)
+
+    return DDIMInversionArguments(**vars(args))
+
+
+class CogVideoXAttnProcessor2_0ForDDIMInversion(CogVideoXAttnProcessor2_0):
+    def __init__(self):
+        super().__init__()
+
+    def calculate_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn: Attention,
+        batch_size: int,
+        image_seq_length: int,
+        text_seq_length: int,
+        attention_mask: Optional[torch.Tensor],
+        image_rotary_emb: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        r"""
+        Core attention computation with inversion-guided RoPE integration.
+
+        Args:
+            query (`torch.Tensor`): `[batch_size, seq_len, dim]` query tensor
+            key (`torch.Tensor`): `[batch_size, seq_len, dim]` key tensor
+            value (`torch.Tensor`): `[batch_size, seq_len, dim]` value tensor
+            attn (`Attention`): Parent attention module with projection layers
+            batch_size (`int`): Effective batch size (after chunk splitting)
+            image_seq_length (`int`): Length of image feature sequence
+            text_seq_length (`int`): Length of text feature sequence
+            attention_mask (`Optional[torch.Tensor]`): Attention mask tensor
+            image_rotary_emb (`Optional[torch.Tensor]`): Rotary embeddings for image positions
+
+        Returns:
+            `Tuple[torch.Tensor, torch.Tensor]`:
+                (1) hidden_states: [batch_size, image_seq_length, dim] processed image features
+                (2) encoder_hidden_states: [batch_size, text_seq_length, dim] processed text features
+        """
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if image_rotary_emb is not None:
+            query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
+            if not attn.is_cross_attention:
+                if key.size(2) == query.size(2):  # Attention for reference hidden states
+                    key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
+                else:  # RoPE should be applied to each group of image tokens
+                    key[:, :, text_seq_length : text_seq_length + image_seq_length] = apply_rotary_emb(
+                        key[:, :, text_seq_length : text_seq_length + image_seq_length], image_rotary_emb
+                    )
+                    key[:, :, text_seq_length * 2 + image_seq_length :] = apply_rotary_emb(
+                        key[:, :, text_seq_length * 2 + image_seq_length :], image_rotary_emb
+                    )
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        encoder_hidden_states, hidden_states = hidden_states.split(
+            [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
+        )
+        return hidden_states, encoder_hidden_states
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        r"""
+        Process the dual-path attention for the inversion-guided denoising procedure.
+
+        Args:
+            attn (`Attention`): Parent attention module
+            hidden_states (`torch.Tensor`): `[batch_size, image_seq_len, dim]` Image tokens
+            encoder_hidden_states (`torch.Tensor`): `[batch_size, text_seq_len, dim]` Text tokens
+            attention_mask (`Optional[torch.Tensor]`): Optional attention mask
+            image_rotary_emb (`Optional[torch.Tensor]`): Rotary embeddings for image tokens
+
+        Returns:
+            `Tuple[torch.Tensor, torch.Tensor]`:
+                (1) Final hidden states: `[batch_size, image_seq_length, dim]` Resulting image tokens
+                (2) Final encoder states: `[batch_size, text_seq_length, dim]` Resulting text tokens
+        """
+        image_seq_length = hidden_states.size(1)
+        text_seq_length = encoder_hidden_states.size(1)
+
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        query, query_reference = query.chunk(2)
+        key, key_reference = key.chunk(2)
+        value, value_reference = value.chunk(2)
+        batch_size = batch_size // 2
+
+        hidden_states, encoder_hidden_states = self.calculate_attention(
+            query=query,
+            key=torch.cat((key, key_reference), dim=1),
+            value=torch.cat((value, value_reference), dim=1),
+            attn=attn,
+            batch_size=batch_size,
+            image_seq_length=image_seq_length,
+            text_seq_length=text_seq_length,
+            attention_mask=attention_mask,
+            image_rotary_emb=image_rotary_emb,
+        )
+        hidden_states_reference, encoder_hidden_states_reference = self.calculate_attention(
+            query=query_reference,
+            key=key_reference,
+            value=value_reference,
+            attn=attn,
+            batch_size=batch_size,
+            image_seq_length=image_seq_length,
+            text_seq_length=text_seq_length,
+            attention_mask=attention_mask,
+            image_rotary_emb=image_rotary_emb,
+        )
+
+        return (
+            torch.cat((hidden_states, hidden_states_reference)),
+            torch.cat((encoder_hidden_states, encoder_hidden_states_reference)),
+        )
+
+
+class OverrideAttnProcessors:
+    r"""
+    Context manager for temporarily overriding attention processors in CogVideo transformer blocks.
+
+    Designed for DDIM inversion process, replaces original attention processors with
+    `CogVideoXAttnProcessor2_0ForDDIMInversion` and restores them upon exit. Uses Python context manager
+    pattern to safely manage processor replacement.
+
+    Typical usage:
+    ```python
+    with OverrideAttnProcessors(transformer):
+        # Perform DDIM inversion operations
+    ```
+
+    Args:
+        transformer (`CogVideoXTransformer3DModel`):
+            The transformer model containing attention blocks to be modified. Should have
+            `transformer_blocks` attribute containing `CogVideoXBlock` instances.
+    """
+
+    def __init__(self, transformer: CogVideoXTransformer3DModel):
+        self.transformer = transformer
+        self.original_processors = {}
+
+    def __enter__(self):
+        for block in self.transformer.transformer_blocks:
+            block = cast(CogVideoXBlock, block)
+            self.original_processors[id(block)] = block.attn1.get_processor()
+            block.attn1.set_processor(CogVideoXAttnProcessor2_0ForDDIMInversion())
+
+    def __exit__(self, _0, _1, _2):
+        for block in self.transformer.transformer_blocks:
+            block = cast(CogVideoXBlock, block)
+            block.attn1.set_processor(self.original_processors[id(block)])
+
+
+def get_video_frames(
+    video_path: str,
+    width: int,
+    height: int,
+    skip_frames_start: int,
+    skip_frames_end: int,
+    max_num_frames: int,
+    frame_sample_step: Optional[int],
+) -> torch.FloatTensor:
+    """
+    Extract and preprocess video frames from a video file for VAE processing.
+
+    Args:
+        video_path (`str`): Path to input video file
+        width (`int`): Target frame width for decoding
+        height (`int`): Target frame height for decoding
+        skip_frames_start (`int`): Number of frames to skip at video start
+        skip_frames_end (`int`): Number of frames to skip at video end
+        max_num_frames (`int`): Maximum allowed number of output frames
+        frame_sample_step (`Optional[int]`):
+            Frame sampling step size. If None, automatically calculated as:
+            (total_frames - skipped_frames) // max_num_frames
+
+    Returns:
+        `torch.FloatTensor`: Preprocessed frames in `[F, C, H, W]` format where:
+        - `F`: Number of frames (adjusted to 4k + 1 for VAE compatibility)
+        - `C`: Channels (3 for RGB)
+        - `H`: Frame height
+        - `W`: Frame width
+    """
+    with decord.bridge.use_torch():
+        video_reader = decord.VideoReader(uri=video_path, width=width, height=height)
+        video_num_frames = len(video_reader)
+        start_frame = min(skip_frames_start, video_num_frames)
+        end_frame = max(0, video_num_frames - skip_frames_end)
+
+        if end_frame <= start_frame:
+            indices = [start_frame]
+        elif end_frame - start_frame <= max_num_frames:
+            indices = list(range(start_frame, end_frame))
+        else:
+            step = frame_sample_step or (end_frame - start_frame) // max_num_frames
+            indices = list(range(start_frame, end_frame, step))
+
+        frames = video_reader.get_batch(indices=indices)
+        frames = frames[:max_num_frames].float()  # ensure that we don't go over the limit
+
+        # Choose first (4k + 1) frames as this is how many is required by the VAE
+        selected_num_frames = frames.size(0)
+        remainder = (3 + selected_num_frames) % 4
+        if remainder != 0:
+            frames = frames[:-remainder]
+        assert frames.size(0) % 4 == 1
+
+        # Normalize the frames
+        transform = T.Lambda(lambda x: x / 255.0 * 2.0 - 1.0)
+        frames = torch.stack(tuple(map(transform, frames)), dim=0)
+
+        return frames.permute(0, 3, 1, 2).contiguous()  # [F, C, H, W]
+
+
+class CogVideoXDDIMInversionOutput:
+    inverse_latents: torch.FloatTensor
+    recon_latents: torch.FloatTensor
+
+    def __init__(self, inverse_latents: torch.FloatTensor, recon_latents: torch.FloatTensor):
+        self.inverse_latents = inverse_latents
+        self.recon_latents = recon_latents
+
+
+class CogVideoXPipelineForDDIMInversion(CogVideoXPipeline):
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: CogVideoXDDIMScheduler,
+    ):
+        super().__init__(
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            vae=vae,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        self.inverse_scheduler = DDIMInverseScheduler(**scheduler.config)
+
+    def encode_video_frames(self, video_frames: torch.FloatTensor) -> torch.FloatTensor:
+        """
+        Encode video frames into latent space using Variational Autoencoder.
+
+        Args:
+            video_frames (`torch.FloatTensor`):
+                Input frames tensor in `[F, C, H, W]` format from `get_video_frames()`
+
+        Returns:
+            `torch.FloatTensor`: Encoded latents in `[1, F, D, H_latent, W_latent]` format where:
+            - `F`: Number of frames (same as input)
+            - `D`: Latent channel dimension
+            - `H_latent`: Latent space height (H // 2^vae.downscale_factor)
+            - `W_latent`: Latent space width (W // 2^vae.downscale_factor)
+        """
+        vae: AutoencoderKLCogVideoX = self.vae
+        video_frames = video_frames.to(device=vae.device, dtype=vae.dtype)
+        video_frames = video_frames.unsqueeze(0).permute(0, 2, 1, 3, 4)  # [B, C, F, H, W]
+        latent_dist = vae.encode(x=video_frames).latent_dist.sample().transpose(1, 2)
+        return latent_dist * vae.config.scaling_factor
+
+    @torch.no_grad()
+    def export_latents_to_video(self, latents: torch.FloatTensor, video_path: str, fps: int):
+        r"""
+        Decode latent vectors into video and export as video file.
+
+        Args:
+            latents (`torch.FloatTensor`): Encoded latents in `[B, F, D, H_latent, W_latent]` format from
+                `encode_video_frames()`
+            video_path (`str`): Output path for video file
+            fps (`int`): Target frames per second for output video
+        """
+        video = self.decode_latents(latents)
+        frames = self.video_processor.postprocess_video(video=video, output_type="pil")
+        os.makedirs(os.path.dirname(video_path), exist_ok=True)
+        export_to_video(video_frames=frames[0], output_video_path=video_path, fps=fps)
+
+    # Modified from CogVideoXPipeline.__call__
+    @torch.no_grad()
+    def sample(
+        self,
+        latents: torch.FloatTensor,
+        scheduler: Union[DDIMInverseScheduler, CogVideoXDDIMScheduler],
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 6,
+        use_dynamic_cfg: bool = False,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        reference_latents: torch.FloatTensor = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Execute the core sampling loop for video generation/inversion using CogVideoX.
+
+        Implements the full denoising trajectory recording for both DDIM inversion and
+        generation processes. Supports dynamic classifier-free guidance and reference
+        latent conditioning.
+
+        Args:
+            latents (`torch.FloatTensor`):
+                Initial noise tensor of shape `[B, F, C, H, W]`.
+            scheduler (`Union[DDIMInverseScheduler, CogVideoXDDIMScheduler]`):
+                Scheduling strategy for diffusion process. Use:
+                (1) `DDIMInverseScheduler` for inversion
+                (2) `CogVideoXDDIMScheduler` for generation
+            prompt (`Optional[Union[str, List[str]]]`):
+                Text prompt(s) for conditional generation. Defaults to unconditional.
+            negative_prompt (`Optional[Union[str, List[str]]]`):
+                Negative prompt(s) for guidance. Requires `guidance_scale > 1`.
+            num_inference_steps (`int`):
+                Number of denoising steps. Affects quality/compute trade-off.
+            guidance_scale (`float`):
+                Classifier-free guidance weight. 1.0 = no guidance.
+            use_dynamic_cfg (`bool`):
+                Enable time-varying guidance scale (cosine schedule)
+            eta (`float`):
+                DDIM variance parameter (0 = deterministic process)
+            generator (`Optional[Union[torch.Generator, List[torch.Generator]]]`):
+                Random number generator(s) for reproducibility
+            attention_kwargs (`Optional[Dict[str, Any]]`):
+                Custom parameters for attention modules
+            reference_latents (`torch.FloatTensor`):
+                Reference latent trajectory for conditional sampling. Shape should match
+                `[T, B, F, C, H, W]` where `T` is number of timesteps
+
+        Returns:
+            `torch.FloatTensor`:
+                Full denoising trajectory tensor of shape `[T, B, F, C, H, W]`.
+        """
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            do_classifier_free_guidance,
+            device=device,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+        if reference_latents is not None:
+            prompt_embeds = torch.cat([prompt_embeds] * 2, dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(scheduler, num_inference_steps, device)
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latents.
+        latents = latents.to(device=device) * scheduler.init_noise_sigma
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+        if isinstance(scheduler, DDIMInverseScheduler):  # Inverse scheduler does not accept extra kwargs
+            extra_step_kwargs = {}
+
+        # 7. Create rotary embeds if required
+        image_rotary_emb = (
+            self._prepare_rotary_positional_embeddings(
+                height=latents.size(3) * self.vae_scale_factor_spatial,
+                width=latents.size(4) * self.vae_scale_factor_spatial,
+                num_frames=latents.size(1),
+                device=device,
+            )
+            if self.transformer.config.use_rotary_positional_embeddings
+            else None
+        )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * scheduler.order, 0)
+
+        trajectory = torch.zeros_like(latents).unsqueeze(0).repeat(len(timesteps), 1, 1, 1, 1, 1)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                if reference_latents is not None:
+                    reference = reference_latents[i]
+                    reference = torch.cat([reference] * 2) if do_classifier_free_guidance else reference
+                    latent_model_input = torch.cat([latent_model_input, reference], dim=0)
+                latent_model_input = scheduler.scale_model_input(latent_model_input, t)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    image_rotary_emb=image_rotary_emb,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                if reference_latents is not None:  # Recover the original batch size
+                    noise_pred, _ = noise_pred.chunk(2)
+
+                # perform guidance
+                if use_dynamic_cfg:
+                    self._guidance_scale = 1 + guidance_scale * (
+                        (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
+                    )
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the noisy sample x_t-1 -> x_t
+                latents = scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                latents = latents.to(prompt_embeds.dtype)
+                trajectory[i] = latents
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % scheduler.order == 0):
+                    progress_bar.update()
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        return trajectory
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: str,
+        video_path: str,
+        guidance_scale: float,
+        num_inference_steps: int,
+        skip_frames_start: int,
+        skip_frames_end: int,
+        frame_sample_step: Optional[int],
+        max_num_frames: int,
+        width: int,
+        height: int,
+        seed: int,
+    ):
+        """
+        Performs DDIM inversion on a video to reconstruct it with a new prompt.
+
+        Args:
+            prompt (`str`): The text prompt to guide the reconstruction.
+            video_path (`str`): Path to the input video file.
+            guidance_scale (`float`): Scale for classifier-free guidance.
+            num_inference_steps (`int`): Number of denoising steps.
+            skip_frames_start (`int`): Number of frames to skip from the beginning of the video.
+            skip_frames_end (`int`): Number of frames to skip from the end of the video.
+            frame_sample_step (`Optional[int]`): Step size for sampling frames. If None, all frames are used.
+            max_num_frames (`int`): Maximum number of frames to process.
+            width (`int`): Width of the output video frames.
+            height (`int`): Height of the output video frames.
+            seed (`int`): Random seed for reproducibility.
+
+        Returns:
+            `CogVideoXDDIMInversionOutput`: Contains the inverse latents and reconstructed latents.
+        """
+        if not self.transformer.config.use_rotary_positional_embeddings:
+            raise NotImplementedError("This script supports CogVideoX 5B model only.")
+        video_frames = get_video_frames(
+            video_path=video_path,
+            width=width,
+            height=height,
+            skip_frames_start=skip_frames_start,
+            skip_frames_end=skip_frames_end,
+            max_num_frames=max_num_frames,
+            frame_sample_step=frame_sample_step,
+        ).to(device=self.device)
+        video_latents = self.encode_video_frames(video_frames=video_frames)
+        inverse_latents = self.sample(
+            latents=video_latents,
+            scheduler=self.inverse_scheduler,
+            prompt="",
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+            generator=torch.Generator(device=self.device).manual_seed(seed),
+        )
+        with OverrideAttnProcessors(transformer=self.transformer):
+            recon_latents = self.sample(
+                latents=torch.randn_like(video_latents),
+                scheduler=self.scheduler,
+                prompt=prompt,
+                num_inference_steps=num_inference_steps,
+                guidance_scale=guidance_scale,
+                generator=torch.Generator(device=self.device).manual_seed(seed),
+                reference_latents=reversed(inverse_latents),
+            )
+        return CogVideoXDDIMInversionOutput(
+            inverse_latents=inverse_latents,
+            recon_latents=recon_latents,
+        )
+
+
+if __name__ == "__main__":
+    arguments = get_args()
+    pipeline = CogVideoXPipelineForDDIMInversion.from_pretrained(
+        arguments.pop("model_path"),
+        torch_dtype=arguments.pop("dtype"),
+    ).to(device=arguments.pop("device"))
+
+    output_path = arguments.pop("output_path")
+    fps = arguments.pop("fps")
+    inverse_video_path = os.path.join(output_path, f"{arguments.get('video_path')}_inversion.mp4")
+    recon_video_path = os.path.join(output_path, f"{arguments.get('video_path')}_reconstruction.mp4")
+
+    # Run DDIM inversion
+    output = pipeline(**arguments)
+    pipeline.export_latents_to_video(output.inverse_latents[-1], inverse_video_path, fps)
+    pipeline.export_latents_to_video(output.recon_latents[-1], recon_video_path, fps)
@@ -686,7 +686,7 @@ class StableDiffusionHDPainterPipeline(StableDiffusionInpaintPipeline):
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -362,7 +362,7 @@ class ImageToImageInpaintingPipeline(DiffusionPipeline):
                f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                " `pipeline.unet` or your `mask_image` or `image` input."
            )

