update

* update

* update

* update

* update

* update

* update

* update

.github/workflows/build_docker_images.yml

@@ -38,15 +38,16 @@ jobs:
           token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Build Changed Docker Images
+        env: 
+          CHANGED_FILES: ${{ steps.file_changes.outputs.all }}
         run: |
-          CHANGED_FILES="${{ steps.file_changes.outputs.all }}"
-          for FILE in $CHANGED_FILES; do
-            # skip anything that isn’t still on disk
+          echo "$CHANGED_FILES"
+          for FILE in $CHANGED_FILES; do 
+            # skip anything that isn't still on disk
             if [[ ! -f "$FILE" ]]; then
               echo "Skipping removed file $FILE"
               continue
-            fi
-            
+            fi           
             if [[ "$FILE" == docker/*Dockerfile ]]; then
               DOCKER_PATH="${FILE%/Dockerfile}"
               DOCKER_TAG=$(basename "$DOCKER_PATH")

.github/workflows/nightly_tests.yml

@@ -13,8 +13,9 @@ env:
   PYTEST_TIMEOUT: 600
   RUN_SLOW: yes
   RUN_NIGHTLY: yes
-  PIPELINE_USAGE_CUTOFF: 5000
+  PIPELINE_USAGE_CUTOFF: 0
   SLACK_API_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+  CONSOLIDATED_REPORT_PATH: consolidated_test_report.md
 
 jobs:
   setup_torch_cuda_pipeline_matrix:
@@ -99,11 +100,6 @@ jobs:
         with:
           name: pipeline_${{ matrix.module }}_test_reports
           path: reports
-      - name: Generate Report and Notify Channel
-        if: always()
-        run: |
-          pip install slack_sdk tabulate
-          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
   run_nightly_tests_for_other_torch_modules:
     name: Nightly Torch CUDA Tests
@@ -142,7 +138,6 @@ jobs:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
         # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
         CUBLAS_WORKSPACE_CONFIG: :16:8
-        RUN_COMPILE: yes
       run: |
         python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
           -s -v -k "not Flax and not Onnx" \
@@ -175,12 +170,6 @@ jobs:
         name: torch_${{ matrix.module }}_cuda_test_reports
         path: reports
 
-    - name: Generate Report and Notify Channel
-      if: always()
-      run: |
-        pip install slack_sdk tabulate
-        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
-
   run_torch_compile_tests:
     name: PyTorch Compile CUDA tests
 
@@ -224,12 +213,6 @@ jobs:
         name: torch_compile_test_reports
         path: reports
 
-    - name: Generate Report and Notify Channel
-      if: always()
-      run: |
-        pip install slack_sdk tabulate
-        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
-  
   run_big_gpu_torch_tests:
     name: Torch tests on big GPU
     strategy:
@@ -280,12 +263,7 @@ jobs:
         with:
           name: torch_cuda_big_gpu_test_reports
           path: reports
-      - name: Generate Report and Notify Channel
-        if: always()
-        run: |
-          pip install slack_sdk tabulate
-          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
-          
+
   torch_minimum_version_cuda_tests:
     name: Torch Minimum Version CUDA Tests
     runs-on:
@@ -342,63 +320,6 @@ jobs:
         with:
           name: torch_minimum_version_cuda_test_reports
           path: reports
- 
-  run_flax_tpu_tests:
-    name: Nightly Flax TPU Tests
-    runs-on:
-      group: gcp-ct5lp-hightpu-8t
-    if: github.event_name == 'schedule'
-
-    container:
-      image: diffusers/diffusers-flax-tpu
-      options: --shm-size "16gb" --ipc host --privileged ${{ vars.V5_LITEPOD_8_ENV}} -v /mnt/hf_cache:/mnt/hf_cache
-    defaults:
-      run:
-        shell: bash
-    steps:
-    - name: Checkout diffusers
-      uses: actions/checkout@v3
-      with:
-        fetch-depth: 2
-
-    - name: Install dependencies
-      run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test]
-        pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-        python -m uv pip install pytest-reportlog
-
-    - name: Environment
-      run: python utils/print_env.py
-
-    - name: Run nightly Flax TPU tests
-      env:
-        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
-      run: |
-        python -m pytest -n 0 \
-          -s -v -k "Flax" \
-          --make-reports=tests_flax_tpu \
-          --report-log=tests_flax_tpu.log \
-          tests/
-
-    - name: Failure short reports
-      if: ${{ failure() }}
-      run: |
-        cat reports/tests_flax_tpu_stats.txt
-        cat reports/tests_flax_tpu_failures_short.txt
-
-    - name: Test suite reports artifacts
-      if: ${{ always() }}
-      uses: actions/upload-artifact@v4
-      with:
-        name: flax_tpu_test_reports
-        path: reports
-
-    - name: Generate Report and Notify Channel
-      if: always()
-      run: |
-        pip install slack_sdk tabulate
-        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
   run_nightly_onnx_tests:
     name: Nightly ONNXRuntime CUDA tests on Ubuntu
@@ -449,18 +370,12 @@ jobs:
         name: tests_onnx_cuda_reports
         path: reports
 
-    - name: Generate Report and Notify Channel
-      if: always()
-      run: |
-        pip install slack_sdk tabulate
-        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
-
   run_nightly_quantization_tests:
     name: Torch quantization nightly tests
     strategy:
       fail-fast: false
       max-parallel: 2
-      matrix: 
+      matrix:
         config:
           - backend: "bitsandbytes"
             test_location: "bnb"
@@ -520,12 +435,7 @@ jobs:
         with:
           name: torch_cuda_${{ matrix.config.backend }}_reports
           path: reports
-      - name: Generate Report and Notify Channel
-        if: always()
-        run: |
-          pip install slack_sdk tabulate
-          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
-
+          
   run_nightly_pipeline_level_quantization_tests:
     name: Torch quantization nightly tests
     strategy:
@@ -574,12 +484,117 @@ jobs:
         with:
           name: torch_cuda_pipeline_level_quant_reports
           path: reports
-      - name: Generate Report and Notify Channel
-        if: always()
+
+  run_flax_tpu_tests:
+    name: Nightly Flax TPU Tests
+    runs-on:
+      group: gcp-ct5lp-hightpu-8t
+    if: github.event_name == 'schedule'
+
+    container:
+      image: diffusers/diffusers-flax-tpu
+      options: --shm-size "16gb" --ipc host --privileged ${{ vars.V5_LITEPOD_8_ENV}} -v /mnt/hf_cache:/mnt/hf_cache
+    defaults:
+      run:
+        shell: bash
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+        python -m uv pip install pytest-reportlog
+
+    - name: Environment
+      run: python utils/print_env.py
+
+    - name: Run nightly Flax TPU tests
+      env:
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
+      run: |
+        python -m pytest -n 0 \
+          -s -v -k "Flax" \
+          --make-reports=tests_flax_tpu \
+          --report-log=tests_flax_tpu.log \
+          tests/
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_flax_tpu_stats.txt
+        cat reports/tests_flax_tpu_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v4
+      with:
+        name: flax_tpu_test_reports
+        path: reports
+
+  generate_consolidated_report:
+    name: Generate Consolidated Test Report
+    needs: [
+      run_nightly_tests_for_torch_pipelines,
+      run_nightly_tests_for_other_torch_modules,
+      run_torch_compile_tests,
+      run_big_gpu_torch_tests,
+      run_nightly_quantization_tests,
+      run_nightly_pipeline_level_quantization_tests,
+      run_nightly_onnx_tests,
+      torch_minimum_version_cuda_tests,
+      run_flax_tpu_tests
+    ]
+    if: always()
+    runs-on:
+      group: aws-general-8-plus
+    container:
+      image: diffusers/diffusers-pytorch-cpu
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+
+      - name: Create reports directory
+        run: mkdir -p combined_reports
+
+      - name: Download all test reports
+        uses: actions/download-artifact@v4
+        with:
+          path: artifacts
+
+      - name: Prepare reports
         run: |
+          # Move all report files to a single directory for processing
+          find artifacts -name "*.txt" -exec cp {} combined_reports/ \;
+
+      - name: Install dependencies
+        run: |
+          pip install -e .[test]
           pip install slack_sdk tabulate
-          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
-  
+
+      - name: Generate consolidated report
+        run: |
+          python utils/consolidated_test_report.py \
+            --reports_dir combined_reports \
+            --output_file $CONSOLIDATED_REPORT_PATH \
+            --slack_channel_name diffusers-ci-nightly
+
+      - name: Show consolidated report
+        run: |
+          cat $CONSOLIDATED_REPORT_PATH >> $GITHUB_STEP_SUMMARY
+
+      - name: Upload consolidated report
+        uses: actions/upload-artifact@v4
+        with:
+          name: consolidated_test_report
+          path: ${{ env.CONSOLIDATED_REPORT_PATH }}
+
 # M1 runner currently not well supported
 # TODO: (Dhruv) add these back when we setup better testing for Apple Silicon
 #  run_nightly_tests_apple_m1:

.github/workflows/pr_tests.yml

@@ -291,8 +291,8 @@ jobs:
     - name: Failure short reports
       if: ${{ failure() }}
       run: |
-        cat reports/tests_lora_failures_short.txt
-        cat reports/tests_models_lora_failures_short.txt
+        cat reports/tests_peft_main_failures_short.txt
+        cat reports/tests_models_lora_peft_main_failures_short.txt
 
     - name: Test suite reports artifacts
       if: ${{ always() }}

docs/source/en/_toctree.yml

@@ -92,8 +92,6 @@
     title: API Reference
   title: Hybrid Inference
 - sections:
-  - local: using-diffusers/cogvideox
-    title: CogVideoX
   - local: using-diffusers/consisid
     title: ConsisID
   - local: using-diffusers/sdxl
@@ -178,10 +176,10 @@
 - sections:
   - local: optimization/fp16
     title: Accelerate inference
+  - local: optimization/cache
+    title: Caching
   - local: optimization/memory
     title: Reduce memory usage
-  - local: optimization/torch2.0
-    title: PyTorch 2.0
   - local: optimization/xformers
     title: xFormers
   - local: optimization/tome

docs/source/en/api/cache.md

@@ -11,71 +11,19 @@ specific language governing permissions and limitations under the License. -->
 
 # Caching methods
 
-## Pyramid Attention Broadcast
+Cache methods speedup diffusion transformers by storing and reusing intermediate outputs of specific layers, such as attention and feedforward layers, instead of recalculating them at each inference step.
 
-[Pyramid Attention Broadcast](https://huggingface.co/papers/2408.12588) from Xuanlei Zhao, Xiaolong Jin, Kai Wang, Yang You.
-
-Pyramid Attention Broadcast (PAB) is a method that speeds up inference in diffusion models by systematically skipping attention computations between successive inference steps and reusing cached attention states. The attention states are not very different between successive inference steps. The most prominent difference is in the spatial attention blocks, not as much in the temporal attention blocks, and finally the least in the cross attention blocks. Therefore, many cross attention computation blocks can be skipped, followed by the temporal and spatial attention blocks. By combining other techniques like sequence parallelism and classifier-free guidance parallelism, PAB achieves near real-time video generation.
-
-Enable PAB with [`~PyramidAttentionBroadcastConfig`] on any pipeline. For some benchmarks, refer to [this](https://github.com/huggingface/diffusers/pull/9562) pull request.
-
-```python
-import torch
-from diffusers import CogVideoXPipeline, PyramidAttentionBroadcastConfig
-
-pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
-pipe.to("cuda")
-
-# Increasing the value of `spatial_attention_timestep_skip_range[0]` or decreasing the value of
-# `spatial_attention_timestep_skip_range[1]` will decrease the interval in which pyramid attention
-# broadcast is active, leader to slower inference speeds. However, large intervals can lead to
-# poorer quality of generated videos.
-config = PyramidAttentionBroadcastConfig(
-    spatial_attention_block_skip_range=2,
-    spatial_attention_timestep_skip_range=(100, 800),
-    current_timestep_callback=lambda: pipe.current_timestep,
-)
-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
+## CacheMixin
 
 [[autodoc]] CacheMixin
 
-### PyramidAttentionBroadcastConfig
+## PyramidAttentionBroadcastConfig
 
 [[autodoc]] PyramidAttentionBroadcastConfig
 
 [[autodoc]] apply_pyramid_attention_broadcast
 
-### FasterCacheConfig
+## FasterCacheConfig
 
 [[autodoc]] FasterCacheConfig
 

docs/source/en/api/loaders/lora.md

@@ -98,4 +98,8 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse
 
 ## LoraBaseMixin
 
-[[autodoc]] loaders.lora_base.LoraBaseMixin
+[[autodoc]] loaders.lora_base.LoraBaseMixin
+
+## WanLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.WanLoraLoaderMixin

docs/source/en/api/pipelines/cogvideox.md

@@ -13,150 +13,181 @@
 # limitations under the License.
 -->
 
-# CogVideoX
-
-<div class="flex flex-wrap space-x-1">
-  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <a href="https://huggingface.co/docs/diffusers/main/en/tutorials/using_peft_for_inference" target="_blank" rel="noopener">
+      <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+    </a>
+  </div>
 </div>
 
-[CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://huggingface.co/papers/2408.06072) from Tsinghua University & ZhipuAI, by Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, Jie Tang.
+# CogVideoX
 
-The abstract from the paper is:
+[CogVideoX](https://huggingface.co/papers/2408.06072) is a large diffusion transformer model - available in 2B and 5B parameters - designed to generate longer and more consistent videos from text. This model uses a 3D causal variational autoencoder to more efficiently process video data by reducing sequence length (and associated training compute) and preventing flickering in generated videos. An "expert" transformer with adaptive LayerNorm improves alignment between text and video, and 3D full attention helps accurately capture motion and time in generated videos.
 
-*We introduce CogVideoX, a large-scale diffusion transformer model designed for generating videos based on text prompts. To efficently model video data, we propose to levearge a 3D Variational Autoencoder (VAE) to compresses videos along both spatial and temporal dimensions. To improve the text-video alignment, we propose an expert transformer with the expert adaptive LayerNorm to facilitate the deep fusion between the two modalities. By employing a progressive training technique, CogVideoX is adept at producing coherent, long-duration videos characterized by significant motion. In addition, we develop an effectively text-video data processing pipeline that includes various data preprocessing strategies and a video captioning method. It significantly helps enhance the performance of CogVideoX, improving both generation quality and semantic alignment. Results show that CogVideoX demonstrates state-of-the-art performance across both multiple machine metrics and human evaluations. The model weight of CogVideoX-2B is publicly available at https://github.com/THUDM/CogVideo.*
+You can find all the original CogVideoX checkpoints under the [CogVideoX](https://huggingface.co/collections/THUDM/cogvideo-66c08e62f1685a3ade464cce) collection.
 
-<Tip>
+> [!TIP]
+> Click on the CogVideoX models in the right sidebar for more examples of other video generation tasks.
 
-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.
+The example below demonstrates how to generate a video optimized for memory or inference speed.
 