@@ -1120,7 +1120,7 @@ class LLMGroundedDiffusionPipeline(

            if verbose:
                logger.info(
-                    f"time index {index}, loss: {loss.item()/loss_scale:.3f} (de-scaled with scale {loss_scale:.1f}), loss threshold: {loss_threshold:.3f}"
+                    f"time index {index}, loss: {loss.item() / loss_scale:.3f} (de-scaled with scale {loss_scale:.1f}), loss threshold: {loss_threshold:.3f}"
                )

            try:
@@ -1184,7 +1184,7 @@ class LLMGroundedDiffusionPipeline(

                    if verbose:
                        logger.info(
-                            f"time index {index}, loss: {loss.item()/loss_scale:.3f}, loss threshold: {loss_threshold:.3f}, iteration: {iteration}"
+                            f"time index {index}, loss: {loss.item() / loss_scale:.3f}, loss threshold: {loss_threshold:.3f}, iteration: {iteration}"
                        )

            finally:
@@ -1773,7 +1773,7 @@ class SDXLLongPromptWeightingPipeline(
                        f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                        f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                        f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                        f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                        f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                        " `pipeline.unet` or your `mask_image` or `image` input."
                    )
            elif num_channels_unet != 4:
@@ -1924,7 +1924,22 @@ class SDXLLongPromptWeightingPipeline(
                self.upcast_vae()
                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)

-            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]

            # cast back to fp16 if needed
            if needs_upcasting:
@@ -43,7 +43,7 @@ from diffusers.utils import BaseOutput, check_min_version