-</Tip>
+<hfoptions id="usage">
+<hfoption id="memory">
 
-This pipeline was contributed by [zRzRzRzRzRzRzR](https://github.com/zRzRzRzRzRzRzR). The original codebase can be found [here](https://huggingface.co/THUDM). The original weights can be found under [hf.co/THUDM](https://huggingface.co/THUDM).
+Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
 
-There are three official CogVideoX checkpoints for text-to-video and video-to-video.
-
-| checkpoints | recommended inference dtype |
-|:---:|:---:|
-| [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b) | torch.float16 |
-| [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b) | torch.bfloat16 |
-| [`THUDM/CogVideoX1.5-5b`](https://huggingface.co/THUDM/CogVideoX1.5-5b) | torch.bfloat16 |
-
-There are two official CogVideoX checkpoints available for image-to-video.
-
-| checkpoints | recommended inference dtype |
-|:---:|:---:|
-| [`THUDM/CogVideoX-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-5b-I2V) | torch.bfloat16 |
-| [`THUDM/CogVideoX-1.5-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-1.5-5b-I2V) | torch.bfloat16 |
-
-For the CogVideoX 1.5 series:
-- Text-to-video (T2V) works best at a resolution of 1360x768 because it was trained with that specific resolution.
-- Image-to-video (I2V) works for multiple resolutions. The width can vary from 768 to 1360, but the height must be 768. The height/width must be divisible by 16.
-- Both T2V and I2V models support generation with 81 and 161 frames and work best at this value. Exporting videos at 16 FPS is recommended.
-
-There are two official CogVideoX checkpoints that support pose controllable generation (by the [Alibaba-PAI](https://huggingface.co/alibaba-pai) team).
-
-| checkpoints | recommended inference dtype |
-|:---:|:---:|
-| [`alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose) | torch.bfloat16 |
-| [`alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose) | torch.bfloat16 |
-
-## Inference
-
-Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
-
-First, load the pipeline:
-
-```python
-import torch
-from diffusers import CogVideoXPipeline, CogVideoXImageToVideoPipeline
-from diffusers.utils import export_to_video,load_image
-pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b").to("cuda") # or "THUDM/CogVideoX-2b" 
-```
-
-If you are using the image-to-video pipeline, load it as follows:
-
-```python
-pipe = CogVideoXImageToVideoPipeline.from_pretrained("THUDM/CogVideoX-5b-I2V").to("cuda")
-```
-
-Then change the memory layout of the pipelines `transformer` component to `torch.channels_last`:
-
-```python
-pipe.transformer.to(memory_format=torch.channels_last)
-```
-
-Compile the components and run inference:
-
-```python
-pipe.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
-
-# CogVideoX works well with long and well-described prompts
-prompt = "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance."
-video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
-```
-
-The [T2V benchmark](https://gist.github.com/a-r-r-o-w/5183d75e452a368fd17448fcc810bd3f) results on an 80GB A100 machine are:
-
-```
-Without torch.compile(): Average inference time: 96.89 seconds.
-With torch.compile(): Average inference time: 76.27 seconds.
-```
-
-### Memory optimization
-
-CogVideoX-2b requires about 19 GB of GPU memory to decode 49 frames (6 seconds of video at 8 FPS) with output resolution 720x480 (W x H), which makes it not possible to run on consumer GPUs or free-tier T4 Colab. The following memory optimizations could be used to reduce the memory footprint. For replication, you can refer to [this](https://gist.github.com/a-r-r-o-w/3959a03f15be5c9bd1fe545b09dfcc93) script.
-
-- `pipe.enable_model_cpu_offload()`:
-  - Without enabling cpu offloading, memory usage is `33 GB`
-  - With enabling cpu offloading, memory usage is `19 GB`
-- `pipe.enable_sequential_cpu_offload()`:
-  - Similar to `enable_model_cpu_offload` but can significantly reduce memory usage at the cost of slow inference
-  - When enabled, memory usage is under `4 GB`
-- `pipe.vae.enable_tiling()`:
-  - With enabling cpu offloading and tiling, memory usage is `11 GB`
-- `pipe.vae.enable_slicing()`
-
-## 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 [`CogVideoXPipeline`] for inference with bitsandbytes.
+The quantized CogVideoX 5B model below requires ~16GB of VRAM.
 
 ```py
 import torch
-from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, CogVideoXTransformer3DModel, CogVideoXPipeline
+from diffusers import CogVideoXPipeline, AutoModel
+from diffusers.quantizers import PipelineQuantizationConfig
+from diffusers.hooks import apply_group_offloading
 from diffusers.utils import export_to_video
-from transformers import BitsAndBytesConfig as BitsAndBytesConfig, T5EncoderModel
 
-quant_config = BitsAndBytesConfig(load_in_8bit=True)
-text_encoder_8bit = T5EncoderModel.from_pretrained(
-    "THUDM/CogVideoX-2b",
-    subfolder="text_encoder",
-    quantization_config=quant_config,
-    torch_dtype=torch.float16,
+# quantize weights to int8 with torchao
+pipeline_quant_config = PipelineQuantizationConfig(
+  quant_backend="torchao",
+  quant_kwargs={"quant_type": "int8wo"},
+  components_to_quantize=["transformer"]
 )
 
-quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
-transformer_8bit = CogVideoXTransformer3DModel.from_pretrained(
-    "THUDM/CogVideoX-2b",
+# fp8 layerwise weight-casting
+transformer = AutoModel.from_pretrained(
+    "THUDM/CogVideoX-5b",
     subfolder="transformer",
-    quantization_config=quant_config,
-    torch_dtype=torch.float16,
+    torch_dtype=torch.bfloat16
+)
+transformer.enable_layerwise_casting(
+    storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16
 )
 
 pipeline = CogVideoXPipeline.from_pretrained(
-    "THUDM/CogVideoX-2b",
-    text_encoder=text_encoder_8bit,
-    transformer=transformer_8bit,
-    torch_dtype=torch.float16,
-    device_map="balanced",
+    "THUDM/CogVideoX-5b",
+    transformer=transformer,
+    quantization_config=pipeline_quant_config,
+    torch_dtype=torch.bfloat16
 )
+pipeline.to("cuda")
 
-prompt = "A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. The ship's hull is painted a rich brown, with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an oceanic expanse. Surrounding the ship are various other toys and children's items, hinting at a playful environment. The scene captures the innocence and imagination of childhood, with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting."
-video = pipeline(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
-export_to_video(video, "ship.mp4", fps=8)
+# model-offloading
+pipeline.enable_model_cpu_offload()
+
+prompt = """
+A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. 
+The ship's hull is painted a rich brown, with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an oceanic expanse. 
+Surrounding the ship are various other toys and children's items, hinting at a playful environment. The scene captures the innocence and imagination of childhood, 
+with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting.
+"""
+
+video = pipeline(
+    prompt=prompt,
+    guidance_scale=6,
+    num_inference_steps=50
+).frames[0]
+export_to_video(video, "output.mp4", fps=8)
 ```
 
+</hfoption>
+<hfoption id="inference speed">
+
+[Compilation](../../optimization/fp16#torchcompile) is slow the first time but subsequent calls to the pipeline are faster.
+
+The average inference time with torch.compile on a 80GB A100 is 76.27 seconds compared to 96.89 seconds for an uncompiled model.
+
+```py
+import torch
+from diffusers import CogVideoXPipeline
+from diffusers.utils import export_to_video
+
+pipeline = CogVideoXPipeline.from_pretrained(
+    "THUDM/CogVideoX-2b",
+    torch_dtype=torch.float16
+).to("cuda")
+
+# torch.compile
+pipeline.transformer.to(memory_format=torch.channels_last)
+pipeline.transformer = torch.compile(
+    pipeline.transformer, mode="max-autotune", fullgraph=True
+)
+
+prompt = """
+A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. 
+The ship's hull is painted a rich brown, with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an oceanic expanse. 
+Surrounding the ship are various other toys and children's items, hinting at a playful environment. The scene captures the innocence and imagination of childhood, 
+with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting.
+"""
+
+video = pipeline(
+    prompt=prompt,
+    guidance_scale=6,
+    num_inference_steps=50
+).frames[0]
+export_to_video(video, "output.mp4", fps=8)
+```
+
+</hfoption>
+</hfoptions>
+
+## Notes
+
+- CogVideoX supports LoRAs with [`~loaders.CogVideoXLoraLoaderMixin.load_lora_weights`].
+
+  <details>
+  <summary>Show example code</summary>
+
+  ```py
+  import torch
+  from diffusers import CogVideoXPipeline
+  from diffusers.hooks import apply_group_offloading
+  from diffusers.utils import export_to_video
+
+  pipeline = CogVideoXPipeline.from_pretrained(
+      "THUDM/CogVideoX-5b",
+      torch_dtype=torch.bfloat16
+  )
+  pipeline.to("cuda")
+
+  # load LoRA weights
+  pipeline.load_lora_weights("finetrainers/CogVideoX-1.5-crush-smol-v0", adapter_name="crush-lora")
+  pipeline.set_adapters("crush-lora", 0.9)
+
+  # model-offloading
+  pipeline.enable_model_cpu_offload()
+
+  prompt = """
+  PIKA_CRUSH A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.
+  """
+  negative_prompt = "inconsistent motion, blurry motion, worse quality, degenerate outputs, deformed outputs"
+
+  video = pipeline(
+      prompt=prompt, 
+      negative_prompt=negative_prompt, 
+      num_frames=81, 
+      height=480,
+      width=768,
+      num_inference_steps=50
+  ).frames[0]
+  export_to_video(video, "output.mp4", fps=16)
+  ```
+
+  </details>
+
+- The text-to-video (T2V) checkpoints work best with a resolution of 1360x768 because that was the resolution it was pretrained on.
+
+- The image-to-video (I2V) checkpoints work with multiple resolutions. The width can vary from 768 to 1360, but the height must be 758. Both height and width must be divisible by 16.
+
+- Both T2V and I2V checkpoints work best with 81 and 161 frames. It is recommended to export the generated video at 16fps.
+
+- Refer to the table below to view memory usage when various memory-saving techniques are enabled.
+
+  | method | memory usage (enabled) | memory usage (disabled) |
+  |---|---|---|
+  | enable_model_cpu_offload | 19GB | 33GB |
+  | enable_sequential_cpu_offload | <4GB | ~33GB (very slow inference speed) |
+  | enable_tiling | 11GB (with enable_model_cpu_offload) | --- |
+ 
 ## CogVideoXPipeline
 
 [[autodoc]] CogVideoXPipeline

docs/source/en/api/pipelines/deepfloyd_if.md

@@ -347,7 +347,7 @@ pipe.to("cuda")
 image = pipe(image=image, prompt="<prompt>", strength=0.3).images
 ```
 
-You can also use [`torch.compile`](../../optimization/torch2.0). Note that we have not exhaustively tested `torch.compile`
+You can also use [`torch.compile`](../../optimization/fp16#torchcompile). Note that we have not exhaustively tested `torch.compile`
 with IF and it might not give expected results.
 
 ```py

docs/source/en/api/pipelines/hunyuan_video.md

@@ -12,78 +12,171 @@
 # See the License for the specific language governing permissions and
 # limitations under the License. -->
 
-# HunyuanVideo
-
-<div class="flex flex-wrap space-x-1">
-  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <a href="https://huggingface.co/docs/diffusers/main/en/tutorials/using_peft_for_inference" target="_blank" rel="noopener">
+      <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+    </a>
+  </div>
 </div>
 
-[HunyuanVideo](https://www.arxiv.org/abs/2412.03603) by Tencent.
+# HunyuanVideo
 
-*Recent advancements in video generation have significantly impacted daily life for both individuals and industries. However, the leading video generation models remain closed-source, resulting in a notable performance gap between industry capabilities and those available to the public. In this report, we introduce HunyuanVideo, an innovative open-source video foundation model that demonstrates performance in video generation comparable to, or even surpassing, that of leading closed-source models. HunyuanVideo encompasses a comprehensive framework that integrates several key elements, including data curation, advanced architectural design, progressive model scaling and training, and an efficient infrastructure tailored for large-scale model training and inference. As a result, we successfully trained a video generative model with over 13 billion parameters, making it the largest among all open-source models. We conducted extensive experiments and implemented a series of targeted designs to ensure high visual quality, motion dynamics, text-video alignment, and advanced filming techniques. According to evaluations by professionals, HunyuanVideo outperforms previous state-of-the-art models, including Runway Gen-3, Luma 1.6, and three top-performing Chinese video generative models. By releasing the code for the foundation model and its applications, we aim to bridge the gap between closed-source and open-source communities. This initiative will empower individuals within the community to experiment with their ideas, fostering a more dynamic and vibrant video generation ecosystem. The code is publicly available at [this https URL](https://github.com/tencent/HunyuanVideo).*
+[HunyuanVideo](https://huggingface.co/papers/2412.03603) is a 13B parameter diffusion transformer model designed to be competitive with closed-source video foundation models and enable wider community access. This model uses a "dual-stream to single-stream" architecture to separately process the video and text tokens first, before concatenating and feeding them to the transformer to fuse the multimodal information. A pretrained multimodal large language model (MLLM) is used as the encoder because it has better image-text alignment, better image detail description and reasoning, and it can be used as a zero-shot learner if system instructions are added to user prompts. Finally, HunyuanVideo uses a 3D causal variational autoencoder to more efficiently process video data at the original resolution and frame rate.
 
-<Tip>
+You can find all the original HunyuanVideo checkpoints under the [Tencent](https://huggingface.co/tencent) organization.
 
-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]
+> Click on the HunyuanVideo models in the right sidebar for more examples of video generation tasks.
+>
+> The examples below use a checkpoint from [hunyuanvideo-community](https://huggingface.co/hunyuanvideo-community) because the weights are stored in a layout compatible with Diffusers.
 
-</Tip>
+The example below demonstrates how to generate a video optimized for memory or inference speed.
 
-Recommendations for inference:
-- Both text encoders should be in `torch.float16`.
-- Transformer should be in `torch.bfloat16`.
-- VAE should be in `torch.float16`.
-- `num_frames` should be of the form `4 * k + 1`, for example `49` or `129`.
-- 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 images, try higher values (between `7.0` and `12.0`). The default value is `7.0` for HunyuanVideo.
-- For more information about supported resolutions and other details, please refer to the original repository [here](https://github.com/Tencent/HunyuanVideo/).
+<hfoptions id="usage">
+<hfoption id="memory">
 
-## Available models
+Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
 
-The following models are available for the [`HunyuanVideoPipeline`](text-to-video) pipeline:
-
-| Model name | Description |
-|:---|:---|
-| [`hunyuanvideo-community/HunyuanVideo`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo) | Official HunyuanVideo (guidance-distilled). Performs best at multiple resolutions and frames. Performs best with `guidance_scale=6.0`, `true_cfg_scale=1.0` and without a negative prompt. |
-| [`https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-T2V`](https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-T2V) | Skywork's custom finetune of HunyuanVideo (de-distilled). Performs best with `97x544x960` resolution, `guidance_scale=1.0`, `true_cfg_scale=6.0` and a negative prompt. |
-
-The following models are available for the image-to-video pipeline:
-
-| Model name | Description |
-|:---|:---|
-| [`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
-
-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 [`HunyuanVideoPipeline`] for inference with bitsandbytes.
+The quantized HunyuanVideo model below requires ~14GB of VRAM.
 
 ```py
 import torch
-from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, HunyuanVideoTransformer3DModel, HunyuanVideoPipeline
+from diffusers import AutoModel, HunyuanVideoPipeline
+from diffusers.quantizers import PipelineQuantizationConfig
 from diffusers.utils import export_to_video
 
-quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
-transformer_8bit = HunyuanVideoTransformer3DModel.from_pretrained(
-    "hunyuanvideo-community/HunyuanVideo",
-    subfolder="transformer",
-    quantization_config=quant_config,
-    torch_dtype=torch.bfloat16,
+# quantize weights to int4 with bitsandbytes
+pipeline_quant_config = PipelineQuantizationConfig(
+    quant_backend="bitsandbytes_4bit",
+    quant_kwargs={
+      "load_in_4bit": True,
+      "bnb_4bit_quant_type": "nf4",
+      "bnb_4bit_compute_dtype": torch.bfloat16
+      },
+    components_to_quantize=["transformer"]
 )
 
 pipeline = HunyuanVideoPipeline.from_pretrained(
     "hunyuanvideo-community/HunyuanVideo",
-    transformer=transformer_8bit,
-    torch_dtype=torch.float16,
-    device_map="balanced",
+    quantization_config=pipeline_quant_config,
+    torch_dtype=torch.bfloat16,
 )
 
-prompt = "A cat walks on the grass, realistic style."
+# model-offloading and tiling
+pipeline.enable_model_cpu_offload()
+pipeline.vae.enable_tiling()
+
+prompt = "A fluffy teddy bear sits on a bed of soft pillows surrounded by children's toys."
 video = pipeline(prompt=prompt, num_frames=61, num_inference_steps=30).frames[0]
-export_to_video(video, "cat.mp4", fps=15)
+export_to_video(video, "output.mp4", fps=15)
 ```
 