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")


 class MarigoldDepthOutput(BaseOutput):
@@ -1,4 +1,4 @@
-# Copyright 2025 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The DEVAIEXP Team and The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -1070,32 +1070,32 @@ class StableDiffusionXLTilingPipeline(
                    text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
                else:
                    text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
-                    add_time_ids = self._get_add_time_ids(
-                        original_size,
-                        crops_coords_top_left[row][col],
-                        target_size,
+                add_time_ids = self._get_add_time_ids(
+                    original_size,
+                    crops_coords_top_left[row][col],
+                    target_size,
+                    dtype=prompt_embeds.dtype,
+                    text_encoder_projection_dim=text_encoder_projection_dim,
+                )
+                if negative_original_size is not None and negative_target_size is not None:
+                    negative_add_time_ids = self._get_add_time_ids(
+                        negative_original_size,
+                        negative_crops_coords_top_left[row][col],
+                        negative_target_size,
                        dtype=prompt_embeds.dtype,
                        text_encoder_projection_dim=text_encoder_projection_dim,
                    )
-                    if negative_original_size is not None and negative_target_size is not None:
-                        negative_add_time_ids = self._get_add_time_ids(
-                            negative_original_size,
-                            negative_crops_coords_top_left[row][col],
-                            negative_target_size,
-                            dtype=prompt_embeds.dtype,
-                            text_encoder_projection_dim=text_encoder_projection_dim,
-                        )
-                    else:
-                        negative_add_time_ids = add_time_ids
+                else:
+                    negative_add_time_ids = add_time_ids

-                    if self.do_classifier_free_guidance:
-                        prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
-                        add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
-                        add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+                if self.do_classifier_free_guidance:
+                    prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+                    add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+                    add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)

-                    prompt_embeds = prompt_embeds.to(device)
-                    add_text_embeds = add_text_embeds.to(device)
-                    add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+                prompt_embeds = prompt_embeds.to(device)
+                add_text_embeds = add_text_embeds.to(device)
+                add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
                addition_embed_type_row.append((prompt_embeds, add_text_embeds, add_time_ids))
            embeddings_and_added_time.append(addition_embed_type_row)

@@ -488,7 +488,7 @@ class FluxDifferentialImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
        if padding_mask_crop is not None:
            if not isinstance(image, PIL.Image.Image):
                raise ValueError(
-                    f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
+                    f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}."
                )
            if not isinstance(mask_image, PIL.Image.Image):
                raise ValueError(
@@ -496,7 +496,7 @@ class FluxDifferentialImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                    f" {type(mask_image)}."
                )
            if output_type != "pil":
-                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
+                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.")

        if max_sequence_length is not None and max_sequence_length > 512:
            raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")
@@ -907,12 +907,12 @@ def create_controller(

    # reweight
    if edit_type == "reweight":
-        assert (
-            equalizer_words is not None and equalizer_strengths is not None
-        ), "To use reweight edit, please specify equalizer_words and equalizer_strengths."
-        assert len(equalizer_words) == len(
-            equalizer_strengths
-        ), "equalizer_words and equalizer_strengths must be of same length."
+        assert equalizer_words is not None and equalizer_strengths is not None, (
+            "To use reweight edit, please specify equalizer_words and equalizer_strengths."
+        )
+        assert len(equalizer_words) == len(equalizer_strengths), (
+            "equalizer_words and equalizer_strengths must be of same length."
+        )
        equalizer = get_equalizer(prompts[1], equalizer_words, equalizer_strengths, tokenizer=tokenizer)
        return AttentionReweight(
            prompts,
@@ -1738,7 +1738,7 @@ class StyleAlignedSDXLPipeline(
                        f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                        f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                        f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                        f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                        f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                        " `pipeline.unet` or your `mask_image` or `image` input."
                    )
            elif num_channels_unet != 4:
@@ -689,7 +689,7 @@ class StableDiffusionUpscaleLDM3DPipeline(
                f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                f" `num_channels_image`: {num_channels_image} "
-                f" = {num_channels_latents+num_channels_image}. Please verify the config of"
+                f" = {num_channels_latents + num_channels_image}. Please verify the config of"
                " `pipeline.unet` or your `image` input."
            )

@@ -1028,7 +1028,7 @@ class StableDiffusionXL_AE_Pipeline(
        if padding_mask_crop is not None:
            if not isinstance(image, PIL.Image.Image):
                raise ValueError(
-                    f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
+                    f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}."
                )
            if not isinstance(mask_image, PIL.Image.Image):
                raise ValueError(
@@ -1036,7 +1036,7 @@ class StableDiffusionXL_AE_Pipeline(
                    f" {type(mask_image)}."
                )
            if output_type != "pil":
-                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
+                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.")

        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
            raise ValueError(
@@ -2050,7 +2050,7 @@ class StableDiffusionXL_AE_Pipeline(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -1578,7 +1578,7 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -0,0 +1,876 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import CogVideoXLoraLoaderMixin
+from diffusers.models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
+from diffusers.models.embeddings import get_3d_rotary_pos_embed
+from diffusers.pipelines.cogvideo.pipeline_output import CogVideoXPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from examples.community.pipeline_stg_cogvideox import CogVideoXSTGPipeline
+
+        >>> # Models: "THUDM/CogVideoX-2b" or "THUDM/CogVideoX-5b"
+        >>> pipe = CogVideoXSTGPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.float16).to("cuda")
+        >>> prompt = (
+        ...     "A father and son building a treehouse together, their hands covered in sawdust and smiles on their faces, realistic style."
+        ... )
+        >>> pipe.transformer.to(memory_format=torch.channels_last)
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [11]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0  # Set to 0.0 for CFG
+        >>> do_rescaling = False
+
+        >>> # Generate video frames with STG parameters
+        >>> frames = pipe(
+        ...     prompt=prompt,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ...     do_rescaling=do_rescaling,
+        >>> ).frames[0]
+        >>> export_to_video(frames, "output.mp4", fps=8)
+        ```
+"""
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+) -> torch.Tensor:
+    hidden_states_ptb = hidden_states[2:]
+    encoder_hidden_states_ptb = encoder_hidden_states[2:]
+
+    text_seq_length = encoder_hidden_states.size(1)
+
+    # norm & modulate
+    norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
+        hidden_states, encoder_hidden_states, temb
+    )
+
+    # attention
+    attn_hidden_states, attn_encoder_hidden_states = self.attn1(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=norm_encoder_hidden_states,
+        image_rotary_emb=image_rotary_emb,
+    )
+
+    hidden_states = hidden_states + gate_msa * attn_hidden_states
+    encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder_hidden_states
+
+    # norm & modulate
+    norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
+        hidden_states, encoder_hidden_states, temb
+    )
+
+    # feed-forward
+    norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
+    ff_output = self.ff(norm_hidden_states)
+
+    hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
+    encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
+
+    hidden_states[2:] = hidden_states_ptb
+    encoder_hidden_states[2:] = encoder_hidden_states_ptb
+
+    return hidden_states, encoder_hidden_states
+
+
+# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
+def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
+    tw = tgt_width
+    th = tgt_height
+    h, w = src
+    r = h / w
+    if r > (th / tw):
+        resize_height = th
+        resize_width = int(round(th / h * w))
+    else:
+        resize_width = tw
+        resize_height = int(round(tw / w * h))
+
+    crop_top = int(round((th - resize_height) / 2.0))
+    crop_left = int(round((tw - resize_width) / 2.0))
+
+    return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class CogVideoXSTGPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation using CogVideoX.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. CogVideoX uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
+            [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`CogVideoXTransformer3DModel`]):
+            A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
+    """
+
+    _optional_components = []
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+    ]
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+        self.vae_scale_factor_spatial = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
+        )
+        self.vae_scale_factor_temporal = (
+            self.vae.config.temporal_compression_ratio if getattr(self, "vae", None) else 4
+        )
+        self.vae_scaling_factor_image = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.7
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+    ):
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        shape = (
+            batch_size,
+            (num_frames - 1) // self.vae_scale_factor_temporal + 1,
+            num_channels_latents,
+            height // self.vae_scale_factor_spatial,
+            width // self.vae_scale_factor_spatial,
+        )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
+        latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
+        latents = 1 / self.vae_scaling_factor_image * latents
+
+        frames = self.vae.decode(latents).sample
+        return frames
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_on_step_end_tensor_inputs,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    def fuse_qkv_projections(self) -> None:
+        r"""Enables fused QKV projections."""
+        self.fusing_transformer = True
+        self.transformer.fuse_qkv_projections()
+
+    def unfuse_qkv_projections(self) -> None:
+        r"""Disable QKV projection fusion if enabled."""
+        if not self.fusing_transformer:
+            logger.warning("The Transformer was not initially fused for QKV projections. Doing nothing.")
+        else:
+            self.transformer.unfuse_qkv_projections()
+            self.fusing_transformer = False
+
+    def _prepare_rotary_positional_embeddings(
+        self,
+        height: int,
+        width: int,
+        num_frames: int,
+        device: torch.device,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        grid_height = height // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+
+        p = self.transformer.config.patch_size
+        p_t = self.transformer.config.patch_size_t
+
+        base_size_width = self.transformer.config.sample_width // p
+        base_size_height = self.transformer.config.sample_height // p
+
+        if p_t is None:
+            # CogVideoX 1.0
+            grid_crops_coords = get_resize_crop_region_for_grid(
+                (grid_height, grid_width), base_size_width, base_size_height
+            )
+            freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
+                embed_dim=self.transformer.config.attention_head_dim,
+                crops_coords=grid_crops_coords,
+                grid_size=(grid_height, grid_width),
+                temporal_size=num_frames,
+                device=device,
+            )
+        else:
+            # CogVideoX 1.5
+            base_num_frames = (num_frames + p_t - 1) // p_t
+
+            freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
+                embed_dim=self.transformer.config.attention_head_dim,
+                crops_coords=None,
+                grid_size=(grid_height, grid_width),
+                temporal_size=base_num_frames,
+                grid_type="slice",
+                max_size=(base_size_height, base_size_width),
+                device=device,
+            )
+
+        return freqs_cos, freqs_sin
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_frames: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: Optional[List[int]] = None,
+        guidance_scale: float = 6,
+        use_dynamic_cfg: bool = False,
+        num_videos_per_prompt: int = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: str = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 226,
+        stg_applied_layers_idx: Optional[List[int]] = [11],
+        stg_scale: Optional[float] = 0.0,
+        do_rescaling: Optional[bool] = False,
+    ) -> Union[CogVideoXPipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            height (`int`, *optional*, defaults to self.transformer.config.sample_height * self.vae_scale_factor_spatial):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, *optional*, defaults to self.transformer.config.sample_height * self.vae_scale_factor_spatial):
+                The width in pixels of the generated image. This is set to 720 by default for the best results.
+            num_frames (`int`, defaults to `48`):
+                Number of frames to generate. Must be divisible by self.vae_scale_factor_temporal. Generated video will
+                contain 1 extra frame because CogVideoX is conditioned with (num_seconds * fps + 1) frames where
+                num_seconds is 6 and fps is 8. However, since videos can be saved at any fps, the only condition that
+                needs to be satisfied is that of divisibility mentioned above.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, defaults to `226`):
+                Maximum sequence length in encoded prompt. Must be consistent with
+                `self.transformer.config.max_text_seq_length` otherwise may lead to poor results.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] or `tuple`:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        height = height or self.transformer.config.sample_height * self.vae_scale_factor_spatial
+        width = width or self.transformer.config.sample_width * self.vae_scale_factor_spatial
+        num_frames = num_frames or self.transformer.config.sample_frames
+
+        num_videos_per_prompt = 1
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_on_step_end_tensor_inputs,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+        self._stg_scale = stg_scale
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        if self.do_spatio_temporal_guidance:
+            for i in stg_applied_layers_idx:
+                self.transformer.transformer_blocks[i].forward = types.MethodType(
+                    forward_with_stg, self.transformer.transformer_blocks[i]
+                )
+
+        # 2. Default call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+        elif do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latents
+        latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+
+        # For CogVideoX 1.5, the latent frames should be padded to make it divisible by patch_size_t
+        patch_size_t = self.transformer.config.patch_size_t
+        additional_frames = 0
+        if patch_size_t is not None and latent_frames % patch_size_t != 0:
+            additional_frames = patch_size_t - latent_frames % patch_size_t
+            num_frames += additional_frames * self.vae_scale_factor_temporal
+
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            latent_channels,
+            num_frames,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Create rotary embeds if required
+        image_rotary_emb = (
+            self._prepare_rotary_positional_embeddings(height, width, latents.size(1), device)
+            if self.transformer.config.use_rotary_positional_embeddings
+            else None
+        )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            # for DPM-solver++
+            old_pred_original_sample = None
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                if do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 2)
+                elif do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 3)
+                else:
+                    latent_model_input = latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    image_rotary_emb=image_rotary_emb,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                # perform guidance
+                if use_dynamic_cfg:
+                    self._guidance_scale = 1 + guidance_scale * (
+                        (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
+                    )
+                if do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                elif do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
+                    noise_pred = (
+                        noise_pred_uncond
+                        + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                        + self._stg_scale * (noise_pred_text - noise_pred_perturb)
+                    )
+
+                if do_rescaling:
+                    rescaling_scale = 0.7
+                    factor = noise_pred_text.std() / noise_pred.std()
+                    factor = rescaling_scale * factor + (1 - rescaling_scale)
+                    noise_pred = noise_pred * factor
+
+                # compute the previous noisy sample x_t -> x_t-1
+                if not isinstance(self.scheduler, CogVideoXDPMScheduler):
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                else:
+                    latents, old_pred_original_sample = self.scheduler.step(
+                        noise_pred,
+                        old_pred_original_sample,
+                        t,
+                        timesteps[i - 1] if i > 0 else None,
+                        latents,
+                        **extra_step_kwargs,
+                        return_dict=False,
+                    )
+                latents = latents.to(prompt_embeds.dtype)
+
+                # call the callback, if provided
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            # Discard any padding frames that were added for CogVideoX 1.5
+            latents = latents[:, additional_frames:]
+            video = self.decode_latents(latents)
+            video = self.video_processor.postprocess_video(video=video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return CogVideoXPipelineOutput(frames=video)
@@ -0,0 +1,794 @@
+# Copyright 2024 The HunyuanVideo Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer, LlamaModel, LlamaTokenizerFast
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import HunyuanVideoLoraLoaderMixin
+from diffusers.models import AutoencoderKLHunyuanVideo, HunyuanVideoTransformer3DModel
+from diffusers.pipelines.hunyuan_video.pipeline_output import HunyuanVideoPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from diffusers import HunyuanVideoTransformer3DModel
+        >>> from examples.community.pipeline_stg_hunyuan_video import HunyuanVideoSTGPipeline
+
+        >>> model_id = "hunyuanvideo-community/HunyuanVideo"
+        >>> transformer = HunyuanVideoTransformer3DModel.from_pretrained(
+        ...     model_id, subfolder="transformer", torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe = HunyuanVideoSTGPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch.float16)
+        >>> pipe.vae.enable_tiling()
+        >>> pipe.to("cuda")
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [2]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+
+        >>> output = pipe(
+        ...     prompt="A wolf howling at the moon, with the moon subtly resembling a giant clock face, realistic style.",
+        ...     height=320,
+        ...     width=512,
+        ...     num_frames=61,
+        ...     num_inference_steps=30,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        >>> ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=15)
+        ```
+"""
+
+
+DEFAULT_PROMPT_TEMPLATE = {
+    "template": (
+        "<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: "
+        "1. The main content and theme of the video."
+        "2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects."
+        "3. Actions, events, behaviors temporal relationships, physical movement changes of the objects."
+        "4. background environment, light, style and atmosphere."
+        "5. camera angles, movements, and transitions used in the video:<|eot_id|>"
+        "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>"
+    ),
+    "crop_start": 95,
+}
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    return hidden_states, encoder_hidden_states
+
+
+def forward_without_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    # 1. Input normalization
+    norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
+    norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+        encoder_hidden_states, emb=temb
+    )
+
+    # 2. Joint attention
+    attn_output, context_attn_output = self.attn(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=norm_encoder_hidden_states,
+        attention_mask=attention_mask,
+        image_rotary_emb=freqs_cis,
+    )
+
+    # 3. Modulation and residual connection
+    hidden_states = hidden_states + attn_output * gate_msa.unsqueeze(1)
+    encoder_hidden_states = encoder_hidden_states + context_attn_output * c_gate_msa.unsqueeze(1)
+
+    norm_hidden_states = self.norm2(hidden_states)
+    norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+
+    norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+    norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+
+    # 4. Feed-forward
+    ff_output = self.ff(norm_hidden_states)
+    context_ff_output = self.ff_context(norm_encoder_hidden_states)
+
+    hidden_states = hidden_states + gate_mlp.unsqueeze(1) * ff_output
+    encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+
+    return hidden_states, encoder_hidden_states
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class HunyuanVideoSTGPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation using HunyuanVideo.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        text_encoder ([`LlamaModel`]):
+            [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers).
+        tokenizer (`LlamaTokenizer`):
+            Tokenizer from [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers).
+        transformer ([`HunyuanVideoTransformer3DModel`]):
+            Conditional Transformer to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLHunyuanVideo`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder_2 ([`CLIPTextModel`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer_2 (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        text_encoder: LlamaModel,
+        tokenizer: LlamaTokenizerFast,
+        transformer: HunyuanVideoTransformer3DModel,
+        vae: AutoencoderKLHunyuanVideo,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        text_encoder_2: CLIPTextModel,
+        tokenizer_2: CLIPTokenizer,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+            text_encoder_2=text_encoder_2,
+            tokenizer_2=tokenizer_2,
+        )