+</hfoption>
+<hfoption id="inference speed">
+
+[Compilation](../../optimization/fp16#torchcompile) is slow the first time but subsequent calls to the pipeline are faster.
+
+```py
+import torch
+from diffusers import AutoModel, HunyuanVideoPipeline
+from diffusers.quantizers import PipelineQuantizationConfig
+from diffusers.utils import export_to_video
+
+# quantize weights to int4 with bitsandbytes
+pipeline_quant_config = PipelineQuantizationConfig(
+    quant_backend="bitsandbytes_4bit",
+    quant_kwargs={
+      "load_in_4bit": True,
+      "bnb_4bit_quant_type": "nf4",
+      "bnb_4bit_compute_dtype": torch.bfloat16
+      },
+    components_to_quantize=["transformer"]
+)
+
+pipeline = HunyuanVideoPipeline.from_pretrained(
+    "hunyuanvideo-community/HunyuanVideo",
+    quantization_config=pipeline_quant_config,
+    torch_dtype=torch.bfloat16,
+)
+
+# model-offloading and tiling
+pipeline.enable_model_cpu_offload()
+pipeline.vae.enable_tiling()
+
+# torch.compile
+pipeline.transformer.to(memory_format=torch.channels_last)
+pipeline.transformer = torch.compile(
+    pipeline.transformer, mode="max-autotune", fullgraph=True
+)
+
+prompt = "A fluffy teddy bear sits on a bed of soft pillows surrounded by children's toys."
+video = pipeline(prompt=prompt, num_frames=61, num_inference_steps=30).frames[0]
+export_to_video(video, "output.mp4", fps=15)
+```
+
+</hfoption>
+</hfoptions>
+
+## Notes
+
+- HunyuanVideo supports LoRAs with [`~loaders.HunyuanVideoLoraLoaderMixin.load_lora_weights`].
+
+  <details>
+  <summary>Show example code</summary>
+
+  ```py
+  import torch
+  from diffusers import AutoModel, HunyuanVideoPipeline
+  from diffusers.quantizers import PipelineQuantizationConfig
+  from diffusers.utils import export_to_video
+
+  # quantize weights to int4 with bitsandbytes
+  pipeline_quant_config = PipelineQuantizationConfig(
+      quant_backend="bitsandbytes_4bit",
+      quant_kwargs={
+        "load_in_4bit": True,
+        "bnb_4bit_quant_type": "nf4",
+        "bnb_4bit_compute_dtype": torch.bfloat16
+        },
+      components_to_quantize=["transformer"]
+  )
+
+  pipeline = HunyuanVideoPipeline.from_pretrained(
+      "hunyuanvideo-community/HunyuanVideo",
+      quantization_config=pipeline_quant_config,
+      torch_dtype=torch.bfloat16,
+  )
+
+  # load LoRA weights
+  pipeline.load_lora_weights("https://huggingface.co/lucataco/hunyuan-steamboat-willie-10", adapter_name="steamboat-willie")
+  pipeline.set_adapters("steamboat-willie", 0.9)
+
+  # model-offloading and tiling
+  pipeline.enable_model_cpu_offload()
+  pipeline.vae.enable_tiling()
+
+  # use "In the style of SWR" to trigger the LoRA
+  prompt = """
+  In the style of SWR. A black and white animated scene featuring a fluffy teddy bear sits on a bed of soft pillows surrounded by children's toys.
+  """
+  video = pipeline(prompt=prompt, num_frames=61, num_inference_steps=30).frames[0]
+  export_to_video(video, "output.mp4", fps=15)
+  ```
+
+  </details>
+
+- Refer to the table below for recommended inference values.
+
+  | parameter | recommended value |
+  |---|---|
+  | text encoder dtype | `torch.float16` |
+  | transformer dtype | `torch.bfloat16` |
+  | vae dtype | `torch.float16` |
+  | `num_frames (k)` | 4 * `k` + 1 |
+
+- Try lower `shift` values (`2.0` to `5.0`) for lower resolution videos and higher `shift` values (`7.0` to `12.0`) for higher resolution images.
+
 ## HunyuanVideoPipeline
 
 [[autodoc]] HunyuanVideoPipeline

docs/source/en/api/pipelines/ltx_video.md

@@ -12,322 +12,67 @@
 # See the License for the specific language governing permissions and
 # limitations under the License. -->
 
-# LTX Video
-
-<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 style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <a href="https://huggingface.co/docs/diffusers/main/en/tutorials/using_peft_for_inference" target="_blank" rel="noopener">
+      <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+    </a>
+    <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
+  </div>
 </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.
+# LTX-Video
 
-<Tip>
+[LTX-Video](https://huggingface.co/Lightricks/LTX-Video) is a diffusion transformer designed for fast and real-time generation of high-resolution videos from text and images. The main feature of LTX-Video is the Video-VAE. The Video-VAE has a higher pixel to latent compression ratio (1:192) which enables more efficient video data processing and faster generation speed. To support and prevent finer details from being lost during generation, the Video-VAE decoder performs the latent to pixel conversion *and* the last denoising step.
 
-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.
+You can find all the original LTX-Video checkpoints under the [Lightricks](https://huggingface.co/Lightricks) organization.
 
-</Tip>
+> [!TIP]
+> Click on the LTX-Video models in the right sidebar for more examples of other video generation tasks.
 
-Available models:
+The example below demonstrates how to generate a video optimized for memory or inference speed.
 
-|  Model name   | Recommended dtype |
-|:-------------:|:-----------------:|
-| [`LTX Video 2B 0.9.0`](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.safetensors) | `torch.bfloat16` |
-| [`LTX Video 2B 0.9.1`](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.1.safetensors) | `torch.bfloat16` |
-| [`LTX Video 2B 0.9.5`](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.5.safetensors) | `torch.bfloat16` |
-| [`LTX Video 13B 0.9.7`](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltxv-13b-0.9.7-dev.safetensors) | `torch.bfloat16` |
-| [`LTX Video 13B 0.9.7 (distilled)`](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltxv-13b-0.9.7-distilled.safetensors) | `torch.bfloat16` |
-| [`LTX Video Spatial Upscaler 0.9.7`](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltxv-spatial-upscaler-0.9.7.safetensors) | `torch.bfloat16` |
+<hfoptions id="usage">
+<hfoption id="memory">
 
-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.
+Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
 
-## Recommended settings for generation
-
-For the best results, it is recommended to follow the guidelines mentioned in the official LTX Video [repository](https://github.com/Lightricks/LTX-Video).
-
-- Some variants of LTX Video are guidance-distilled. For guidance-distilled models, `guidance_scale` must be set to `1.0`. For any other models, `guidance_scale` should be set higher (e.g., `5.0`) for good generation quality.
-- For variants with a timestep-aware VAE (LTXV 0.9.1 and above), it is recommended to set `decode_timestep` to `0.05` and `image_cond_noise_scale` to `0.025`.
-- For variants that support interpolation between multiple conditioning images and videos (LTXV 0.9.5 and above), it is recommended to use similar looking images/videos for the best results. High divergence between the conditionings may lead to abrupt transitions in the generated video.
-
-<!-- TODO(aryan): remove this warning when modular diffusers is ready -->
-
-<Tip warning={true}>
-
-The examples below show some recommended generation settings, but note that all features supported in the original [LTX Video repository](https://github.com/Lightricks/LTX-Video) are not supported in `diffusers` yet (for example, Spatio-temporal Guidance and CRF compression for image inputs). This will gradually be supported in the future. For the best possible generation quality, we recommend using the code from the original repository.
-
-</Tip>
-
-## Using LTX Video 13B 0.9.7
-
-LTX Video 0.9.7 comes with a spatial latent upscaler and a 13B parameter transformer. The inference involves generating a low resolution video first, which is very fast, followed by upscaling and refining the generated video.
-
-<!-- TODO(aryan): modify when official checkpoints are available -->
-
-```python
-import torch
-from diffusers import LTXConditionPipeline, LTXLatentUpsamplePipeline
-from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXVideoCondition
-from diffusers.utils import export_to_video, load_video
-
-pipe = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.7-dev", torch_dtype=torch.bfloat16)
-pipe_upsample = LTXLatentUpsamplePipeline.from_pretrained("Lightricks/ltxv-spatial-upscaler-0.9.7", vae=pipe.vae, torch_dtype=torch.bfloat16)
-pipe.to("cuda")
-pipe_upsample.to("cuda")
-pipe.vae.enable_tiling()
-
-def round_to_nearest_resolution_acceptable_by_vae(height, width):
-    height = height - (height % pipe.vae_temporal_compression_ratio)
-    width = width - (width % pipe.vae_temporal_compression_ratio)
-    return height, width
-
-video = load_video(
-    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cosmos/cosmos-video2world-input-vid.mp4"
-)[:21]  # Use only the first 21 frames as conditioning
-condition1 = LTXVideoCondition(video=video, frame_index=0)
-
-prompt = "The video depicts a winding mountain road covered in snow, with a single vehicle traveling along it. The road is flanked by steep, rocky cliffs and sparse vegetation. The landscape is characterized by rugged terrain and a river visible in the distance. The scene captures the solitude and beauty of a winter drive through a mountainous region."
-negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
-expected_height, expected_width = 768, 1152
-downscale_factor = 2 / 3
-num_frames = 161
-
-# Part 1. Generate video at smaller resolution
-# Text-only conditioning is also supported without the need to pass `conditions`
-downscaled_height, downscaled_width = int(expected_height * downscale_factor), int(expected_width * downscale_factor)
-downscaled_height, downscaled_width = round_to_nearest_resolution_acceptable_by_vae(downscaled_height, downscaled_width)
-latents = pipe(
-    conditions=[condition1],
-    prompt=prompt,
-    negative_prompt=negative_prompt,
-    width=downscaled_width,
-    height=downscaled_height,
-    num_frames=num_frames,
-    num_inference_steps=30,
-    decode_timestep=0.05,
-    decode_noise_scale=0.025,
-    image_cond_noise_scale=0.0,
-    guidance_scale=5.0,
-    guidance_rescale=0.7,
-    generator=torch.Generator().manual_seed(0),
-    output_type="latent",
-).frames
-
-# Part 2. Upscale generated video using latent upsampler with fewer inference steps
-# The available latent upsampler upscales the height/width by 2x
-upscaled_height, upscaled_width = downscaled_height * 2, downscaled_width * 2
-upscaled_latents = pipe_upsample(
-    latents=latents,
-    output_type="latent"
-).frames
-
-# Part 3. Denoise the upscaled video with few steps to improve texture (optional, but recommended)
-video = pipe(
-    conditions=[condition1],
-    prompt=prompt,
-    negative_prompt=negative_prompt,
-    width=upscaled_width,
-    height=upscaled_height,
-    num_frames=num_frames,
-    denoise_strength=0.4,  # Effectively, 4 inference steps out of 10
-    num_inference_steps=10,
-    latents=upscaled_latents,
-    decode_timestep=0.05,
-    decode_noise_scale=0.025,
-    image_cond_noise_scale=0.0,
-    guidance_scale=5.0,
-    guidance_rescale=0.7,
-    generator=torch.Generator().manual_seed(0),
-    output_type="pil",
-).frames[0]
-
-# Part 4. Downscale the video to the expected resolution
-video = [frame.resize((expected_width, expected_height)) for frame in video]
-
-export_to_video(video, "output.mp4", fps=24)
-```
-
-## Using LTX Video 0.9.7 (distilled)
-
-The same example as above can be used with the exception of the `guidance_scale` parameter. The model is both guidance and timestep distilled in order to speedup generation. It requires `guidance_scale` to be set to `1.0`. Additionally, to benefit from the timestep distillation, `num_inference_steps` can be set between `4` and `10` for good generation quality.
-
-Additionally, custom timesteps can also be used for conditioning the generation. The authors recommend using the following timesteps for best results:
-- Base model inference to prepare for upscaling: `[1000, 993, 987, 981, 975, 909, 725, 0.03]`
-- Upscaling: `[1000, 909, 725, 421, 0]`
-
-<details>
-<summary> Full example </summary>
-
-```python
-import torch
-from diffusers import LTXConditionPipeline, LTXLatentUpsamplePipeline
-from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXVideoCondition
-from diffusers.utils import export_to_video, load_video
-
-pipe = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.7-distilled", torch_dtype=torch.bfloat16)
-pipe_upsample = LTXLatentUpsamplePipeline.from_pretrained("Lightricks/ltxv-spatial-upscaler-0.9.7", vae=pipe.vae, torch_dtype=torch.bfloat16)
-pipe.to("cuda")
-pipe_upsample.to("cuda")
-pipe.vae.enable_tiling()
-
-def round_to_nearest_resolution_acceptable_by_vae(height, width):
-    height = height - (height % pipe.vae_temporal_compression_ratio)
-    width = width - (width % pipe.vae_temporal_compression_ratio)
-    return height, width
-
-prompt = "artistic anatomical 3d render, utlra quality, human half full male body with transparent skin revealing structure instead of organs, muscular, intricate creative patterns, monochromatic with backlighting, lightning mesh, scientific concept art, blending biology with botany, surreal and ethereal quality, unreal engine 5, ray tracing, ultra realistic, 16K UHD, rich details. camera zooms out in a rotating fashion"
-negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
-expected_height, expected_width = 768, 1152
-downscale_factor = 2 / 3
-num_frames = 161
-
-# Part 1. Generate video at smaller resolution
-downscaled_height, downscaled_width = int(expected_height * downscale_factor), int(expected_width * downscale_factor)
-downscaled_height, downscaled_width = round_to_nearest_resolution_acceptable_by_vae(downscaled_height, downscaled_width)
-latents = pipe(
-    prompt=prompt,
-    negative_prompt=negative_prompt,
-    width=downscaled_width,
-    height=downscaled_height,
-    num_frames=num_frames,
-    timesteps=[1000, 993, 987, 981, 975, 909, 725, 0.03],
-    decode_timestep=0.05,
-    decode_noise_scale=0.025,
-    image_cond_noise_scale=0.0,
-    guidance_scale=1.0,
-    guidance_rescale=0.7,
-    generator=torch.Generator().manual_seed(0),
-    output_type="latent",
-).frames
-
-# Part 2. Upscale generated video using latent upsampler with fewer inference steps
-# The available latent upsampler upscales the height/width by 2x
-upscaled_height, upscaled_width = downscaled_height * 2, downscaled_width * 2
-upscaled_latents = pipe_upsample(
-    latents=latents,
-    adain_factor=1.0,
-    output_type="latent"
-).frames
-
-# Part 3. Denoise the upscaled video with few steps to improve texture (optional, but recommended)
-video = pipe(
-    prompt=prompt,
-    negative_prompt=negative_prompt,
-    width=upscaled_width,
-    height=upscaled_height,
-    num_frames=num_frames,
-    denoise_strength=0.999,  # Effectively, 4 inference steps out of 5
-    timesteps=[1000, 909, 725, 421, 0],
-    latents=upscaled_latents,
-    decode_timestep=0.05,
-    decode_noise_scale=0.025,
-    image_cond_noise_scale=0.0,
-    guidance_scale=1.0,
-    guidance_rescale=0.7,
-    generator=torch.Generator().manual_seed(0),
-    output_type="pil",
-).frames[0]
-
-# Part 4. Downscale the video to the expected resolution
-video = [frame.resize((expected_width, expected_height)) for frame in video]
-
-export_to_video(video, "output.mp4", fps=24)
-```
-
-</details>
-
-## Loading Single Files
-
-Loading the original LTX Video checkpoints is also possible with [`~ModelMixin.from_single_file`]. We recommend using `from_single_file` for the Lightricks series of models, as they plan to release multiple models in the future in the single file format.
-
-```python
-import torch
-from diffusers import AutoencoderKLLTXVideo, LTXImageToVideoPipeline, LTXVideoTransformer3DModel
-
-# `single_file_url` could also be https://huggingface.co/Lightricks/LTX-Video/ltx-video-2b-v0.9.1.safetensors
-single_file_url = "https://huggingface.co/Lightricks/LTX-Video/ltx-video-2b-v0.9.safetensors"
-transformer = LTXVideoTransformer3DModel.from_single_file(
-  single_file_url, torch_dtype=torch.bfloat16
-)
-vae = AutoencoderKLLTXVideo.from_single_file(single_file_url, torch_dtype=torch.bfloat16)
-pipe = LTXImageToVideoPipeline.from_pretrained(
-  "Lightricks/LTX-Video", transformer=transformer, vae=vae, torch_dtype=torch.bfloat16
-)
-
-# ... inference code ...
-```
-
-Alternatively, the pipeline can be used to load the weights with [`~FromSingleFileMixin.from_single_file`].
-
-```python
-import torch
-from diffusers import LTXImageToVideoPipeline
-from transformers import T5EncoderModel, T5Tokenizer
-
-single_file_url = "https://huggingface.co/Lightricks/LTX-Video/ltx-video-2b-v0.9.safetensors"
-text_encoder = T5EncoderModel.from_pretrained(
-  "Lightricks/LTX-Video", subfolder="text_encoder", torch_dtype=torch.bfloat16
-)
-tokenizer = T5Tokenizer.from_pretrained(
-  "Lightricks/LTX-Video", subfolder="tokenizer", torch_dtype=torch.bfloat16
-)
-pipe = LTXImageToVideoPipeline.from_single_file(
-  single_file_url, text_encoder=text_encoder, tokenizer=tokenizer, torch_dtype=torch.bfloat16
-)
-```
-
-Loading [LTX GGUF checkpoints](https://huggingface.co/city96/LTX-Video-gguf) are also supported:
+The LTX-Video model below requires ~10GB of VRAM.
 
 ```py
 import torch
+from diffusers import LTXPipeline, AutoModel
+from diffusers.hooks import apply_group_offloading
 from diffusers.utils import export_to_video
-from diffusers import LTXPipeline, LTXVideoTransformer3DModel, GGUFQuantizationConfig
 
-ckpt_path = (
-    "https://huggingface.co/city96/LTX-Video-gguf/blob/main/ltx-video-2b-v0.9-Q3_K_S.gguf"
-)
-transformer = LTXVideoTransformer3DModel.from_single_file(
-    ckpt_path,
-    quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
-    torch_dtype=torch.bfloat16,
-)
-pipe = LTXPipeline.from_pretrained(
+# fp8 layerwise weight-casting
+transformer = AutoModel.from_pretrained(
     "Lightricks/LTX-Video",
-    transformer=transformer,
-    torch_dtype=torch.bfloat16,
+    subfolder="transformer",
+    torch_dtype=torch.bfloat16
+)
+transformer.enable_layerwise_casting(
+    storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16
 )
-pipe.enable_model_cpu_offload()
 
-prompt = "A woman with long brown hair and light skin smiles at another woman with long blonde hair. The woman with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek. The camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be real-life footage"
+pipeline = LTXPipeline.from_pretrained("Lightricks/LTX-Video", transformer=transformer, torch_dtype=torch.bfloat16)
+
+# group-offloading
+onload_device = torch.device("cuda")
+offload_device = torch.device("cpu")
+pipeline.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True)
+apply_group_offloading(pipeline.text_encoder, onload_device=onload_device, offload_type="block_level", num_blocks_per_group=2)
+apply_group_offloading(pipeline.vae, onload_device=onload_device, offload_type="leaf_level")
+
+prompt = """
+A woman with long brown hair and light skin smiles at another woman with long blonde hair.
+The woman with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek.
+The camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and 
+natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be real-life footage
+"""
 negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
 
-video = pipe(
-    prompt=prompt,
-    negative_prompt=negative_prompt,
-    width=704,
-    height=480,
-    num_frames=161,
-    num_inference_steps=50,
-).frames[0]
-export_to_video(video, "output_gguf_ltx.mp4", fps=24)
-```
-
-Make sure to read the [documentation on GGUF](../../quantization/gguf) to learn more about our GGUF support.
-
-<!-- TODO(aryan): Update this when official weights are supported -->
-
-Loading and running inference with [LTX Video 0.9.1](https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.1.safetensors) weights.
-
-```python
-import torch
-from diffusers import LTXPipeline
-from diffusers.utils import export_to_video
-
-pipe = LTXPipeline.from_pretrained("a-r-r-o-w/LTX-Video-0.9.1-diffusers", torch_dtype=torch.bfloat16)
-pipe.to("cuda")
-
-prompt = "A woman with long brown hair and light skin smiles at another woman with long blonde hair. The woman with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek. The camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be real-life footage"
-negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
-
-video = pipe(
+video = pipeline(
     prompt=prompt,
     negative_prompt=negative_prompt,
     width=768,
@@ -340,49 +85,306 @@ video = pipe(
 export_to_video(video, "output.mp4", fps=24)
 ```
 