+
+        self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_llama_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_template: Dict[str, Any],
+        num_videos_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 256,
+        num_hidden_layers_to_skip: int = 2,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        prompt = [prompt_template["template"].format(p) for p in prompt]
+
+        crop_start = prompt_template.get("crop_start", None)
+        if crop_start is None:
+            prompt_template_input = self.tokenizer(
+                prompt_template["template"],
+                padding="max_length",
+                return_tensors="pt",
+                return_length=False,
+                return_overflowing_tokens=False,
+                return_attention_mask=False,
+            )
+            crop_start = prompt_template_input["input_ids"].shape[-1]
+            # Remove <|eot_id|> token and placeholder {}
+            crop_start -= 2
+
+        max_sequence_length += crop_start
+        text_inputs = self.tokenizer(
+            prompt,
+            max_length=max_sequence_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt",
+            return_length=False,
+            return_overflowing_tokens=False,
+            return_attention_mask=True,
+        )
+        text_input_ids = text_inputs.input_ids.to(device=device)
+        prompt_attention_mask = text_inputs.attention_mask.to(device=device)
+
+        prompt_embeds = self.text_encoder(
+            input_ids=text_input_ids,
+            attention_mask=prompt_attention_mask,
+            output_hidden_states=True,
+        ).hidden_states[-(num_hidden_layers_to_skip + 1)]
+        prompt_embeds = prompt_embeds.to(dtype=dtype)
+
+        if crop_start is not None and crop_start > 0:
+            prompt_embeds = prompt_embeds[:, crop_start:]
+            prompt_attention_mask = prompt_attention_mask[:, crop_start:]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(1, num_videos_per_prompt)
+        prompt_attention_mask = prompt_attention_mask.view(batch_size * num_videos_per_prompt, seq_len)
+
+        return prompt_embeds, prompt_attention_mask
+
+    def _get_clip_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        num_videos_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 77,
+    ) -> torch.Tensor:
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder_2.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer_2(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False).pooler_output
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_2: Union[str, List[str]] = None,
+        prompt_template: Dict[str, Any] = DEFAULT_PROMPT_TEMPLATE,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 256,
+    ):
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds(
+                prompt,
+                prompt_template,
+                num_videos_per_prompt,
+                device=device,
+                dtype=dtype,
+                max_sequence_length=max_sequence_length,
+            )
+
+        if pooled_prompt_embeds is None:
+            if prompt_2 is None and pooled_prompt_embeds is None:
+                prompt_2 = prompt
+            pooled_prompt_embeds = self._get_clip_prompt_embeds(
+                prompt,
+                num_videos_per_prompt,
+                device=device,
+                dtype=dtype,
+                max_sequence_length=77,
+            )
+
+        return prompt_embeds, pooled_prompt_embeds, prompt_attention_mask
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        height,
+        width,
+        prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_template=None,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+        if prompt_template is not None:
+            if not isinstance(prompt_template, dict):
+                raise ValueError(f"`prompt_template` has to be of type `dict` but is {type(prompt_template)}")
+            if "template" not in prompt_template:
+                raise ValueError(
+                    f"`prompt_template` has to contain a key `template` but only found {prompt_template.keys()}"
+                )
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: 32,
+        height: int = 720,
+        width: int = 1280,
+        num_frames: int = 129,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_frames,
+            int(height) // self.vae_scale_factor_spatial,
+            int(width) // self.vae_scale_factor_spatial,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Union[str, List[str]] = None,
+        height: int = 720,
+        width: int = 1280,
+        num_frames: int = 129,
+        num_inference_steps: int = 50,
+        sigmas: List[float] = None,
+        guidance_scale: float = 6.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        prompt_template: Dict[str, Any] = DEFAULT_PROMPT_TEMPLATE,
+        max_sequence_length: int = 256,
+        stg_applied_layers_idx: Optional[List[int]] = [2],
+        stg_scale: Optional[float] = 0.0,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                will be used instead.
+            height (`int`, defaults to `720`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `1280`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `129`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, defaults to `6.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality. Note that the only available HunyuanVideo model is
+                CFG-distilled, which means that traditional guidance between unconditional and conditional latent is
+                not applied.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`HunyuanVideoPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~HunyuanVideoPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`HunyuanVideoPipelineOutput`] is returned, otherwise a `tuple` is returned
+                where the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+            prompt_template,
+        )
+
+        self._stg_scale = stg_scale
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # 3. Encode input prompt
+        prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_template=prompt_template,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            device=device,
+            max_sequence_length=max_sequence_length,
+        )
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        prompt_attention_mask = prompt_attention_mask.to(transformer_dtype)
+        if pooled_prompt_embeds is not None:
+            pooled_prompt_embeds = pooled_prompt_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+        )
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_latent_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare guidance condition
+        guidance = torch.tensor([guidance_scale] * latents.shape[0], dtype=transformer_dtype, device=device) * 1000.0
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                latent_model_input = latents.to(transformer_dtype)
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                if self.do_spatio_temporal_guidance:
+                    for i in stg_applied_layers_idx:
+                        self.transformer.transformer_blocks[i].forward = types.MethodType(
+                            forward_without_stg, self.transformer.transformer_blocks[i]
+                        )
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_attention_mask=prompt_attention_mask,
+                    pooled_projections=pooled_prompt_embeds,
+                    guidance=guidance,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_spatio_temporal_guidance:
+                    for i in stg_applied_layers_idx:
+                        self.transformer.transformer_blocks[i].forward = types.MethodType(
+                            forward_with_stg, self.transformer.transformer_blocks[i]
+                        )
+
+                    noise_pred_perturb = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=prompt_embeds,
+                        encoder_attention_mask=prompt_attention_mask,
+                        pooled_projections=pooled_prompt_embeds,
+                        guidance=guidance,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_pred + self._stg_scale * (noise_pred - noise_pred_perturb)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return HunyuanVideoPipelineOutput(frames=video)
@@ -0,0 +1,886 @@
+# Copyright 2024 Lightricks and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin
+from diffusers.models.autoencoders import AutoencoderKLLTXVideo
+from diffusers.models.transformers import LTXVideoTransformer3DModel
+from diffusers.pipelines.ltx.pipeline_output import LTXPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from examples.community.pipeline_stg_ltx import LTXSTGPipeline
+
+        >>> pipe = LTXSTGPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A woman with light skin, wearing a blue jacket and a black hat with a veil, looks down and to her right, then back up as she speaks; she has brown hair styled in an updo, light brown eyebrows, and is wearing a white collared shirt under her jacket; the camera remains stationary on her face as she speaks; the background is out of focus, but shows trees and people in period clothing; the scene is captured in real-life footage."
+        >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [19]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+        >>> do_rescaling = False
+
+        >>> video = pipe(
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     width=704,
+        ...     height=480,
+        ...     num_frames=161,
+        ...     num_inference_steps=50,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ...     do_rescaling=do_rescaling,
+        >>> ).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=24)
+        ```
+"""
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    encoder_attention_mask: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    hidden_states_ptb = hidden_states[2:]
+    encoder_hidden_states_ptb = encoder_hidden_states[2:]
+
+    batch_size = hidden_states.size(0)
+    norm_hidden_states = self.norm1(hidden_states)
+
+    num_ada_params = self.scale_shift_table.shape[0]
+    ada_values = self.scale_shift_table[None, None] + temb.reshape(batch_size, temb.size(1), num_ada_params, -1)
+    shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ada_values.unbind(dim=2)
+    norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+
+    attn_hidden_states = self.attn1(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=None,
+        image_rotary_emb=image_rotary_emb,
+    )
+    hidden_states = hidden_states + attn_hidden_states * gate_msa
+
+    attn_hidden_states = self.attn2(
+        hidden_states,
+        encoder_hidden_states=encoder_hidden_states,
+        image_rotary_emb=None,
+        attention_mask=encoder_attention_mask,
+    )
+    hidden_states = hidden_states + attn_hidden_states
+    norm_hidden_states = self.norm2(hidden_states) * (1 + scale_mlp) + shift_mlp
+
+    ff_output = self.ff(norm_hidden_states)
+    hidden_states = hidden_states + ff_output * gate_mlp
+
+    hidden_states[2:] = hidden_states_ptb
+    encoder_hidden_states[2:] = encoder_hidden_states_ptb
+
+    return hidden_states
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class LTXSTGPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation.
+
+    Reference: https://github.com/Lightricks/LTX-Video
+
+    Args:
+        transformer ([`LTXVideoTransformer3DModel`]):
+            Conditional Transformer architecture to denoise the encoded video latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLLTXVideo`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLLTXVideo,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: LTXVideoTransformer3DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_spatial_compression_ratio = (
+            self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32
+        )
+        self.vae_temporal_compression_ratio = (
+            self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8
+        )
+        self.transformer_spatial_patch_size = (
+            self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1
+        )
+        self.transformer_temporal_patch_size = (
+            self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128
+        )
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 32 != 0 or width % 32 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    @staticmethod
+    def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor:
+        # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p].
+        # The patch dimensions are then permuted and collapsed into the channel dimension of shape:
+        # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor).
+        # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features
+        batch_size, num_channels, num_frames, height, width = latents.shape
+        post_patch_num_frames = num_frames // patch_size_t
+        post_patch_height = height // patch_size
+        post_patch_width = width // patch_size
+        latents = latents.reshape(
+            batch_size,
+            -1,
+            post_patch_num_frames,
+            patch_size_t,
+            post_patch_height,
+            patch_size,
+            post_patch_width,
+            patch_size,
+        )
+        latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
+        return latents
+
+    @staticmethod
+    def _unpack_latents(
+        latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1
+    ) -> torch.Tensor:
+        # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions)
+        # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of
+        # what happens in the `_pack_latents` method.
+        batch_size = latents.size(0)
+        latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size)
+        latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3)
+        return latents
+
+    @staticmethod
+    def _normalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Normalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = (latents - latents_mean) * scaling_factor / latents_std
+        return latents
+
+    @staticmethod
+    def _denormalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Denormalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = latents * latents_std / scaling_factor + latents_mean
+        return latents
+
+    def prepare_latents(
+        self,
+        batch_size: int = 1,
+        num_channels_latents: int = 128,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        height = height // self.vae_spatial_compression_ratio
+        width = width // self.vae_spatial_compression_ratio
+        num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
+
+        shape = (batch_size, num_channels_latents, num_frames, height, width)
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        latents = self._pack_latents(
+            latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        )
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        frame_rate: int = 25,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 3,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        decode_timestep: Union[float, List[float]] = 0.0,
+        decode_noise_scale: Optional[Union[float, List[float]]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 128,
+        stg_applied_layers_idx: Optional[List[int]] = [19],
+        stg_scale: Optional[float] = 1.0,
+        do_rescaling: Optional[bool] = False,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `512`):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, defaults to `704`):
+                The width in pixels of the generated image. This is set to 848 by default for the best results.
+            num_frames (`int`, defaults to `161`):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, defaults to `3 `):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            decode_timestep (`float`, defaults to `0.0`):
+                The timestep at which generated video is decoded.
+            decode_noise_scale (`float`, defaults to `None`):
+                The interpolation factor between random noise and denoised latents at the decode timestep.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to `128 `):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            height=height,
+            width=width,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+        )
+
+        self._stg_scale = stg_scale
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._interrupt = False
+
+        if self.do_spatio_temporal_guidance:
+            for i in stg_applied_layers_idx:
+                self.transformer.transformer_blocks[i].forward = types.MethodType(
+                    forward_with_stg, self.transformer.transformer_blocks[i]
+                )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Prepare text embeddings
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+        elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat(
+                [negative_prompt_attention_mask, prompt_attention_mask, prompt_attention_mask], dim=0
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 5. Prepare timesteps
+        latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
+        latent_height = height // self.vae_spatial_compression_ratio
+        latent_width = width // self.vae_spatial_compression_ratio
+        video_sequence_length = latent_num_frames * latent_height * latent_width
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        mu = calculate_shift(
+            video_sequence_length,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.16),
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Prepare micro-conditions
+        latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
+        rope_interpolation_scale = (
+            1 / latent_frame_rate,
+            self.vae_spatial_compression_ratio,
+            self.vae_spatial_compression_ratio,
+        )
+
+        # 7. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 2)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 3)
+                else:
+                    latent_model_input = latents
+
+                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    encoder_attention_mask=prompt_attention_mask,
+                    num_frames=latent_num_frames,
+                    height=latent_height,
+                    width=latent_width,
+                    rope_interpolation_scale=rope_interpolation_scale,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
+                    noise_pred = (
+                        noise_pred_uncond
+                        + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                        + self._stg_scale * (noise_pred_text - noise_pred_perturb)
+                    )
+
+                if do_rescaling:
+                    rescaling_scale = 0.7
+                    factor = noise_pred_text.std() / noise_pred.std()
+                    factor = rescaling_scale * factor + (1 - rescaling_scale)
+                    noise_pred = noise_pred * factor
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            video = latents
+        else:
+            latents = self._unpack_latents(
+                latents,
+                latent_num_frames,
+                latent_height,
+                latent_width,
+                self.transformer_spatial_patch_size,
+                self.transformer_temporal_patch_size,
+            )
+            latents = self._denormalize_latents(
+                latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+            )
+            latents = latents.to(prompt_embeds.dtype)
+
+            if not self.vae.config.timestep_conditioning:
+                timestep = None
+            else:
+                noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype)
+                if not isinstance(decode_timestep, list):
+                    decode_timestep = [decode_timestep] * batch_size
+                if decode_noise_scale is None:
+                    decode_noise_scale = decode_timestep
+                elif not isinstance(decode_noise_scale, list):
+                    decode_noise_scale = [decode_noise_scale] * batch_size
+
+                timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype)
+                decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[
+                    :, None, None, None, None
+                ]
+                latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise
+
+            video = self.vae.decode(latents, timestep, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return LTXPipelineOutput(frames=video)
@@ -0,0 +1,985 @@
+# Copyright 2024 Lightricks and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.image_processor import PipelineImageInput
+from diffusers.loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin
+from diffusers.models.autoencoders import AutoencoderKLLTXVideo
+from diffusers.models.transformers import LTXVideoTransformer3DModel
+from diffusers.pipelines.ltx.pipeline_output import LTXPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers.utils import export_to_video, load_image
+        >>> from examples.community.pipeline_stg_ltx_image2video import LTXImageToVideoSTGPipeline
+
+        >>> pipe = LTXImageToVideoSTGPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/a-r-r-o-w/tiny-meme-dataset-captioned/resolve/main/images/11.png"
+        >>> )
+        >>> prompt = "A medieval fantasy scene featuring a rugged man with shoulder-length brown hair and a beard. He wears a dark leather tunic over a maroon shirt with intricate metal details. His facial expression is serious and intense, and he is making a gesture with his right hand, forming a small circle with his thumb and index finger. The warm golden lighting casts dramatic shadows on his face. The background includes an ornate stone arch and blurred medieval-style decor, creating an epic atmosphere."
+        >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [19]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+        >>> do_rescaling = False
+
+        >>> video = pipe(
+        ...     image=image,
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     width=704,
+        ...     height=480,
+        ...     num_frames=161,
+        ...     num_inference_steps=50,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ...     do_rescaling=do_rescaling,
+        >>> ).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=24)
+        ```
+"""
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    encoder_attention_mask: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    hidden_states_ptb = hidden_states[2:]
+    encoder_hidden_states_ptb = encoder_hidden_states[2:]
+
+    batch_size = hidden_states.size(0)
+    norm_hidden_states = self.norm1(hidden_states)
+
+    num_ada_params = self.scale_shift_table.shape[0]
+    ada_values = self.scale_shift_table[None, None] + temb.reshape(batch_size, temb.size(1), num_ada_params, -1)
+    shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ada_values.unbind(dim=2)
+    norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+
+    attn_hidden_states = self.attn1(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=None,
+        image_rotary_emb=image_rotary_emb,
+    )
+    hidden_states = hidden_states + attn_hidden_states * gate_msa
+
+    attn_hidden_states = self.attn2(
+        hidden_states,
+        encoder_hidden_states=encoder_hidden_states,
+        image_rotary_emb=None,
+        attention_mask=encoder_attention_mask,
+    )
+    hidden_states = hidden_states + attn_hidden_states
+    norm_hidden_states = self.norm2(hidden_states) * (1 + scale_mlp) + shift_mlp
+
+    ff_output = self.ff(norm_hidden_states)
+    hidden_states = hidden_states + ff_output * gate_mlp
+
+    hidden_states[2:] = hidden_states_ptb
+    encoder_hidden_states[2:] = encoder_hidden_states_ptb
+
+    return hidden_states
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class LTXImageToVideoSTGPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin):
+    r"""
+    Pipeline for image-to-video generation.
+
+    Reference: https://github.com/Lightricks/LTX-Video
+
+    Args:
+        transformer ([`LTXVideoTransformer3DModel`]):
+            Conditional Transformer architecture to denoise the encoded video latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLLTXVideo`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLLTXVideo,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: LTXVideoTransformer3DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_spatial_compression_ratio = (
+            self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32
+        )
+        self.vae_temporal_compression_ratio = (
+            self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8
+        )
+        self.transformer_spatial_patch_size = (
+            self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1
+        )
+        self.transformer_temporal_patch_size = (
+            self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128
+        )
+
+        self.default_height = 512
+        self.default_width = 704
+        self.default_frames = 121
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 32 != 0 or width % 32 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._pack_latents
+    def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor:
+        # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p].
+        # The patch dimensions are then permuted and collapsed into the channel dimension of shape:
+        # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor).
+        # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features
+        batch_size, num_channels, num_frames, height, width = latents.shape
+        post_patch_num_frames = num_frames // patch_size_t
+        post_patch_height = height // patch_size
+        post_patch_width = width // patch_size
+        latents = latents.reshape(
+            batch_size,
+            -1,
+            post_patch_num_frames,
+            patch_size_t,
+            post_patch_height,
+            patch_size,
+            post_patch_width,
+            patch_size,
+        )
+        latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._unpack_latents
+    def _unpack_latents(
+        latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1
+    ) -> torch.Tensor:
+        # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions)
+        # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of
+        # what happens in the `_pack_latents` method.
+        batch_size = latents.size(0)
+        latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size)
+        latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3)
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents
+    def _normalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Normalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = (latents - latents_mean) * scaling_factor / latents_std
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents
+    def _denormalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Denormalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = latents * latents_std / scaling_factor + latents_mean
+        return latents
+
+    def prepare_latents(
+        self,
+        image: Optional[torch.Tensor] = None,
+        batch_size: int = 1,
+        num_channels_latents: int = 128,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        height = height // self.vae_spatial_compression_ratio
+        width = width // self.vae_spatial_compression_ratio
+        num_frames = (
+            (num_frames - 1) // self.vae_temporal_compression_ratio + 1 if latents is None else latents.size(2)
+        )
+
+        shape = (batch_size, num_channels_latents, num_frames, height, width)
+        mask_shape = (batch_size, 1, num_frames, height, width)
+
+        if latents is not None:
+            conditioning_mask = latents.new_zeros(shape)
+            conditioning_mask[:, :, 0] = 1.0
+            conditioning_mask = self._pack_latents(
+                conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+            )
+            return latents.to(device=device, dtype=dtype), conditioning_mask
+
+        if isinstance(generator, list):
+            if len(generator) != batch_size:
+                raise ValueError(
+                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                )
+
+            init_latents = [
+                retrieve_latents(self.vae.encode(image[i].unsqueeze(0).unsqueeze(2)), generator[i])
+                for i in range(batch_size)
+            ]
+        else:
+            init_latents = [
+                retrieve_latents(self.vae.encode(img.unsqueeze(0).unsqueeze(2)), generator) for img in image
+            ]
+
+        init_latents = torch.cat(init_latents, dim=0).to(dtype)
+        init_latents = self._normalize_latents(init_latents, self.vae.latents_mean, self.vae.latents_std)
+        init_latents = init_latents.repeat(1, 1, num_frames, 1, 1)
+        conditioning_mask = torch.zeros(mask_shape, device=device, dtype=dtype)
+        conditioning_mask[:, :, 0] = 1.0
+
+        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        latents = init_latents * conditioning_mask + noise * (1 - conditioning_mask)
+
+        conditioning_mask = self._pack_latents(
+            conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        ).squeeze(-1)
+        latents = self._pack_latents(
+            latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        )
+
+        return latents, conditioning_mask
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: PipelineImageInput = None,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        frame_rate: int = 25,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 3,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        decode_timestep: Union[float, List[float]] = 0.0,
+        decode_noise_scale: Optional[Union[float, List[float]]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 128,
+        stg_applied_layers_idx: Optional[List[int]] = [19],
+        stg_scale: Optional[float] = 1.0,
+        do_rescaling: Optional[bool] = False,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            image (`PipelineImageInput`):
+                The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`.
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `512`):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, defaults to `704`):
+                The width in pixels of the generated image. This is set to 848 by default for the best results.
+            num_frames (`int`, defaults to `161`):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, defaults to `3 `):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            decode_timestep (`float`, defaults to `0.0`):
+                The timestep at which generated video is decoded.
+            decode_noise_scale (`float`, defaults to `None`):
+                The interpolation factor between random noise and denoised latents at the decode timestep.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to `128 `):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            height=height,
+            width=width,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+        )
+
+        self._stg_scale = stg_scale
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._interrupt = False
+
+        if self.do_spatio_temporal_guidance:
+            for i in stg_applied_layers_idx:
+                self.transformer.transformer_blocks[i].forward = types.MethodType(
+                    forward_with_stg, self.transformer.transformer_blocks[i]
+                )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Prepare text embeddings
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+        elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat(
+                [negative_prompt_attention_mask, prompt_attention_mask, prompt_attention_mask], dim=0
+            )
+
+        # 4. Prepare latent variables
+        if latents is None:
+            image = self.video_processor.preprocess(image, height=height, width=width)
+            image = image.to(device=device, dtype=prompt_embeds.dtype)
+
+        num_channels_latents = self.transformer.config.in_channels
+        latents, conditioning_mask = self.prepare_latents(
+            image,
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            conditioning_mask = torch.cat([conditioning_mask, conditioning_mask])
+        elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            conditioning_mask = torch.cat([conditioning_mask, conditioning_mask, conditioning_mask])
+
+        # 5. Prepare timesteps
+        latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
+        latent_height = height // self.vae_spatial_compression_ratio
+        latent_width = width // self.vae_spatial_compression_ratio
+        video_sequence_length = latent_num_frames * latent_height * latent_width
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        mu = calculate_shift(
+            video_sequence_length,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.16),
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Prepare micro-conditions
+        latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
+        rope_interpolation_scale = (
+            1 / latent_frame_rate,
+            self.vae_spatial_compression_ratio,
+            self.vae_spatial_compression_ratio,
+        )
+
+        # 7. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 2)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 3)
+                else:
+                    latent_model_input = latents
+
+                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+                timestep = timestep.unsqueeze(-1) * (1 - conditioning_mask)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    encoder_attention_mask=prompt_attention_mask,
+                    num_frames=latent_num_frames,
+                    height=latent_height,
+                    width=latent_width,
+                    rope_interpolation_scale=rope_interpolation_scale,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                    timestep, _ = timestep.chunk(2)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
+                    noise_pred = (
+                        noise_pred_uncond
+                        + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                        + self._stg_scale * (noise_pred_text - noise_pred_perturb)
+                    )
+                    timestep, _, _ = timestep.chunk(3)
+
+                if do_rescaling:
+                    rescaling_scale = 0.7
+                    factor = noise_pred_text.std() / noise_pred.std()
+                    factor = rescaling_scale * factor + (1 - rescaling_scale)
+                    noise_pred = noise_pred * factor
+
+                # compute the previous noisy sample x_t -> x_t-1
+                noise_pred = self._unpack_latents(
+                    noise_pred,
+                    latent_num_frames,
+                    latent_height,
+                    latent_width,
+                    self.transformer_spatial_patch_size,
+                    self.transformer_temporal_patch_size,
+                )
+                latents = self._unpack_latents(
+                    latents,
+                    latent_num_frames,
+                    latent_height,
+                    latent_width,
+                    self.transformer_spatial_patch_size,
+                    self.transformer_temporal_patch_size,
+                )
+
+                noise_pred = noise_pred[:, :, 1:]
+                noise_latents = latents[:, :, 1:]
+                pred_latents = self.scheduler.step(noise_pred, t, noise_latents, return_dict=False)[0]
+
+                latents = torch.cat([latents[:, :, :1], pred_latents], dim=2)
+                latents = self._pack_latents(
+                    latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+                )
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            video = latents
+        else:
+            latents = self._unpack_latents(
+                latents,
+                latent_num_frames,
+                latent_height,
+                latent_width,
+                self.transformer_spatial_patch_size,
+                self.transformer_temporal_patch_size,
+            )
+            latents = self._denormalize_latents(