-Refer to [this section](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox#memory-optimization) to learn more about optimizing memory consumption.
+</hfoption>
+<hfoption id="inference speed">
 
-## 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 [`LTXPipeline`] for inference with bitsandbytes.
+[Compilation](../../optimization/fp16#torchcompile) is slow the first time but subsequent calls to the pipeline are faster.
 
 ```py
 import torch
-from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, LTXVideoTransformer3DModel, LTXPipeline
+from diffusers import LTXPipeline
 from diffusers.utils import export_to_video
-from transformers import BitsAndBytesConfig as BitsAndBytesConfig, T5EncoderModel
-
-quant_config = BitsAndBytesConfig(load_in_8bit=True)
-text_encoder_8bit = T5EncoderModel.from_pretrained(
-    "Lightricks/LTX-Video",
-    subfolder="text_encoder",
-    quantization_config=quant_config,
-    torch_dtype=torch.float16,
-)
-
-quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
-transformer_8bit = LTXVideoTransformer3DModel.from_pretrained(
-    "Lightricks/LTX-Video",
-    subfolder="transformer",
-    quantization_config=quant_config,
-    torch_dtype=torch.float16,
-)
 
 pipeline = LTXPipeline.from_pretrained(
-    "Lightricks/LTX-Video",
-    text_encoder=text_encoder_8bit,
-    transformer=transformer_8bit,
-    torch_dtype=torch.float16,
-    device_map="balanced",
+    "Lightricks/LTX-Video", torch_dtype=torch.bfloat16
 )
 
-prompt = "A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. The ship's hull is painted a rich brown, with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an oceanic expanse. Surrounding the ship are various other toys and children's items, hinting at a playful environment. The scene captures the innocence and imagination of childhood, with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting."
-video = pipeline(prompt=prompt, num_frames=161, num_inference_steps=50).frames[0]
-export_to_video(video, "ship.mp4", fps=24)
+# torch.compile
+pipeline.transformer.to(memory_format=torch.channels_last)
+pipeline.transformer = torch.compile(
+    pipeline.transformer, mode="max-autotune", fullgraph=True
+)
+
+prompt = """
+A woman with long brown hair and light skin smiles at another woman with long blonde hair.
+The woman with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek.
+The camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and 
+natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be real-life footage
+"""
+negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+
+video = pipeline(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    width=768,
+    height=512,
+    num_frames=161,
+    decode_timestep=0.03,
+    decode_noise_scale=0.025,
+    num_inference_steps=50,
+).frames[0]
+export_to_video(video, "output.mp4", fps=24)
 ```
 
+</hfoption>
+</hfoptions>
+
+## Notes
+
+- Refer to the following recommended settings for generation from the [LTX-Video](https://github.com/Lightricks/LTX-Video) repository.
+
+  - The recommended dtype for the transformer, VAE, and text encoder is `torch.bfloat16`. The VAE and text encoder can also be `torch.float32` or `torch.float16`.
+  - For guidance-distilled variants of LTX-Video, set `guidance_scale` to `1.0`. The `guidance_scale` for any other model should be set higher, like `5.0`, for good generation quality.
+  - For timestep-aware VAE variants (LTX-Video 0.9.1 and above), set `decode_timestep` to `0.05` and `image_cond_noise_scale` to `0.025`.
+  - For variants that support interpolation between multiple conditioning images and videos (LTX-Video 0.9.5 and above), use similar images and videos for the best results. Divergence from the conditioning inputs may lead to abrupt transitionts in the generated video.
+
+- LTX-Video 0.9.7 includes a spatial latent upscaler and a 13B parameter transformer. During inference, a low resolution video is quickly generated first and then upscaled and refined.
+
+  <details>
+  <summary>Show example code</summary>
+
+  ```py
+  import torch
+  from diffusers import LTXConditionPipeline, LTXLatentUpsamplePipeline
+  from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXVideoCondition
+  from diffusers.utils import export_to_video, load_video
+
+  pipeline = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.7-dev", torch_dtype=torch.bfloat16)
+  pipeline_upsample = LTXLatentUpsamplePipeline.from_pretrained("Lightricks/ltxv-spatial-upscaler-0.9.7", vae=pipeline.vae, torch_dtype=torch.bfloat16)
+  pipeline.to("cuda")
+  pipe_upsample.to("cuda")
+  pipeline.vae.enable_tiling()
+
+  def round_to_nearest_resolution_acceptable_by_vae(height, width):
+      height = height - (height % pipeline.vae_temporal_compression_ratio)
+      width = width - (width % pipeline.vae_temporal_compression_ratio)
+      return height, width
+
+  video = load_video(
+      "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cosmos/cosmos-video2world-input-vid.mp4"
+  )[:21]  # only use the first 21 frames as conditioning
+  condition1 = LTXVideoCondition(video=video, frame_index=0)
+
+  prompt = """
+  The video depicts a winding mountain road covered in snow, with a single vehicle 
+  traveling along it. The road is flanked by steep, rocky cliffs and sparse vegetation. 
+  The landscape is characterized by rugged terrain and a river visible in the distance. 
+  The scene captures the solitude and beauty of a winter drive through a mountainous region.
+  """
+  negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+  expected_height, expected_width = 768, 1152
+  downscale_factor = 2 / 3
+  num_frames = 161
+
+  # 1. Generate video at smaller resolution
+  # Text-only conditioning is also supported without the need to pass `conditions`
+  downscaled_height, downscaled_width = int(expected_height * downscale_factor), int(expected_width * downscale_factor)
+  downscaled_height, downscaled_width = round_to_nearest_resolution_acceptable_by_vae(downscaled_height, downscaled_width)
+  latents = pipeline(
+      conditions=[condition1],
+      prompt=prompt,
+      negative_prompt=negative_prompt,
+      width=downscaled_width,
+      height=downscaled_height,
+      num_frames=num_frames,
+      num_inference_steps=30,
+      decode_timestep=0.05,
+      decode_noise_scale=0.025,
+      image_cond_noise_scale=0.0,
+      guidance_scale=5.0,
+      guidance_rescale=0.7,
+      generator=torch.Generator().manual_seed(0),
+      output_type="latent",
+  ).frames
+
+  # 2. Upscale generated video using latent upsampler with fewer inference steps
+  # The available latent upsampler upscales the height/width by 2x
+  upscaled_height, upscaled_width = downscaled_height * 2, downscaled_width * 2
+  upscaled_latents = pipe_upsample(
+      latents=latents,
+      output_type="latent"
+  ).frames
+
+  # 3. Denoise the upscaled video with few steps to improve texture (optional, but recommended)
+  video = pipeline(
+      conditions=[condition1],
+      prompt=prompt,
+      negative_prompt=negative_prompt,
+      width=upscaled_width,
+      height=upscaled_height,
+      num_frames=num_frames,
+      denoise_strength=0.4,  # Effectively, 4 inference steps out of 10
+      num_inference_steps=10,
+      latents=upscaled_latents,
+      decode_timestep=0.05,
+      decode_noise_scale=0.025,
+      image_cond_noise_scale=0.0,
+      guidance_scale=5.0,
+      guidance_rescale=0.7,
+      generator=torch.Generator().manual_seed(0),
+      output_type="pil",
+  ).frames[0]
+
+  # 4. Downscale the video to the expected resolution
+  video = [frame.resize((expected_width, expected_height)) for frame in video]
+
+  export_to_video(video, "output.mp4", fps=24)
+  ```
+
+  </details>
+
+- LTX-Video 0.9.7 distilled model is guidance and timestep-distilled to speedup generation. It requires `guidance_scale` to be set to `1.0` and `num_inference_steps` should be set between `4` and `10` for good generation quality. You should also use the following custom timesteps for the best results.
+
+  - Base model inference to prepare for upscaling: `[1000, 993, 987, 981, 975, 909, 725, 0.03]`.
+  - Upscaling: `[1000, 909, 725, 421, 0]`.
+
+  <details>
+  <summary>Show example code</summary>
+
+  ```py
+  import torch
+  from diffusers import LTXConditionPipeline, LTXLatentUpsamplePipeline
+  from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXVideoCondition
+  from diffusers.utils import export_to_video, load_video
+
+  pipeline = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.7-distilled", torch_dtype=torch.bfloat16)
+  pipe_upsample = LTXLatentUpsamplePipeline.from_pretrained("Lightricks/ltxv-spatial-upscaler-0.9.7", vae=pipeline.vae, torch_dtype=torch.bfloat16)
+  pipeline.to("cuda")
+  pipe_upsample.to("cuda")
+  pipeline.vae.enable_tiling()
+
+  def round_to_nearest_resolution_acceptable_by_vae(height, width):
+      height = height - (height % pipeline.vae_temporal_compression_ratio)
+      width = width - (width % pipeline.vae_temporal_compression_ratio)
+      return height, width
+
+  prompt = """
+  artistic anatomical 3d render, utlra quality, human half full male body with transparent 
+  skin revealing structure instead of organs, muscular, intricate creative patterns, 
+  monochromatic with backlighting, lightning mesh, scientific concept art, blending biology 
+  with botany, surreal and ethereal quality, unreal engine 5, ray tracing, ultra realistic, 
+  16K UHD, rich details. camera zooms out in a rotating fashion
+  """
+  negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+  expected_height, expected_width = 768, 1152
+  downscale_factor = 2 / 3
+  num_frames = 161
+
+  # 1. Generate video at smaller resolution
+  downscaled_height, downscaled_width = int(expected_height * downscale_factor), int(expected_width * downscale_factor)
+  downscaled_height, downscaled_width = round_to_nearest_resolution_acceptable_by_vae(downscaled_height, downscaled_width)
+  latents = pipeline(
+      prompt=prompt,
+      negative_prompt=negative_prompt,
+      width=downscaled_width,
+      height=downscaled_height,
+      num_frames=num_frames,
+      timesteps=[1000, 993, 987, 981, 975, 909, 725, 0.03],
+      decode_timestep=0.05,
+      decode_noise_scale=0.025,
+      image_cond_noise_scale=0.0,
+      guidance_scale=1.0,
+      guidance_rescale=0.7,
+      generator=torch.Generator().manual_seed(0),
+      output_type="latent",
+  ).frames
+
+  # 2. Upscale generated video using latent upsampler with fewer inference steps
+  # The available latent upsampler upscales the height/width by 2x
+  upscaled_height, upscaled_width = downscaled_height * 2, downscaled_width * 2
+  upscaled_latents = pipe_upsample(
+      latents=latents,
+      adain_factor=1.0,
+      output_type="latent"
+  ).frames
+
+  # 3. Denoise the upscaled video with few steps to improve texture (optional, but recommended)
+  video = pipeline(
+      prompt=prompt,
+      negative_prompt=negative_prompt,
+      width=upscaled_width,
+      height=upscaled_height,
+      num_frames=num_frames,
+      denoise_strength=0.999,  # Effectively, 4 inference steps out of 5
+      timesteps=[1000, 909, 725, 421, 0],
+      latents=upscaled_latents,
+      decode_timestep=0.05,
+      decode_noise_scale=0.025,
+      image_cond_noise_scale=0.0,
+      guidance_scale=1.0,
+      guidance_rescale=0.7,
+      generator=torch.Generator().manual_seed(0),
+      output_type="pil",
+  ).frames[0]
+
+  # 4. Downscale the video to the expected resolution
+  video = [frame.resize((expected_width, expected_height)) for frame in video]
+
+  export_to_video(video, "output.mp4", fps=24)
+  ```
+
+  </details>
+
+- LTX-Video supports LoRAs with [`~loaders.LTXVideoLoraLoaderMixin.load_lora_weights`].
+
+  <details>
+  <summary>Show example code</summary>
+
+  ```py
+  import torch
+  from diffusers import LTXConditionPipeline
+  from diffusers.utils import export_to_video, load_image
+
+  pipeline = LTXConditionPipeline.from_pretrained(
+      "Lightricks/LTX-Video-0.9.5", torch_dtype=torch.bfloat16
+  )
+
+  pipeline.load_lora_weights("Lightricks/LTX-Video-Cakeify-LoRA", adapter_name="cakeify")
+  pipeline.set_adapters("cakeify")
+
+  # use "CAKEIFY" to trigger the LoRA
+  prompt = "CAKEIFY a person using a knife to cut a cake shaped like a Pikachu plushie"
+  image = load_image("https://huggingface.co/Lightricks/LTX-Video-Cakeify-LoRA/resolve/main/assets/images/pikachu.png")
+
+  video = pipeline(
+      prompt=prompt,
+      image=image,
+      width=576,
+      height=576,
+      num_frames=161,
+      decode_timestep=0.03,
+      decode_noise_scale=0.025,
+      num_inference_steps=50,
+  ).frames[0]
+  export_to_video(video, "output.mp4", fps=26)
+  ```
+
+  </details>
+
+- LTX-Video supports loading from single files, such as [GGUF checkpoints](../../quantization/gguf), with [`loaders.FromOriginalModelMixin.from_single_file`] or [`loaders.FromSingleFileMixin.from_single_file`].
+
+  <details>
+  <summary>Show example code</summary>
+
+  ```py
+  import torch
+  from diffusers.utils import export_to_video
+  from diffusers import LTXPipeline, AutoModel, GGUFQuantizationConfig
+
+  transformer = AutoModel.from_single_file(
+      "https://huggingface.co/city96/LTX-Video-gguf/blob/main/ltx-video-2b-v0.9-Q3_K_S.gguf"
+      quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
+      torch_dtype=torch.bfloat16
+  )
+  pipeline = LTXPipeline.from_pretrained(
+      "Lightricks/LTX-Video",
+      transformer=transformer,
+      torch_dtype=torch.bfloat16
+  )
+  ```
+
+  </details>
+
 ## LTXPipeline
 
 [[autodoc]] LTXPipeline

docs/source/en/api/pipelines/sana_sprint.md

@@ -88,12 +88,46 @@ image.save("sana.png")
 
 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.
 
+## Image to Image 
+
+The [`SanaSprintImg2ImgPipeline`] is a pipeline for image-to-image generation. It takes an input image and a prompt, and generates a new image based on the input image and the prompt.
+
+```py
+import torch
+from diffusers import SanaSprintImg2ImgPipeline
+from diffusers.utils.loading_utils import load_image
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/penguin.png"
+)
+
+pipe = SanaSprintImg2ImgPipeline.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers", 
+    torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+image = pipe(
+    prompt="a cute pink bear", 
+    image=image, 
+    strength=0.5, 
+    height=832, 
+    width=480
+).images[0]
+image.save("output.png")
+```
+
 ## SanaSprintPipeline
 
 [[autodoc]] SanaSprintPipeline
   - all
   - __call__
 
+## SanaSprintImg2ImgPipeline
+
+[[autodoc]] SanaSprintImg2ImgPipeline
+  - all
+  - __call__
+
 
 ## SanaPipelineOutput
 

docs/source/en/api/pipelines/wan.md

@@ -12,128 +12,170 @@
 # 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 style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <a href="https://huggingface.co/docs/diffusers/main/en/tutorials/using_peft_for_inference" target="_blank" rel="noopener">
+      <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+    </a>
+  </div>
 </div>
 
-[Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+# Wan2.1
 
-<!-- TODO(aryan): update abstract once paper is out -->
+[Wan-2.1](https://huggingface.co/papers/2503.20314) by the Wan Team.
 
-## Generating Videos with Wan 2.1
+*This report presents Wan, a comprehensive and open suite of video foundation models designed to push the boundaries of video generation. Built upon the mainstream diffusion transformer paradigm, Wan achieves significant advancements in generative capabilities through a series of innovations, including our novel VAE, scalable pre-training strategies, large-scale data curation, and automated evaluation metrics. These contributions collectively enhance the model's performance and versatility. Specifically, Wan is characterized by four key features: Leading Performance: The 14B model of Wan, trained on a vast dataset comprising billions of images and videos, demonstrates the scaling laws of video generation with respect to both data and model size. It consistently outperforms the existing open-source models as well as state-of-the-art commercial solutions across multiple internal and external benchmarks, demonstrating a clear and significant performance superiority. Comprehensiveness: Wan offers two capable models, i.e., 1.3B and 14B parameters, for efficiency and effectiveness respectively. It also covers multiple downstream applications, including image-to-video, instruction-guided video editing, and personal video generation, encompassing up to eight tasks. Consumer-Grade Efficiency: The 1.3B model demonstrates exceptional resource efficiency, requiring only 8.19 GB VRAM, making it compatible with a wide range of consumer-grade GPUs. Openness: We open-source the entire series of Wan, including source code and all models, with the goal of fostering the growth of the video generation community. This openness seeks to significantly expand the creative possibilities of video production in the industry and provide academia with high-quality video foundation models. All the code and models are available at [this https URL](https://github.com/Wan-Video/Wan2.1).*
 
-We will first need to install some additional dependencies.
+You can find all the original Wan2.1 checkpoints under the [Wan-AI](https://huggingface.co/Wan-AI) organization.
 
-```shell
-pip install -u ftfy imageio-ffmpeg imageio
-```
+The following Wan models are supported in Diffusers:
+- [Wan 2.1 T2V 1.3B](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers)
+- [Wan 2.1 T2V 14B](https://huggingface.co/Wan-AI/Wan2.1-T2V-14B-Diffusers)
+- [Wan 2.1 I2V 14B - 480P](https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-480P-Diffusers)
+- [Wan 2.1 I2V 14B - 720P](https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-720P-Diffusers)
+- [Wan 2.1 FLF2V 14B - 720P](https://huggingface.co/Wan-AI/Wan2.1-FLF2V-14B-720P-diffusers)
+- [Wan 2.1 VACE 1.3B](https://huggingface.co/Wan-AI/Wan2.1-VACE-1.3B-diffusers)
+- [Wan 2.1 VACE 14B](https://huggingface.co/Wan-AI/Wan2.1-VACE-14B-diffusers)
 
-### Text to Video Generation
+> [!TIP]
+> Click on the Wan2.1 models in the right sidebar for more examples of 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.
+### Text-to-Video Generation
 
-```python
-from diffusers import WanPipeline
-from diffusers.utils import export_to_video
+The example below demonstrates how to generate a video from text optimized for memory or inference speed.
 
-# 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"
+<hfoptions id="T2V usage">
+<hfoption id="T2V memory">
 
-pipe = WanPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
-pipe.enable_model_cpu_offload()
+Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
 
-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
+The Wan2.1 text-to-video model below requires ~13GB of VRAM.
 
-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
+```py
+# pip install ftfy
 import torch
 import numpy as np
-from diffusers import AutoencoderKLWan, WanImageToVideoPipeline
+from diffusers import AutoModel, WanPipeline
+from diffusers.quantizers import PipelineQuantizationConfig
+from diffusers.hooks.group_offloading import apply_group_offloading
 from diffusers.utils import export_to_video, load_image
-from transformers import CLIPVisionModel
+from transformers import UMT5EncoderModel
 
-# 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
+text_encoder = UMT5EncoderModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
+transformer = AutoModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+
+# group-offloading
+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
 )
-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
+transformer.enable_group_offload(
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="leaf_level",
+    use_stream=True
 )
 
-# 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"
+pipeline = WanPipeline.from_pretrained(
+    "Wan-AI/Wan2.1-T2V-14B-Diffusers",
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    torch_dtype=torch.bfloat16
 )
+pipeline.to("cuda")
 
-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 = """
+The camera rushes from far to near in a low-angle shot, 
+revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
+for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
+Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic 
+shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
+"""
+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
+"""
 
-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,
+output = pipeline(
     prompt=prompt,
     negative_prompt=negative_prompt,
-    height=height,
-    width=width,
-    num_frames=num_frames,
+    num_frames=81,
     guidance_scale=5.0,
 ).frames[0]
-export_to_video(output, "wan-i2v.mp4", fps=16)
+export_to_video(output, "output.mp4", fps=16)
 ```
 
-### First and Last Frame Interpolation
+</hfoption>
+<hfoption id="T2V inference speed">
+
+[Compilation](../../optimization/fp16#torchcompile) is slow the first time but subsequent calls to the pipeline are faster.
+
+```py
+# pip install ftfy
+import torch
+import numpy as np
+from diffusers import AutoModel, WanPipeline
+from diffusers.hooks.group_offloading import apply_group_offloading
+from diffusers.utils import export_to_video, load_image
+from transformers import UMT5EncoderModel
+
+text_encoder = UMT5EncoderModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
+transformer = AutoModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+
+pipeline = WanPipeline.from_pretrained(
+    "Wan-AI/Wan2.1-T2V-14B-Diffusers",
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    torch_dtype=torch.bfloat16
+)
+pipeline.to("cuda")
+
+# torch.compile
+pipeline.transformer.to(memory_format=torch.channels_last)
+pipeline.transformer = torch.compile(
+    pipeline.transformer, mode="max-autotune", fullgraph=True
+)
+
+prompt = """
+The camera rushes from far to near in a low-angle shot, 
+revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
+for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
+Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic 
+shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
+"""
+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
+"""
+
+output = pipeline(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    num_frames=81,
+    guidance_scale=5.0,
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
+```
+
+</hfoption>
+</hfoptions>
+
+### First-Last-Frame-to-Video Generation
+
+The example below demonstrates how to use the image-to-video pipeline to generate a video using a text description, a starting frame, and an ending frame.
+
+<hfoptions id="FLF2V usage">
+<hfoption id="usage">
 
 ```python
 import numpy as np
@@ -166,13 +208,13 @@ def aspect_ratio_resize(image, pipe, max_area=720 * 1280):
 def center_crop_resize(image, height, width):
     # Calculate resize ratio to match first frame dimensions
     resize_ratio = max(width / image.width, height / image.height)
-    
+
     # Resize the image
     width = round(image.width * resize_ratio)
     height = round(image.height * resize_ratio)
     size = [width, height]
     image = TF.center_crop(image, size)
-    
+
     return image, height, width
 
 first_frame, height, width = aspect_ratio_resize(first_frame, pipe)
@@ -187,320 +229,103 @@ output = pipe(
 export_to_video(output, "output.mp4", fps=16)
 ```
 
-### Video to Video Generation
+</hfoption>
+</hfoptions>
 
-```python
-import torch
-from diffusers.utils import load_video, export_to_video
-from diffusers import AutoencoderKLWan, WanVideoToVideoPipeline, UniPCMultistepScheduler
+### Any-to-Video Controllable Generation
 
-# 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()
+Wan VACE supports various generation techniques which achieve controllable video generation. Some of the capabilities include:
+- Control to Video (Depth, Pose, Sketch, Flow, Grayscale, Scribble, Layout, Boundary Box, etc.). Recommended library for preprocessing videos to obtain control videos: [huggingface/controlnet_aux]()
+- Image/Video to Video (first frame, last frame, starting clip, ending clip, random clips)
+- Inpainting and Outpainting
+- Subject to Video (faces, object, characters, etc.)
+- Composition to Video (reference anything, animate anything, swap anything, expand anything, move anything, etc.)
 
-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]
+The code snippets available in [this](https://github.com/huggingface/diffusers/pull/11582) pull request demonstrate some examples of how videos can be generated with controllability signals.
 
-export_to_video(output, "wan-v2v.mp4", fps=16)
-```
+The general rule of thumb to keep in mind when preparing inputs for the VACE pipeline is that the input images, or frames of a video that you want to use for conditioning, should have a corresponding mask that is black in color. The black mask signifies that the model will not generate new content for that area, and only use those parts for conditioning the generation process. For parts/frames that should be generated by the model, the mask should be white in color.
 
-## Memory Optimizations for Wan 2.1
+## Notes
 
-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.
+- Wan2.1 supports LoRAs with [`~loaders.WanLoraLoaderMixin.load_lora_weights`].
 
-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.
+  <details>
+  <summary>Show example code</summary>
 
-### Group Offloading the Transformer and UMT5 Text Encoder
+  ```py
+  # pip install ftfy
+  import torch
+  from diffusers import AutoModel, WanPipeline
+  from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
+  from diffusers.utils import export_to_video
 
-Find more information about group offloading [here](../optimization/memory.md)
+  vae = AutoModel.from_pretrained(
+      "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="vae", torch_dtype=torch.float32
+  )
+  pipeline = WanPipeline.from_pretrained(
+      "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", vae=vae, torch_dtype=torch.bfloat16
+  )
+  pipeline.scheduler = UniPCMultistepScheduler.from_config(
+      pipeline.scheduler.config, flow_shift=5.0
+  )
+  pipeline.to("cuda")
 
-#### Block Level Group Offloading
+  pipeline.load_lora_weights("benjamin-paine/steamboat-willie-1.3b", adapter_name="steamboat-willie")
+  pipeline.set_adapters("steamboat-willie")
 
-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`.
+  pipeline.enable_model_cpu_offload()
 
-The following example will now only require 14GB of VRAM to run, but will take approximately 30 minutes to generate a video.
+  # use "steamboat willie style" to trigger the LoRA
+  prompt = """
+  steamboat willie style, golden era animation, The camera rushes from far to near in a low-angle shot, 
+  revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
+  for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
+  Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic 
+  shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
+  """
 
-```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
+  output = pipeline(
+      prompt=prompt,
+      num_frames=81,
+      guidance_scale=5.0,
+  ).frames[0]
+  export_to_video(output, "output.mp4", fps=16)
+  ```
 
-# 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
-)
+  </details>
 
-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)
+- [`WanTransformer3DModel`] and [`AutoencoderKLWan`] supports loading from single files with [`~loaders.FromSingleFileMixin.from_single_file`].
 
-onload_device = torch.device("cuda")
-offload_device = torch.device("cpu")
+  <details>
+  <summary>Show example code</summary>
 
-apply_group_offloading(text_encoder,
-    onload_device=onload_device,
-    offload_device=offload_device,
-    offload_type="block_level",
-    num_blocks_per_group=4
-)
+  ```py
+  # pip install ftfy
+  import torch
+  from diffusers import WanPipeline, AutoModel
 