+                latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+            )
+            latents = latents.to(prompt_embeds.dtype)
+
+            if not self.vae.config.timestep_conditioning:
+                timestep = None
+            else:
+                noise = torch.randn(latents.shape, generator=generator, device=device, dtype=latents.dtype)
+                if not isinstance(decode_timestep, list):
+                    decode_timestep = [decode_timestep] * batch_size
+                if decode_noise_scale is None:
+                    decode_noise_scale = decode_timestep
+                elif not isinstance(decode_noise_scale, list):
+                    decode_noise_scale = [decode_noise_scale] * batch_size
+
+                timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype)
+                decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[
+                    :, None, None, None, None
+                ]
+                latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise
+
+            video = self.vae.decode(latents, timestep, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return LTXPipelineOutput(frames=video)
@@ -0,0 +1,843 @@
+# Copyright 2024 Genmo and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import Mochi1LoraLoaderMixin
+from diffusers.models import AutoencoderKLMochi, MochiTransformer3DModel
+from diffusers.pipelines.mochi.pipeline_output import MochiPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import (
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from examples.community.pipeline_stg_mochi import MochiSTGPipeline
+
+        >>> pipe = MochiSTGPipeline.from_pretrained("genmo/mochi-1-preview", torch_dtype=torch.bfloat16)
+        >>> pipe.enable_model_cpu_offload()
+        >>> pipe.enable_vae_tiling()
+        >>> prompt = "A close-up of a beautiful woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues, realistic style."
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [34]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+        >>> do_rescaling = False
+
+        >>> frames = pipe(
+        ...     prompt=prompt,
+        ...     num_inference_steps=28,
+        ...     guidance_scale=3.5,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ...     do_rescaling=do_rescaling).frames[0]
+        >>> export_to_video(frames, "mochi.mp4")
+        ```
+"""
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    encoder_attention_mask: torch.Tensor,
+    image_rotary_emb: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    hidden_states_ptb = hidden_states[2:]
+    encoder_hidden_states_ptb = encoder_hidden_states[2:]
+    norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
+
+    if not self.context_pre_only:
+        norm_encoder_hidden_states, enc_gate_msa, enc_scale_mlp, enc_gate_mlp = self.norm1_context(
+            encoder_hidden_states, temb
+        )
+    else:
+        norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb)
+
+    attn_hidden_states, context_attn_hidden_states = self.attn1(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=norm_encoder_hidden_states,
+        image_rotary_emb=image_rotary_emb,
+        attention_mask=encoder_attention_mask,
+    )
+
+    hidden_states = hidden_states + self.norm2(attn_hidden_states, torch.tanh(gate_msa).unsqueeze(1))
+    norm_hidden_states = self.norm3(hidden_states, (1 + scale_mlp.unsqueeze(1).to(torch.float32)))
+    ff_output = self.ff(norm_hidden_states)
+    hidden_states = hidden_states + self.norm4(ff_output, torch.tanh(gate_mlp).unsqueeze(1))
+
+    if not self.context_pre_only:
+        encoder_hidden_states = encoder_hidden_states + self.norm2_context(
+            context_attn_hidden_states, torch.tanh(enc_gate_msa).unsqueeze(1)
+        )
+        norm_encoder_hidden_states = self.norm3_context(
+            encoder_hidden_states, (1 + enc_scale_mlp.unsqueeze(1).to(torch.float32))
+        )
+        context_ff_output = self.ff_context(norm_encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states + self.norm4_context(
+            context_ff_output, torch.tanh(enc_gate_mlp).unsqueeze(1)
+        )
+
+        hidden_states[2:] = hidden_states_ptb
+        encoder_hidden_states[2:] = encoder_hidden_states_ptb
+
+    return hidden_states, encoder_hidden_states
+
+
+# from: https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77
+def linear_quadratic_schedule(num_steps, threshold_noise, linear_steps=None):
+    if linear_steps is None:
+        linear_steps = num_steps // 2
+    linear_sigma_schedule = [i * threshold_noise / linear_steps for i in range(linear_steps)]
+    threshold_noise_step_diff = linear_steps - threshold_noise * num_steps
+    quadratic_steps = num_steps - linear_steps
+    quadratic_coef = threshold_noise_step_diff / (linear_steps * quadratic_steps**2)
+    linear_coef = threshold_noise / linear_steps - 2 * threshold_noise_step_diff / (quadratic_steps**2)
+    const = quadratic_coef * (linear_steps**2)
+    quadratic_sigma_schedule = [
+        quadratic_coef * (i**2) + linear_coef * i + const for i in range(linear_steps, num_steps)
+    ]
+    sigma_schedule = linear_sigma_schedule + quadratic_sigma_schedule
+    sigma_schedule = [1.0 - x for x in sigma_schedule]
+    return sigma_schedule
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom value")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class MochiSTGPipeline(DiffusionPipeline, Mochi1LoraLoaderMixin):
+    r"""
+    The mochi pipeline for text-to-video generation.
+
+    Reference: https://github.com/genmoai/models
+
+    Args:
+        transformer ([`MochiTransformer3DModel`]):
+            Conditional Transformer architecture to denoise the encoded video latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLMochi`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLMochi,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: MochiTransformer3DModel,
+        force_zeros_for_empty_prompt: bool = False,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        # TODO: determine these scaling factors from model parameters
+        self.vae_spatial_scale_factor = 8
+        self.vae_temporal_scale_factor = 6
+        self.patch_size = 2
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_scale_factor)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 256
+        )
+        self.default_height = 480
+        self.default_width = 848
+        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 256,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        # The original Mochi implementation zeros out empty negative prompts
+        # but this can lead to overflow when placing the entire pipeline under the autocast context
+        # adding this here so that we can enable zeroing prompts if necessary
+        if self.config.force_zeros_for_empty_prompt and (prompt == "" or prompt[-1] == ""):
+            text_input_ids = torch.zeros_like(text_input_ids, device=device)
+            prompt_attention_mask = torch.zeros_like(prompt_attention_mask, dtype=torch.bool, device=device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Adapted from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 256,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        num_frames,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        height = height // self.vae_spatial_scale_factor
+        width = width // self.vae_spatial_scale_factor
+        num_frames = (num_frames - 1) // self.vae_temporal_scale_factor + 1
+
+        shape = (batch_size, num_channels_latents, num_frames, height, width)
+
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=torch.float32)
+        latents = latents.to(dtype)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_frames: int = 19,
+        num_inference_steps: int = 64,
+        timesteps: List[int] = None,
+        guidance_scale: float = 4.5,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 256,
+        stg_applied_layers_idx: Optional[List[int]] = [34],
+        stg_scale: Optional[float] = 0.0,
+        do_rescaling: Optional[bool] = False,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, *optional*, defaults to `self.default_height`):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, *optional*, defaults to `self.default_width`):
+                The width in pixels of the generated image. This is set to 848 by default for the best results.
+            num_frames (`int`, defaults to `19`):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, defaults to `4.5`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.mochi.MochiPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to `256`):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.mochi.MochiPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.mochi.MochiPipelineOutput`] is returned, otherwise a `tuple`
+                is returned where the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        height = height or self.default_height
+        width = width or self.default_width
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            height=height,
+            width=width,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._stg_scale = stg_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        if self.do_spatio_temporal_guidance:
+            for i in stg_applied_layers_idx:
+                self.transformer.transformer_blocks[i].forward = types.MethodType(
+                    forward_with_stg, self.transformer.transformer_blocks[i]
+                )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # 3. Prepare text embeddings
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+        elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat(
+                [negative_prompt_attention_mask, prompt_attention_mask, prompt_attention_mask], dim=0
+            )
+
+        # 5. Prepare timestep
+        # from https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77
+        threshold_noise = 0.025
+        sigmas = linear_quadratic_schedule(num_inference_steps, threshold_noise)
+        sigmas = np.array(sigmas)
+
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # Note: Mochi uses reversed timesteps. To ensure compatibility with methods like FasterCache, we need
+                # to make sure we're using the correct non-reversed timestep value.
+                self._current_timestep = 1000 - t
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 2)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 3)
+                else:
+                    latent_model_input = latents
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    encoder_attention_mask=prompt_attention_mask,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                # Mochi CFG + Sampling runs in FP32
+                noise_pred = noise_pred.to(torch.float32)
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
+                    noise_pred = (
+                        noise_pred_uncond
+                        + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                        + self._stg_scale * (noise_pred_text - noise_pred_perturb)
+                    )
+
+                if do_rescaling:
+                    rescaling_scale = 0.7
+                    factor = noise_pred_text.std() / noise_pred.std()
+                    factor = rescaling_scale * factor + (1 - rescaling_scale)
+                    noise_pred = noise_pred * factor
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents.to(torch.float32), return_dict=False)[0]
+                latents = latents.to(latents_dtype)
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if output_type == "latent":
+            video = latents
+        else:
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return MochiPipelineOutput(frames=video)
@@ -0,0 +1,661 @@
+# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+import types
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import ftfy
+import regex as re
+import torch
+from transformers import AutoTokenizer, UMT5EncoderModel
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import WanLoraLoaderMixin
+from diffusers.models import AutoencoderKLWan, WanTransformer3DModel
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.wan.pipeline_output import WanPipelineOutput
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from diffusers import AutoencoderKLWan
+        >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
+        >>> from examples.community.pipeline_stg_wan import WanSTGPipeline
+
+        >>> # Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+        >>> model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
+        >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+        >>> pipe = WanSTGPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+        >>> flow_shift = 5.0  # 5.0 for 720P, 3.0 for 480P
+        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+        >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [8] # Layer indices from 0 to 39 for 14b or 0 to 29 for 1.3b
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+
+        >>> output = pipe(
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     height=720,
+        ...     width=1280,
+        ...     num_frames=81,
+        ...     guidance_scale=5.0,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ... ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=16)
+        ```
+"""
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def prompt_clean(text):
+    text = whitespace_clean(basic_clean(text))
+    return text
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    rotary_emb: torch.Tensor,
+) -> torch.Tensor:
+    return hidden_states
+
+
+def forward_without_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    rotary_emb: torch.Tensor,
+) -> torch.Tensor:
+    shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+        self.scale_shift_table + temb.float()
+    ).chunk(6, dim=1)
+
+    # 1. Self-attention
+    norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
+    attn_output = self.attn1(hidden_states=norm_hidden_states, rotary_emb=rotary_emb)
+    hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
+
+    # 2. Cross-attention
+    norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
+    attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+    hidden_states = hidden_states + attn_output
+
+    # 3. Feed-forward
+    norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(hidden_states)
+    ff_output = self.ffn(norm_hidden_states)
+    hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
+
+    return hidden_states
+
+
+class WanSTGPipeline(DiffusionPipeline, WanLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation using Wan.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        tokenizer ([`T5Tokenizer`]):
+            Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer),
+            specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        transformer ([`WanTransformer3DModel`]):
+            Conditional Transformer to denoise the input latents.
+        scheduler ([`UniPCMultistepScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: UMT5EncoderModel,
+        transformer: WanTransformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt = [prompt_clean(u) for u in prompt]
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
+        seq_lens = mask.gt(0).sum(dim=1).long()
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
+        )
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        negative_prompt,
+        height,
+        width,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif negative_prompt is not None and (
+            not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list)
+        ):
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: int = 16,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_latent_frames,
+            int(height) // self.vae_scale_factor_spatial,
+            int(width) // self.vae_scale_factor_spatial,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "np",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+        stg_applied_layers_idx: Optional[List[int]] = [3, 8, 16],
+        stg_scale: Optional[float] = 0.0,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `480`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `832`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `81`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `5.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
+                The dtype to use for the torch.amp.autocast.
+
+        Examples:
+
+        Returns:
+            [`~WanPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where
+                the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._stg_scale = stg_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                latent_model_input = latents.to(transformer_dtype)
+                timestep = t.expand(latents.shape[0])
+
+                if self.do_spatio_temporal_guidance:
+                    for idx, block in enumerate(self.transformer.blocks):
+                        block.forward = types.MethodType(forward_without_stg, block)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    if self.do_spatio_temporal_guidance:
+                        for idx, block in enumerate(self.transformer.blocks):
+                            if idx in stg_applied_layers_idx:
+                                block.forward = types.MethodType(forward_with_stg, block)
+                        noise_perturb = self.transformer(
+                            hidden_states=latent_model_input,
+                            timestep=timestep,
+                            encoder_hidden_states=prompt_embeds,
+                            attention_kwargs=attention_kwargs,
+                            return_dict=False,
+                        )[0]
+                        noise_pred = (
+                            noise_uncond
+                            + guidance_scale * (noise_pred - noise_uncond)
+                            + self._stg_scale * (noise_pred - noise_perturb)
+                        )
+                    else:
+                        noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return WanPipelineOutput(frames=video)
@@ -288,8 +288,7 @@ class UFOGenScheduler(SchedulerMixin, ConfigMixin):