-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 already, we can move the rest of the model components to GPU
-pipe.to("cuda")
+  vae = AutoModel.from_single_file(
+      "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors"
+  )
+  transformer = AutoModel.from_single_file(
+      "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/diffusion_models/wan2.1_t2v_1.3B_bf16.safetensors",
+      torch_dtype=torch.bfloat16
+  )
+  pipeline = WanPipeline.from_pretrained(
+      "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+      vae=vae,
+      transformer=transformer,
+      torch_dtype=torch.bfloat16
+  )
+  ```
 
-image = load_image(
-    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
-)
+  </details>
 
-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))
+- Set the [`AutoencoderKLWan`] dtype to `torch.float32` for better decoding quality.
 
-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"
+- The number of frames per second (fps) or `k` should be calculated by `4 * k + 1`.
 
-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 already, 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.
+- Try lower `shift` values (`2.0` to `5.0`) for lower resolution videos and higher `shift` values (`7.0` to `12.0`) for higher resolution images.
 
 ## WanPipeline
 
@@ -514,6 +339,18 @@ pipe = WanPipeline.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", transform
   - all
   - __call__
 
+## WanVACEPipeline
+
+[[autodoc]] WanVACEPipeline
+  - all
+  - __call__
+
+## WanVideoToVideoPipeline
+
+[[autodoc]] WanVideoToVideoPipeline
+  - all
+  - __call__
+
 ## WanPipelineOutput
 
-[[autodoc]] pipelines.wan.pipeline_output.WanPipelineOutput
+[[autodoc]] pipelines.wan.pipeline_output.WanPipelineOutput

docs/source/en/optimization/cache.md

@@ -0,0 +1,69 @@
+<!-- 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. -->
+
+# Caching
+
+Caching accelerates inference by storing and reusing intermediate outputs of different layers, such as attention and feedforward layers, instead of performing the entire computation at each inference step. It significantly improves generation speed at the expense of more memory and doesn't require additional training.
+
+This guide shows you how to use the caching methods supported in Diffusers.
+
+## Pyramid Attention Broadcast
+
+[Pyramid Attention Broadcast (PAB)](https://huggingface.co/papers/2408.12588) is based on the observation that attention outputs aren't that different between successive timesteps of the generation process. The attention differences are smallest in the cross attention layers and are generally cached over a longer timestep range. This is followed by temporal attention and spatial attention layers.
+
+> [!TIP]
+> Not all video models have three types of attention (cross, temporal, and spatial)!
+
+PAB can be combined with other techniques like sequence parallelism and classifier-free guidance parallelism (data parallelism) for near real-time video generation.
+
+Set up and pass a [`PyramidAttentionBroadcastConfig`] to a pipeline's transformer to enable it. The `spatial_attention_block_skip_range` controls how often to skip attention calculations in the spatial attention blocks and the `spatial_attention_timestep_skip_range` is the range of timesteps to skip. Take care to choose an appropriate range because a smaller interval can lead to slower inference speeds and a larger interval can result in lower generation quality.
+
+```python
+import torch
+from diffusers import CogVideoXPipeline, PyramidAttentionBroadcastConfig
+
+pipeline = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
+pipeline.to("cuda")
+
+config = PyramidAttentionBroadcastConfig(
+    spatial_attention_block_skip_range=2,
+    spatial_attention_timestep_skip_range=(100, 800),
+    current_timestep_callback=lambda: pipe.current_timestep,
+)
+pipeline.transformer.enable_cache(config)
+```
+
+## FasterCache
+
+[FasterCache](https://huggingface.co/papers/2410.19355) caches and reuses attention features similar to [PAB](#pyramid-attention-broadcast) since output differences are small for each successive timestep.
+
+This method may also choose to skip the unconditional branch prediction, when using classifier-free guidance for sampling (common in most base models), and estimate it from the conditional branch prediction if there is significant redundancy in the predicted latent outputs between successive timesteps.
+
+Set up and pass a [`FasterCacheConfig`] to a pipeline's transformer to enable it.
+
+```python
+import torch
+from diffusers import CogVideoXPipeline, FasterCacheConfig
+
+pipe line= CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
+pipeline.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",
+)
+pipeline.transformer.enable_cache(config)
+```

docs/source/en/optimization/fp16.md

@@ -150,6 +150,24 @@ pipeline(prompt, num_inference_steps=30).images[0]
 
 Compilation is slow the first time, but once compiled, it is significantly faster. Try to only use the compiled pipeline on the same type of inference operations. Calling the compiled pipeline on a different image size retriggers compilation which is slow and inefficient.
 
+### Regional compilation
+
+[Regional compilation](https://docs.pytorch.org/tutorials/recipes/regional_compilation.html) reduces the cold start compilation time by only compiling a specific repeated region (or block) of the model instead of the entire model. The compiler reuses the cached and compiled code for the other blocks.
+
+[Accelerate](https://huggingface.co/docs/accelerate/index) provides the [compile_regions](https://github.com/huggingface/accelerate/blob/273799c85d849a1954a4f2e65767216eb37fa089/src/accelerate/utils/other.py#L78) method for automatically compiling the repeated blocks of a `nn.Module` sequentially. The rest of the model is compiled separately.
+
+```py
+# pip install -U accelerate
+import torch
+from diffusers import StableDiffusionXLPipeline
+from accelerate.utils import compile regions
+
+pipeline = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+).to("cuda")
+pipeline.unet = compile_regions(pipeline.unet, mode="reduce-overhead", fullgraph=True)
+```
+
 ### Graph breaks
 
 It is important to specify `fullgraph=True` in torch.compile to ensure there are no graph breaks in the underlying model. This allows you to take advantage of torch.compile without any performance degradation. For the UNet and VAE, this changes how you access the return variables.
@@ -170,6 +188,12 @@ The `step()` function is [called](https://github.com/huggingface/diffusers/blob/
 
 In general, the `sigmas` should [stay on the CPU](https://github.com/huggingface/diffusers/blob/35a969d297cba69110d175ee79c59312b9f49e1e/src/diffusers/schedulers/scheduling_euler_discrete.py#L240) to avoid the communication sync and latency.
 
+### Benchmarks
+
+Refer to the [diffusers/benchmarks](https://huggingface.co/datasets/diffusers/benchmarks) dataset to see inference latency and memory usage data for compiled pipelines.
+
+The [diffusers-torchao](https://github.com/sayakpaul/diffusers-torchao#benchmarking-results) repository also contains benchmarking results for compiled versions of Flux and CogVideoX.
+
 ## Dynamic quantization
 
 [Dynamic quantization](https://pytorch.org/tutorials/recipes/recipes/dynamic_quantization.html) improves inference speed by reducing precision to enable faster math operations. This particular type of quantization determines how to scale the activations based on the data at runtime rather than using a fixed scaling factor. As a result, the scaling factor is more accurately aligned with the data.

docs/source/en/optimization/tome.md

@@ -93,4 +93,4 @@ To reproduce this benchmark, feel free to use this [script](https://gist.github.
 |          |                |              2 |         OOM |             13 |               10.78 |
 |          |                |              1 |         OOM |           6.66 |                5.54 |
 
-As seen in the tables above, the speed-up from `tomesd` becomes more pronounced for larger image resolutions. It is also interesting to note that with `tomesd`, it is possible to run the pipeline on a higher resolution like 1024x1024. You may be able to speed-up inference even more with [`torch.compile`](torch2.0).
+As seen in the tables above, the speed-up from `tomesd` becomes more pronounced for larger image resolutions. It is also interesting to note that with `tomesd`, it is possible to run the pipeline on a higher resolution like 1024x1024. You may be able to speed-up inference even more with [`torch.compile`](fp16#torchcompile).

docs/source/en/optimization/torch2.0.md

@@ -1,438 +0,0 @@
-<!--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.
--->
-
-# PyTorch 2.0
-
-🤗 Diffusers supports the latest optimizations from [PyTorch 2.0](https://pytorch.org/get-started/pytorch-2.0/) which include:
-
-1. A memory-efficient attention implementation, scaled dot product attention, without requiring any extra dependencies such as xFormers.
-2. [`torch.compile`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html), a just-in-time (JIT) compiler to provide an extra performance boost when individual models are compiled.
-
-Both of these optimizations require PyTorch 2.0 or later and 🤗 Diffusers > 0.13.0.
-
-```bash
-pip install --upgrade torch diffusers
-```
-
-## Scaled dot product attention
-
-[`torch.nn.functional.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention) (SDPA) is an optimized and memory-efficient attention (similar to xFormers) that automatically enables several other optimizations depending on the model inputs and GPU type. SDPA is enabled by default if you're using PyTorch 2.0 and the latest version of 🤗 Diffusers, so you don't need to add anything to your code.
-
-However, if you want to explicitly enable it, you can set a [`DiffusionPipeline`] to use [`~models.attention_processor.AttnProcessor2_0`]:
-
-```diff
-  import torch
-  from diffusers import DiffusionPipeline
-+ from diffusers.models.attention_processor import AttnProcessor2_0
-
-  pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
-+ pipe.unet.set_attn_processor(AttnProcessor2_0())
-
-  prompt = "a photo of an astronaut riding a horse on mars"
-  image = pipe(prompt).images[0]
-```
-
-SDPA should be as fast and memory efficient as `xFormers`; check the [benchmark](#benchmark) for more details.
-
-In some cases - such as making the pipeline more deterministic or converting it to other formats - it may be helpful to use the vanilla attention processor, [`~models.attention_processor.AttnProcessor`]. To revert to [`~models.attention_processor.AttnProcessor`], call the [`~UNet2DConditionModel.set_default_attn_processor`] function on the pipeline:
-
-```diff
-  import torch
-  from diffusers import DiffusionPipeline
-
-  pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
-+ pipe.unet.set_default_attn_processor()
-
-  prompt = "a photo of an astronaut riding a horse on mars"
-  image = pipe(prompt).images[0]
-```
-
-## torch.compile
-
-The `torch.compile` function can often provide an additional speed-up to your PyTorch code. In 🤗 Diffusers, it is usually best to wrap the UNet with `torch.compile` because it does most of the heavy lifting in the pipeline.
-
-```python
-from diffusers import DiffusionPipeline
-import torch
-
-pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
-pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
-images = pipe(prompt, num_inference_steps=steps, num_images_per_prompt=batch_size).images[0]
-```
-
-Depending on GPU type, `torch.compile` can provide an *additional speed-up* of **5-300x** on top of SDPA! If you're using more recent GPU architectures such as Ampere (A100, 3090), Ada (4090), and Hopper (H100), `torch.compile` is able to squeeze even more performance out of these GPUs.
-
-Compilation requires some time to complete, so it is best suited for situations where you prepare your pipeline once and then perform the same type of inference operations multiple times. For example, calling the compiled pipeline on a different image size triggers compilation again which can be expensive.
-
-For more information and different options about `torch.compile`, refer to the [`torch_compile`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) tutorial.
-
-> [!TIP]
-> Learn more about other ways PyTorch 2.0 can help optimize your model in the [Accelerate inference of text-to-image diffusion models](../tutorials/fast_diffusion) tutorial.
-
-### Regional compilation
-
-Compiling the whole model usually has a big problem space for optimization. Models are often composed of multiple repeated blocks. [Regional compilation](https://pytorch.org/tutorials/recipes/regional_compilation.html) compiles the repeated block first (a transformer encoder block, for example), so that the Torch compiler would re-use its cached/optimized generated code for the other blocks, reducing (often massively) the cold start compilation time observed on the first inference call.
-
-Enabling regional compilation might require simple yet intrusive changes to the
-modeling code. However, 🤗 Accelerate provides a utility [`compile_regions()`](https://huggingface.co/docs/accelerate/main/en/usage_guides/compilation#how-to-use-regional-compilation) which automatically compiles
-the repeated blocks of the provided `nn.Module` sequentially, and the rest of the model separately. This helps with reducing cold start time while keeping most (if not all) of the speedup you would get from full compilation.
-
-```py
-# Make sure you're on the latest `accelerate`: `pip install -U accelerate`.
-from accelerate.utils import compile_regions
-
-pipe.unet = compile_regions(pipe.unet, mode="reduce-overhead", fullgraph=True)
-```
-
-As you may have noticed `compile_regions()` takes the same arguments as `torch.compile()`, allowing flexibility.
-
-## Benchmark
-
-We conducted a comprehensive benchmark with PyTorch 2.0's efficient attention implementation and `torch.compile` across different GPUs and batch sizes for five of our most used pipelines. The code is benchmarked on 🤗 Diffusers v0.17.0.dev0 to optimize `torch.compile` usage (see [here](https://github.com/huggingface/diffusers/pull/3313) for more details).
-
-Expand the dropdown below to find the code used to benchmark each pipeline:
-
-<details>
-
-### Stable Diffusion text-to-image
-
-```python
-from diffusers import DiffusionPipeline
-import torch
-
-path = "stable-diffusion-v1-5/stable-diffusion-v1-5"
-
-run_compile = True  # Set True / False
-
-pipe = DiffusionPipeline.from_pretrained(path, torch_dtype=torch.float16, use_safetensors=True)
-pipe = pipe.to("cuda")
-pipe.unet.to(memory_format=torch.channels_last)
-
-if run_compile:
-    print("Run torch compile")
-    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
-
-prompt = "ghibli style, a fantasy landscape with castles"
-
-for _ in range(3):
-    images = pipe(prompt=prompt).images
-```
-
-### Stable Diffusion image-to-image
-
-```python
-from diffusers import StableDiffusionImg2ImgPipeline
-from diffusers.utils import load_image
-import torch
-
-url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
-
-init_image = load_image(url)
-init_image = init_image.resize((512, 512))
-
-path = "stable-diffusion-v1-5/stable-diffusion-v1-5"
-
-run_compile = True  # Set True / False
-
-pipe = StableDiffusionImg2ImgPipeline.from_pretrained(path, torch_dtype=torch.float16, use_safetensors=True)
-pipe = pipe.to("cuda")
-pipe.unet.to(memory_format=torch.channels_last)
-
-if run_compile:
-    print("Run torch compile")
-    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
-
-prompt = "ghibli style, a fantasy landscape with castles"
-
-for _ in range(3):
-    image = pipe(prompt=prompt, image=init_image).images[0]
-```
-
-### Stable Diffusion inpainting
-
-```python
-from diffusers import StableDiffusionInpaintPipeline
-from diffusers.utils import load_image
-import torch
-
-img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
-mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
-
-init_image = load_image(img_url).resize((512, 512))
-mask_image = load_image(mask_url).resize((512, 512))
-
-path = "runwayml/stable-diffusion-inpainting"
-
-run_compile = True  # Set True / False
-
-pipe = StableDiffusionInpaintPipeline.from_pretrained(path, torch_dtype=torch.float16, use_safetensors=True)
-pipe = pipe.to("cuda")
-pipe.unet.to(memory_format=torch.channels_last)
-
-if run_compile:
-    print("Run torch compile")
-    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
-
-prompt = "ghibli style, a fantasy landscape with castles"
-
-for _ in range(3):
-    image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
-```
-
-### ControlNet
-
-```python
-from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
-from diffusers.utils import load_image
-import torch
-
-url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
-
-init_image = load_image(url)
-init_image = init_image.resize((512, 512))
-
-path = "stable-diffusion-v1-5/stable-diffusion-v1-5"
-
-run_compile = True  # Set True / False
-controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16, use_safetensors=True)
-pipe = StableDiffusionControlNetPipeline.from_pretrained(
-    path, controlnet=controlnet, torch_dtype=torch.float16, use_safetensors=True
-)
-
-pipe = pipe.to("cuda")
-pipe.unet.to(memory_format=torch.channels_last)
-pipe.controlnet.to(memory_format=torch.channels_last)
-
-if run_compile:
-    print("Run torch compile")
-    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
-    pipe.controlnet = torch.compile(pipe.controlnet, mode="reduce-overhead", fullgraph=True)
-
-prompt = "ghibli style, a fantasy landscape with castles"
-
-for _ in range(3):
-    image = pipe(prompt=prompt, image=init_image).images[0]
-```
-
-### DeepFloyd IF text-to-image + upscaling
-
-```python
-from diffusers import DiffusionPipeline
-import torch
-
-run_compile = True  # Set True / False
-
-pipe_1 = DiffusionPipeline.from_pretrained("DeepFloyd/IF-I-M-v1.0", variant="fp16", text_encoder=None, torch_dtype=torch.float16, use_safetensors=True)
-pipe_1.to("cuda")
-pipe_2 = DiffusionPipeline.from_pretrained("DeepFloyd/IF-II-M-v1.0", variant="fp16", text_encoder=None, torch_dtype=torch.float16, use_safetensors=True)
-pipe_2.to("cuda")
-pipe_3 = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-x4-upscaler", torch_dtype=torch.float16, use_safetensors=True)
-pipe_3.to("cuda")
-
-
-pipe_1.unet.to(memory_format=torch.channels_last)
-pipe_2.unet.to(memory_format=torch.channels_last)
-pipe_3.unet.to(memory_format=torch.channels_last)
-
-if run_compile:
-    pipe_1.unet = torch.compile(pipe_1.unet, mode="reduce-overhead", fullgraph=True)
-    pipe_2.unet = torch.compile(pipe_2.unet, mode="reduce-overhead", fullgraph=True)
-    pipe_3.unet = torch.compile(pipe_3.unet, mode="reduce-overhead", fullgraph=True)
-
-prompt = "the blue hulk"
-
-prompt_embeds = torch.randn((1, 2, 4096), dtype=torch.float16)
-neg_prompt_embeds = torch.randn((1, 2, 4096), dtype=torch.float16)
-
-for _ in range(3):
-    image_1 = pipe_1(prompt_embeds=prompt_embeds, negative_prompt_embeds=neg_prompt_embeds, output_type="pt").images
-    image_2 = pipe_2(image=image_1, prompt_embeds=prompt_embeds, negative_prompt_embeds=neg_prompt_embeds, output_type="pt").images
-    image_3 = pipe_3(prompt=prompt, image=image_1, noise_level=100).images
-```
-</details>
-
-The graph below highlights the relative speed-ups for the [`StableDiffusionPipeline`] across five GPU families with PyTorch 2.0 and `torch.compile` enabled. The benchmarks for the following graphs are measured in *number of iterations/second*.
-
-![t2i_speedup](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/pt2_benchmarks/t2i_speedup.png)
-
-To give you an even better idea of how this speed-up holds for the other pipelines, consider the following
-graph for an A100 with PyTorch 2.0 and `torch.compile`:
-
-![a100_numbers](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/pt2_benchmarks/a100_numbers.png)
-
-In the following tables, we report our findings in terms of the *number of iterations/second*.
-
-### A100 (batch size: 1)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 21.66 | 23.13 | 44.03 | 49.74 |
-| SD - img2img | 21.81 | 22.40 | 43.92 | 46.32 |
-| SD - inpaint | 22.24 | 23.23 | 43.76 | 49.25 |
-| SD - controlnet | 15.02 | 15.82 | 32.13 | 36.08 |
-| IF | 20.21 / <br>13.84 / <br>24.00 | 20.12 / <br>13.70 / <br>24.03 | ❌ | 97.34 / <br>27.23 / <br>111.66 |
-| SDXL - txt2img | 8.64 | 9.9 | - | - |
-
-### A100 (batch size: 4)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 11.6 | 13.12 | 14.62 | 17.27 |
-| SD - img2img | 11.47 | 13.06 | 14.66 | 17.25 |
-| SD - inpaint | 11.67 | 13.31 | 14.88 | 17.48 |
-| SD - controlnet | 8.28 | 9.38 | 10.51 | 12.41 |
-| IF | 25.02 | 18.04 | ❌ | 48.47 |
-| SDXL - txt2img | 2.44 | 2.74 | - | - |
-
-### A100 (batch size: 16)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 3.04 | 3.6 | 3.83 | 4.68 |
-| SD - img2img | 2.98 | 3.58 | 3.83 | 4.67 |
-| SD - inpaint | 3.04 | 3.66 | 3.9 | 4.76 |
-| SD - controlnet | 2.15 | 2.58 | 2.74 | 3.35 |
-| IF | 8.78 | 9.82 | ❌ | 16.77 |
-| SDXL - txt2img | 0.64 | 0.72 | - | - |
-
-### V100 (batch size: 1)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 18.99 | 19.14 | 20.95 | 22.17 |
-| SD - img2img | 18.56 | 19.18 | 20.95 | 22.11 |
-| SD - inpaint | 19.14 | 19.06 | 21.08 | 22.20 |
-| SD - controlnet | 13.48 | 13.93 | 15.18 | 15.88 |
-| IF |  20.01 / <br>9.08 / <br>23.34 | 19.79 / <br>8.98 / <br>24.10 | ❌ | 55.75 / <br>11.57 / <br>57.67 |
-
-### V100 (batch size: 4)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 5.96 | 5.89 | 6.83 | 6.86 |
-| SD - img2img | 5.90 | 5.91 | 6.81 | 6.82 |
-| SD - inpaint | 5.99 | 6.03 | 6.93 | 6.95 |
-| SD - controlnet | 4.26 | 4.29 | 4.92 | 4.93 |
-| IF | 15.41 | 14.76 | ❌ | 22.95 |
-
-### V100 (batch size: 16)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 1.66 | 1.66 | 1.92 | 1.90 |
-| SD - img2img | 1.65 | 1.65 | 1.91 | 1.89 |
-| SD - inpaint | 1.69 | 1.69 | 1.95 | 1.93 |
-| SD - controlnet | 1.19 | 1.19 | OOM after warmup | 1.36 |
-| IF | 5.43 | 5.29 | ❌ | 7.06 |
-
-### T4 (batch size: 1)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 6.9 | 6.95 | 7.3 | 7.56 |
-| SD - img2img | 6.84 | 6.99 | 7.04 | 7.55 |
-| SD - inpaint | 6.91 | 6.7 | 7.01 | 7.37 |
-| SD - controlnet | 4.89 | 4.86 | 5.35 | 5.48 |
-| IF | 17.42 / <br>2.47 / <br>18.52 | 16.96 / <br>2.45 / <br>18.69 | ❌ | 24.63 / <br>2.47 / <br>23.39 |
-| SDXL - txt2img | 1.15 | 1.16 | - | - |
-
-### T4 (batch size: 4)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 1.79 | 1.79 | 2.03 | 1.99 |
-| SD - img2img | 1.77 | 1.77 | 2.05 | 2.04 |
-| SD - inpaint | 1.81 | 1.82 | 2.09 | 2.09 |
-| SD - controlnet | 1.34 | 1.27 | 1.47 | 1.46 |
-| IF | 5.79 |  5.61 | ❌ | 7.39 |
-| SDXL - txt2img | 0.288 | 0.289 | - | - |
-
-### T4 (batch size: 16)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 2.34s | 2.30s | OOM after 2nd iteration | 1.99s |
-| SD - img2img | 2.35s | 2.31s | OOM after warmup | 2.00s |
-| SD - inpaint | 2.30s | 2.26s | OOM after 2nd iteration | 1.95s |
-| SD - controlnet | OOM after 2nd iteration | OOM after 2nd iteration | OOM after warmup | OOM after warmup |
-| IF * | 1.44 | 1.44 | ❌ | 1.94 |
-| SDXL - txt2img | OOM | OOM | - | - |
-
-### RTX 3090 (batch size: 1)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 22.56 | 22.84 | 23.84 | 25.69 |
-| SD - img2img | 22.25 | 22.61 | 24.1 | 25.83 |
-| SD - inpaint | 22.22 | 22.54 | 24.26 | 26.02 |
-| SD - controlnet | 16.03 | 16.33 | 17.38 | 18.56 |
-| IF | 27.08 / <br>9.07 / <br>31.23 | 26.75 / <br>8.92 / <br>31.47 | ❌ | 68.08 / <br>11.16 / <br>65.29 |
-
-### RTX 3090 (batch size: 4)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 6.46 | 6.35 | 7.29 | 7.3 |
-| SD - img2img | 6.33 | 6.27 | 7.31 | 7.26 |
-| SD - inpaint | 6.47 | 6.4 | 7.44 | 7.39 |
-| SD - controlnet | 4.59 | 4.54 | 5.27 | 5.26 |
-| IF | 16.81 | 16.62 | ❌ | 21.57 |
-
-### RTX 3090 (batch size: 16)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 1.7 | 1.69 | 1.93 | 1.91 |
-| SD - img2img | 1.68 | 1.67 | 1.93 | 1.9 |
-| SD - inpaint | 1.72 | 1.71 | 1.97 | 1.94 |
-| SD - controlnet | 1.23 | 1.22 | 1.4 | 1.38 |
-| IF | 5.01 | 5.00 | ❌ | 6.33 |
-
-### RTX 4090 (batch size: 1)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 40.5 | 41.89 | 44.65 | 49.81 |
-| SD - img2img | 40.39 | 41.95 | 44.46 | 49.8 |
-| SD - inpaint | 40.51 | 41.88 | 44.58 | 49.72 |
-| SD - controlnet | 29.27 | 30.29 | 32.26 | 36.03 |
-| IF | 69.71 / <br>18.78 / <br>85.49 | 69.13 / <br>18.80 / <br>85.56 | ❌ | 124.60 / <br>26.37 / <br>138.79 |
-| SDXL - txt2img | 6.8 | 8.18 | - | - |
-
-### RTX 4090 (batch size: 4)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 12.62 | 12.84 | 15.32 | 15.59 |
-| SD - img2img | 12.61 | 12,.79 | 15.35 | 15.66 |
-| SD - inpaint | 12.65 | 12.81 | 15.3 | 15.58 |
-| SD - controlnet | 9.1 | 9.25 | 11.03 | 11.22 |
-| IF | 31.88 | 31.14 | ❌ | 43.92 |
-| SDXL - txt2img | 2.19 | 2.35 | - | - |
-
-### RTX 4090 (batch size: 16)
-
-| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
-|:---:|:---:|:---:|:---:|:---:|
-| SD - txt2img | 3.17 | 3.2 | 3.84 | 3.85 |
-| SD - img2img | 3.16 | 3.2 | 3.84 | 3.85 |
-| SD - inpaint | 3.17 | 3.2 | 3.85 | 3.85 |
-| SD - controlnet | 2.23 | 2.3 | 2.7 | 2.75 |
-| IF | 9.26 | 9.2 | ❌ | 13.31 |
-| SDXL - txt2img | 0.52 | 0.53 | - | - |
-
-## Notes
-
-* Follow this [PR](https://github.com/huggingface/diffusers/pull/3313) for more details on the environment used for conducting the benchmarks.
-* For the DeepFloyd IF pipeline where batch sizes > 1, we only used a batch size of > 1 in the first IF pipeline for text-to-image generation and NOT for upscaling. That means the two upscaling pipelines received a batch size of 1.
-
-*Thanks to [Horace He](https://github.com/Chillee) from the PyTorch team for their support in improving our support of `torch.compile()` in Diffusers.*

docs/source/en/quantization/torchao.md

@@ -56,7 +56,7 @@ image = pipe(
 image.save("output.png")
 ```
 
-TorchAO is fully compatible with [torch.compile](./optimization/torch2.0#torchcompile), setting it apart from other quantization methods. This makes it easy to speed up inference with just one line of code.
+TorchAO is fully compatible with [torch.compile](../optimization/fp16#torchcompile), setting it apart from other quantization methods. This makes it easy to speed up inference with just one line of code.
 
 ```python
 # In the above code, add the following after initializing the transformer
@@ -91,7 +91,7 @@ The quantization methods supported are as follows:
 
 Some quantization methods are aliases (for example, `int8wo` is the commonly used shorthand for `int8_weight_only`). This allows using the quantization methods described in the torchao docs as-is, while also making it convenient to remember their shorthand notations.
 
-Refer to the official torchao documentation for a better understanding of the available quantization methods and the exhaustive list of configuration options available.
+Refer to the [official torchao documentation](https://docs.pytorch.org/ao/stable/index.html) for a better understanding of the available quantization methods and the exhaustive list of configuration options available.
 
 ## Serializing and Deserializing quantized models
 
@@ -155,5 +155,5 @@ transformer.load_state_dict(state_dict, strict=True, assign=True)
 
 ## Resources
 
-- [TorchAO Quantization API](https://github.com/pytorch/ao/blob/main/torchao/quantization/README.md)
+- [TorchAO Quantization API](https://docs.pytorch.org/ao/stable/index.html)
 - [Diffusers-TorchAO examples](https://github.com/sayakpaul/diffusers-torchao)

docs/source/en/stable_diffusion.md

@@ -256,6 +256,6 @@ make_image_grid(images, 2, 2)
 
 In this tutorial, you learned how to optimize a [`DiffusionPipeline`] for computational and memory efficiency as well as improving the quality of generated outputs. If you're interested in making your pipeline even faster, take a look at the following resources:
 
-- Learn how [PyTorch 2.0](./optimization/torch2.0) and [`torch.compile`](https://pytorch.org/docs/stable/generated/torch.compile.html) can yield 5 - 300% faster inference speed. On an A100 GPU, inference can be up to 50% faster!
+- Learn how [PyTorch 2.0](./optimization/fp16) and [`torch.compile`](https://pytorch.org/docs/stable/generated/torch.compile.html) can yield 5 - 300% faster inference speed. On an A100 GPU, inference can be up to 50% faster!
 - If you can't use PyTorch 2, we recommend you install [xFormers](./optimization/xformers). Its memory-efficient attention mechanism works great with PyTorch 1.13.1 for faster speed and reduced memory consumption.
 - Other optimization techniques, such as model offloading, are covered in [this guide](./optimization/fp16).

docs/source/en/training/overview.md

@@ -59,5 +59,5 @@ pip install -r requirements_sdxl.txt
 
 To speedup training and reduce memory-usage, we recommend:
 
-- using PyTorch 2.0 or higher to automatically use [scaled dot product attention](../optimization/torch2.0#scaled-dot-product-attention) during training (you don't need to make any changes to the training code)
+- using PyTorch 2.0 or higher to automatically use [scaled dot product attention](../optimization/fp16#scaled-dot-product-attention) during training (you don't need to make any changes to the training code)
 - installing [xFormers](../optimization/xformers) to enable memory-efficient attention

docs/source/en/tutorials/using_peft_for_inference.md

@@ -103,7 +103,7 @@ pipeline("A cute cnmt eating a slice of pizza, stunning color scheme, masterpiec
 
 ## torch.compile
 
-[torch.compile](../optimization/torch2.0#torchcompile) speeds up inference by compiling the PyTorch model to use optimized kernels. Before compiling, the LoRA weights need to be fused into the base model and unloaded first.
+[torch.compile](../optimization/fp16#torchcompile) speeds up inference by compiling the PyTorch model to use optimized kernels. Before compiling, the LoRA weights need to be fused into the base model and unloaded first.
 
 ```py
 import torch

docs/source/en/using-diffusers/cogvideox.md

@@ -1,120 +0,0 @@
-<!--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.
--->
-# CogVideoX
-
-CogVideoX is a text-to-video generation model focused on creating more coherent videos aligned with a prompt. It achieves this using several methods.
-
-- a 3D variational autoencoder that compresses videos spatially and temporally, improving compression rate and video accuracy.
-
-- an expert transformer block to help align text and video, and a 3D full attention module for capturing and creating spatially and temporally accurate videos.
- 
-
-
-## Load model checkpoints
-Model weights may be stored in separate subfolders on the Hub or locally, in which case, you should use the [`~DiffusionPipeline.from_pretrained`] method.
-
-
-```py
-from diffusers import CogVideoXPipeline, CogVideoXImageToVideoPipeline
-pipe = CogVideoXPipeline.from_pretrained(
-    "THUDM/CogVideoX-2b",
-    torch_dtype=torch.float16
-)
-
-pipe = CogVideoXImageToVideoPipeline.from_pretrained(
-    "THUDM/CogVideoX-5b-I2V",
-    torch_dtype=torch.bfloat16
-)
-
-```
-
-## Text-to-Video
-For text-to-video, pass a text prompt. By default, CogVideoX generates a 720x480 video for the best results.
-
-```py
-import torch
-from diffusers import CogVideoXPipeline
-from diffusers.utils import export_to_video
-
-prompt = "An elderly gentleman, with a serene expression, sits at the water's edge, a steaming cup of tea by his side. He is engrossed in his artwork, brush in hand, as he renders an oil painting on a canvas that's propped up against a small, weathered table. The sea breeze whispers through his silver hair, gently billowing his loose-fitting white shirt, while the salty air adds an intangible element to his masterpiece in progress. The scene is one of tranquility and inspiration, with the artist's canvas capturing the vibrant hues of the setting sun reflecting off the tranquil sea."
-
-pipe = CogVideoXPipeline.from_pretrained(
-    "THUDM/CogVideoX-5b",
-    torch_dtype=torch.bfloat16
-)
-
-pipe.enable_model_cpu_offload()
-pipe.vae.enable_tiling()
-
-video = pipe(
-    prompt=prompt,
-    num_videos_per_prompt=1,
-    num_inference_steps=50,
-    num_frames=49,
-    guidance_scale=6,
-    generator=torch.Generator(device="cuda").manual_seed(42),
-).frames[0]
-
-export_to_video(video, "output.mp4", fps=8)
-
-```
-
-
-<div class="flex justify-center">
-    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_out.gif" alt="generated image of an astronaut in a jungle"/>
-</div>
-
-
-## Image-to-Video
-
-
-You'll use the [THUDM/CogVideoX-5b-I2V](https://huggingface.co/THUDM/CogVideoX-5b-I2V)  checkpoint for this guide.
-
-```py
-import torch
-from diffusers import CogVideoXImageToVideoPipeline
-from diffusers.utils import export_to_video, load_image
-
-prompt = "A vast, shimmering ocean flows gracefully under a twilight sky, its waves undulating in a mesmerizing dance of blues and greens. The surface glints with the last rays of the setting sun, casting golden highlights that ripple across the water. Seagulls soar above, their cries blending with the gentle roar of the waves. The horizon stretches infinitely, where the ocean meets the sky in a seamless blend of hues. Close-ups reveal the intricate patterns of the waves, capturing the fluidity and dynamic beauty of the sea in motion."
-image = load_image(image="cogvideox_rocket.png")
-pipe = CogVideoXImageToVideoPipeline.from_pretrained(
-    "THUDM/CogVideoX-5b-I2V",
-    torch_dtype=torch.bfloat16
-)
- 
-pipe.vae.enable_tiling()
-pipe.vae.enable_slicing()
-
-video = pipe(
-    prompt=prompt,
-    image=image,
-    num_videos_per_prompt=1,
-    num_inference_steps=50,
-    num_frames=49,
-    guidance_scale=6,
-    generator=torch.Generator(device="cuda").manual_seed(42),
-).frames[0]
-
-export_to_video(video, "output.mp4", fps=8)
-```
-
-<div class="flex gap-4">
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_rocket.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">initial image</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_outrocket.gif"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">generated video</figcaption>
-  </div>
-</div>
-

docs/source/en/using-diffusers/conditional_image_generation.md

@@ -303,7 +303,7 @@ There are many types of conditioning inputs you can use, and 🤗 Diffusers supp
 
 Diffusion models are large, and the iterative nature of denoising an image is computationally expensive and intensive. But this doesn't mean you need access to powerful - or even many - GPUs to use them. There are many optimization techniques for running diffusion models on consumer and free-tier resources. For example, you can load model weights in half-precision to save GPU memory and increase speed or offload the entire model to the GPU to save even more memory.
 