            if timesteps[0] >= self.config.num_train_timesteps:
                raise ValueError(
-                    f"`timesteps` must start before `self.config.train_timesteps`:"
-                    f" {self.config.num_train_timesteps}."
+                    f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}."
                )

            timesteps = np.array(timesteps, dtype=np.int64)
@@ -73,7 +73,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -89,7 +89,7 @@ def get_module_kohya_state_dict(module, prefix: str, dtype: torch.dtype, adapter

        # Set alpha parameter
        if "lora_down" in kohya_key:
-            alpha_key = f'{kohya_key.split(".")[0]}.alpha'
+            alpha_key = f"{kohya_key.split('.')[0]}.alpha"
            kohya_ss_state_dict[alpha_key] = torch.tensor(module.peft_config[adapter_name].lora_alpha).to(dtype)

    return kohya_ss_state_dict
@@ -66,7 +66,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -901,7 +901,7 @@ def main(args):
            unet_ = accelerator.unwrap_model(unet)
            lora_state_dict, _ = StableDiffusionXLPipeline.lora_state_dict(input_dir)
            unet_state_dict = {
-                f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")
+                f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")
            }
            unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
            incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
@@ -79,7 +79,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -95,7 +95,7 @@ def get_module_kohya_state_dict(module, prefix: str, dtype: torch.dtype, adapter

        # Set alpha parameter
        if "lora_down" in kohya_key:
-            alpha_key = f'{kohya_key.split(".")[0]}.alpha'
+            alpha_key = f"{kohya_key.split('.')[0]}.alpha"
            kohya_ss_state_dict[alpha_key] = torch.tensor(module.peft_config[adapter_name].lora_alpha).to(dtype)

    return kohya_ss_state_dict
@@ -72,7 +72,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -78,7 +78,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -60,7 +60,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -152,9 +152,7 @@ def log_validation(
                validation_prompt = log["validation_prompt"]
                validation_image = log["validation_image"]

-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]

                for image in images:
                    formatted_images.append(np.asarray(image))
@@ -929,17 +927,22 @@ def main(args):
    )

    # Scheduler and math around the number of training steps.
-    overrode_max_train_steps = False
-    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
    if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * num_update_steps_per_epoch * accelerator.num_processes
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )
@@ -964,8 +967,14 @@ def main(args):

    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps * accelerator.num_processes:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
    # Afterwards we recalculate our number of training epochs
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

@@ -60,7 +60,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = logging.getLogger(__name__)

@@ -51,7 +51,7 @@ from diffusers import (
    FlowMatchEulerDiscreteScheduler,
    FluxTransformer2DModel,
 )
-from diffusers.models.controlnet_flux import FluxControlNetModel
+from diffusers.models.controlnets.controlnet_flux import FluxControlNetModel
 from diffusers.optimization import get_scheduler
 from diffusers.pipelines.flux.pipeline_flux_controlnet import FluxControlNetPipeline
 from diffusers.training_utils import compute_density_for_timestep_sampling, free_memory
@@ -65,7 +65,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)
 if is_torch_npu_available():
@@ -166,9 +166,7 @@ def log_validation(
                validation_prompt = log["validation_prompt"]
                validation_image = log["validation_image"]

-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]

                for image in images:
                    formatted_images.append(np.asarray(image))
@@ -17,6 +17,7 @@ import argparse
 import contextlib
 import copy
 import functools
+import gc
 import logging
 import math
 import os
@@ -52,6 +53,7 @@ from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_density_for_timestep_sampling, compute_loss_weighting_for_sd3, free_memory
 from diffusers.utils import check_min_version, is_wandb_available, make_image_grid
 from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
+from diffusers.utils.testing_utils import backend_empty_cache
 from diffusers.utils.torch_utils import is_compiled_module


@@ -59,7 +61,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.33.0.dev0")
+check_min_version("0.34.0.dev0")

 logger = get_logger(__name__)

@@ -74,8 +76,9 @@ def log_validation(controlnet, args, accelerator, weight_dtype, step, is_final_v

    pipeline = StableDiffusion3ControlNetPipeline.from_pretrained(
        args.pretrained_model_name_or_path,
-        controlnet=controlnet,
+        controlnet=None,
        safety_checker=None,
+        transformer=None,
        revision=args.revision,
        variant=args.variant,
        torch_dtype=weight_dtype,
@@ -102,18 +105,55 @@ def log_validation(controlnet, args, accelerator, weight_dtype, step, is_final_v
            "number of `args.validation_image` and `args.validation_prompt` should be checked in `parse_args`"
        )

+    with torch.no_grad():
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = pipeline.encode_prompt(
+            validation_prompts,
+            prompt_2=None,
+            prompt_3=None,
+        )
+
+    del pipeline
+    gc.collect()
+    backend_empty_cache(accelerator.device.type)
+
+    pipeline = StableDiffusion3ControlNetPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        controlnet=controlnet,
+        safety_checker=None,
+        text_encoder=None,
+        text_encoder_2=None,
+        text_encoder_3=None,
+        revision=args.revision,
+        variant=args.variant,
+        torch_dtype=weight_dtype,
+    )
+    pipeline.enable_model_cpu_offload(device=accelerator.device.type)
+    pipeline.set_progress_bar_config(disable=True)
+
    image_logs = []
    inference_ctx = contextlib.nullcontext() if is_final_validation else torch.autocast(accelerator.device.type)

-    for validation_prompt, validation_image in zip(validation_prompts, validation_images):
+    for i, validation_image in enumerate(validation_images):
        validation_image = Image.open(validation_image).convert("RGB")
+        validation_prompt = validation_prompts[i]

        images = []

        for _ in range(args.num_validation_images):
            with inference_ctx:
                image = pipeline(
-                    validation_prompt, control_image=validation_image, num_inference_steps=20, generator=generator
+                    prompt_embeds=prompt_embeds[i].unsqueeze(0),
+                    negative_prompt_embeds=negative_prompt_embeds[i].unsqueeze(0),
+                    pooled_prompt_embeds=pooled_prompt_embeds[i].unsqueeze(0),
+                    negative_pooled_prompt_embeds=negative_pooled_prompt_embeds[i].unsqueeze(0),
+                    control_image=validation_image,
+                    num_inference_steps=20,
+                    generator=generator,
                ).images[0]

            images.append(image)
@@ -655,6 +695,7 @@ def make_train_dataset(args, tokenizer_one, tokenizer_two, tokenizer_three, acce
            dataset = load_dataset(
                args.train_data_dir,
                cache_dir=args.cache_dir,
+                trust_remote_code=True,
            )
        # See more about loading custom images at
        # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
@@ -1283,8 +1324,8 @@ def main(args):
                noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise

                # Get the text embedding for conditioning
-                prompt_embeds = batch["prompt_embeds"]
-                pooled_prompt_embeds = batch["pooled_prompt_embeds"]
+                prompt_embeds = batch["prompt_embeds"].to(dtype=weight_dtype)
+                pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(dtype=weight_dtype)

                # controlnet(s) inference
                controlnet_image = batch["conditioning_pixel_values"].to(dtype=weight_dtype)
--- a/Show More
+++ b/Show More