-PyTorch 2.0 also supports a more memory-efficient attention mechanism called [*scaled dot product attention*](../optimization/torch2.0#scaled-dot-product-attention) that is automatically enabled if you're using PyTorch 2.0. You can combine this with [`torch.compile`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) to speed your code up even more:
+PyTorch 2.0 also supports a more memory-efficient attention mechanism called [*scaled dot product attention*](../optimization/fp16#scaled-dot-product-attention) that is automatically enabled if you're using PyTorch 2.0. You can combine this with [`torch.compile`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) to speed your code up even more:
 
 ```py
 from diffusers import AutoPipelineForText2Image
@@ -313,4 +313,4 @@ pipeline = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stab
 pipeline.unet = torch.compile(pipeline.unet, mode="reduce-overhead", fullgraph=True)
 ```
 
-For more tips on how to optimize your code to save memory and speed up inference, read the [Memory and speed](../optimization/fp16) and [Torch 2.0](../optimization/torch2.0) guides.
+For more tips on how to optimize your code to save memory and speed up inference, read the [Accelerate inference](../optimization/fp16) and [Reduce memory usage](../optimization/memory) guides.

docs/source/en/using-diffusers/img2img.md

@@ -35,7 +35,7 @@ pipeline.enable_xformers_memory_efficient_attention()
 
 <Tip>
 
-You'll notice throughout the guide, we use [`~DiffusionPipeline.enable_model_cpu_offload`] and [`~DiffusionPipeline.enable_xformers_memory_efficient_attention`], to save memory and increase inference speed. If you're using PyTorch 2.0, then you don't need to call [`~DiffusionPipeline.enable_xformers_memory_efficient_attention`] on your pipeline because it'll already be using PyTorch 2.0's native [scaled-dot product attention](../optimization/torch2.0#scaled-dot-product-attention).
+You'll notice throughout the guide, we use [`~DiffusionPipeline.enable_model_cpu_offload`] and [`~DiffusionPipeline.enable_xformers_memory_efficient_attention`], to save memory and increase inference speed. If you're using PyTorch 2.0, then you don't need to call [`~DiffusionPipeline.enable_xformers_memory_efficient_attention`] on your pipeline because it'll already be using PyTorch 2.0's native [scaled-dot product attention](../optimization/fp16#scaled-dot-product-attention).
 
 </Tip>
 
@@ -589,17 +589,17 @@ make_image_grid([init_image, depth_image, image_control_net, image_elden_ring],
 
 ## Optimize
 
-Running diffusion models is computationally expensive and intensive, but with a few optimization tricks, it is entirely possible to run them on consumer and free-tier GPUs. For example, you can use a more memory-efficient form of attention such as PyTorch 2.0's [scaled-dot product attention](../optimization/torch2.0#scaled-dot-product-attention) or [xFormers](../optimization/xformers) (you can use one or the other, but there's no need to use both). You can also offload the model to the GPU while the other pipeline components wait on the CPU.
+Running diffusion models is computationally expensive and intensive, but with a few optimization tricks, it is entirely possible to run them on consumer and free-tier GPUs. For example, you can use a more memory-efficient form of attention such as PyTorch 2.0's [scaled-dot product attention](../optimization/fp16#scaled-dot-product-attention) or [xFormers](../optimization/xformers) (you can use one or the other, but there's no need to use both). You can also offload the model to the GPU while the other pipeline components wait on the CPU.
 
 ```diff
 + pipeline.enable_model_cpu_offload()
 + pipeline.enable_xformers_memory_efficient_attention()
 ```
 
-With [`torch.compile`](../optimization/torch2.0#torchcompile), you can boost your inference speed even more by wrapping your UNet with it:
+With [`torch.compile`](../optimization/fp16#torchcompile), you can boost your inference speed even more by wrapping your UNet with it:
 
 ```py
 pipeline.unet = torch.compile(pipeline.unet, mode="reduce-overhead", fullgraph=True)
 ```
 
-To learn more, take a look at the [Reduce memory usage](../optimization/memory) and [Torch 2.0](../optimization/torch2.0) guides.
+To learn more, take a look at the [Reduce memory usage](../optimization/memory) and [Accelerate inference](../optimization/fp16) guides.

docs/source/en/using-diffusers/inpaint.md

@@ -35,7 +35,7 @@ pipeline.enable_xformers_memory_efficient_attention()
 
 <Tip>
 
-You'll notice throughout the guide, we use [`~DiffusionPipeline.enable_model_cpu_offload`] and [`~DiffusionPipeline.enable_xformers_memory_efficient_attention`], to save memory and increase inference speed. If you're using PyTorch 2.0, it's not necessary to call [`~DiffusionPipeline.enable_xformers_memory_efficient_attention`] on your pipeline because it'll already be using PyTorch 2.0's native [scaled-dot product attention](../optimization/torch2.0#scaled-dot-product-attention).
+You'll notice throughout the guide, we use [`~DiffusionPipeline.enable_model_cpu_offload`] and [`~DiffusionPipeline.enable_xformers_memory_efficient_attention`], to save memory and increase inference speed. If you're using PyTorch 2.0, it's not necessary to call [`~DiffusionPipeline.enable_xformers_memory_efficient_attention`] on your pipeline because it'll already be using PyTorch 2.0's native [scaled-dot product attention](../optimization/fp16#scaled-dot-product-attention).
 
 </Tip>
 
@@ -788,7 +788,7 @@ make_image_grid([init_image, mask_image, image, image_elden_ring], rows=2, cols=
 
 ## Optimize
 
-It can be difficult and slow to run diffusion models if you're resource constrained, but it doesn't have to be with a few optimization tricks. One of the biggest (and easiest) optimizations you can enable is switching to memory-efficient attention. If you're using PyTorch 2.0, [scaled-dot product attention](../optimization/torch2.0#scaled-dot-product-attention) is automatically enabled and you don't need to do anything else. For non-PyTorch 2.0 users, you can install and use [xFormers](../optimization/xformers)'s implementation of memory-efficient attention. Both options reduce memory usage and accelerate inference.
+It can be difficult and slow to run diffusion models if you're resource constrained, but it doesn't have to be with a few optimization tricks. One of the biggest (and easiest) optimizations you can enable is switching to memory-efficient attention. If you're using PyTorch 2.0, [scaled-dot product attention](../optimization/fp16#scaled-dot-product-attention) is automatically enabled and you don't need to do anything else. For non-PyTorch 2.0 users, you can install and use [xFormers](../optimization/xformers)'s implementation of memory-efficient attention. Both options reduce memory usage and accelerate inference.
 
 You can also offload the model to the CPU to save even more memory:
 
@@ -797,10 +797,10 @@ You can also offload the model to the CPU to save even more memory:
 + pipeline.enable_model_cpu_offload()
 ```
 
-To speed-up your inference code even more, use [`torch_compile`](../optimization/torch2.0#torchcompile). You should wrap `torch.compile` around the most intensive component in the pipeline which is typically the UNet:
+To speed-up your inference code even more, use [`torch_compile`](../optimization/fp16#torchcompile). You should wrap `torch.compile` around the most intensive component in the pipeline which is typically the UNet:
 
 ```py
 pipeline.unet = torch.compile(pipeline.unet, mode="reduce-overhead", fullgraph=True)
 ```
 
-Learn more in the [Reduce memory usage](../optimization/memory) and [Torch 2.0](../optimization/torch2.0) guides.
+Learn more in the [Reduce memory usage](../optimization/memory) and [Accelerate inference](../optimization/fp16) guides.

docs/source/en/using-diffusers/marigold_usage.md

@@ -288,7 +288,7 @@ Speeding them up can be achieved by using a more efficient attention processor:
   depth = pipe(image, num_inference_steps=1)
 ```
 
-Finally, as suggested in [Optimizations](../optimization/torch2.0#torch.compile), enabling `torch.compile` can further enhance performance depending on 
+Finally, as suggested in [Optimizations](../optimization/fp16#torchcompile), enabling `torch.compile` can further enhance performance depending on 
 the target hardware.
 However, compilation incurs a significant overhead during the first pipeline invocation, making it beneficial only when 
 the same pipeline instance is called repeatedly, such as within a loop.

docs/source/en/using-diffusers/svd.md

@@ -63,7 +63,7 @@ export_to_video(frames, "generated.mp4", fps=7)
 
 ## torch.compile
 
-You can gain a 20-25% speedup at the expense of slightly increased memory by [compiling](../optimization/torch2.0#torchcompile) the UNet.
+You can gain a 20-25% speedup at the expense of slightly increased memory by [compiling](../optimization/fp16#torchcompile) the UNet.
 
 ```diff
 - pipe.enable_model_cpu_offload()

docs/source/en/using-diffusers/text-img2vid.md

@@ -1,4 +1,4 @@
- <!--Copyright 2024 The HuggingFace Team. All rights reserved.
+<!--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
@@ -12,551 +12,436 @@ specific language governing permissions and limitations under the License.
 
 # Video generation
 
-Video generation models include a temporal dimension to bring images, or frames, together to create a video. These models are trained on large-scale datasets of high-quality text-video pairs to learn how to combine the modalities to ensure the generated video is coherent and realistic.
+Video generation models extend image generation (can be considered a 1-frame video) to also process data related to space and time. Making sure all this data - text, space, time - remain consistent and aligned from frame-to-frame is a big challenge in generating long and high-resolution videos.
 
-[Explore](https://huggingface.co/models?other=video-generation) some of the more popular open-source video generation models available from Diffusers below.
+Modern video models tackle this challenge with the diffusion transformer (DiT) architecture. This reduces computational costs and allows more efficient scaling to larger and higher-quality image and video data.
 
-<hfoptions id="popular-models">
-<hfoption id="CogVideoX">
+Check out what some of these video models are capable of below.
 
-[CogVideoX](https://huggingface.co/collections/THUDM/cogvideo-66c08e62f1685a3ade464cce) uses a 3D causal Variational Autoencoder (VAE) to compress videos along the spatial and temporal dimensions, and it includes a stack of expert transformer blocks with a 3D full attention mechanism to better capture visual, semantic, and motion information in the data.
-
-The CogVideoX family also includes models capable of generating videos from images and videos in addition to text. The image-to-video models are indicated by **I2V** in the checkpoint name, and they should be used with the [`CogVideoXImageToVideoPipeline`]. The regular checkpoints support video-to-video through the [`CogVideoXVideoToVideoPipeline`].
-
-The example below demonstrates how to generate a video from an image and text prompt with [THUDM/CogVideoX-5b-I2V](https://huggingface.co/THUDM/CogVideoX-5b-I2V).
+<hfoptions id="popular models">
+<hfoption id="Wan2.1">
 
 ```py
+# pip install ftfy
 import torch
-from diffusers import CogVideoXImageToVideoPipeline
+import numpy as np
+from diffusers import AutoModel, WanPipeline
+from diffusers.hooks.group_offloading import apply_group_offloading
 from diffusers.utils import export_to_video, load_image
+from transformers import UMT5EncoderModel
 
-prompt = "A vast, shimmering ocean flows gracefully under a twilight sky, its waves undulating in a mesmerizing dance of blues and greens. The surface glints with the last rays of the setting sun, casting golden highlights that ripple across the water. Seagulls soar above, their cries blending with the gentle roar of the waves. The horizon stretches infinitely, where the ocean meets the sky in a seamless blend of hues. Close-ups reveal the intricate patterns of the waves, capturing the fluidity and dynamic beauty of the sea in motion."
-image = load_image(image="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_rocket.png")
-pipe = CogVideoXImageToVideoPipeline.from_pretrained(
-    "THUDM/CogVideoX-5b-I2V",
-    torch_dtype=torch.bfloat16
+text_encoder = UMT5EncoderModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
+transformer = AutoModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+
+# group-offloading
+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
 )
 
-# reduce memory requirements 
-pipe.vae.enable_tiling()
-pipe.vae.enable_slicing()
+pipeline = WanPipeline.from_pretrained(
+    "Wan-AI/Wan2.1-T2V-14B-Diffusers",
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    torch_dtype=torch.bfloat16
+)
+pipeline.to("cuda")
 
-video = pipe(
+prompt = """
+The camera rushes from far to near in a low-angle shot, 
+revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
+for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
+Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic 
+shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
+"""
+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
+"""
+
+output = pipeline(
     prompt=prompt,
-    image=image,
-    num_videos_per_prompt=1,
-    num_inference_steps=50,
-    num_frames=49,
-    guidance_scale=6,
-    generator=torch.Generator(device="cuda").manual_seed(42),
+    negative_prompt=negative_prompt,
+    num_frames=81,
+    guidance_scale=5.0,
 ).frames[0]
-export_to_video(video, "output.mp4", fps=8)
+export_to_video(output, "output.mp4", fps=16)
 ```
 
-<div class="flex gap-4">
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_rocket.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">initial image</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_outrocket.gif"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">generated video</figcaption>
-  </div>
-</div>
-
 </hfoption>
 <hfoption id="HunyuanVideo">
 
-> [!TIP]
-> HunyuanVideo is a 13B parameter model and requires a lot of memory. Refer to the HunyuanVideo [Quantization](../api/pipelines/hunyuan_video#quantization) guide to learn how to quantize the model. CogVideoX and LTX-Video are more lightweight options that can still generate high-quality videos.
-
-[HunyuanVideo](https://huggingface.co/tencent/HunyuanVideo) features a dual-stream to single-stream diffusion transformer (DiT) for learning video and text tokens separately, and then subsequently concatenating the video and text tokens to combine their information. A single multimodal large language model (MLLM) serves as the text encoder, and videos are also spatio-temporally compressed with a 3D causal VAE.
-
 ```py
 import torch
-from diffusers import HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
+from diffusers importAutoModel, HunyuanVideoPipeline
+from diffusers.quantizers import PipelineQuantizationConfig
 from diffusers.utils import export_to_video
 
-transformer = HunyuanVideoTransformer3DModel.from_pretrained(
-    "hunyuanvideo-community/HunyuanVideo", subfolder="transformer", torch_dtype=torch.bfloat16
-)
-pipe = HunyuanVideoPipeline.from_pretrained(
-  "hunyuanvideo-community/HunyuanVideo", transformer=transformer, torch_dtype=torch.float16
+# quantize weights to int4 with bitsandbytes
+pipeline_quant_config = PipelineQuantizationConfig(
+  quant_backend="bitsandbytes_4bit",
+  quant_kwargs={
+    "load_in_4bit": True,
+    "bnb_4bit_quant_type": "nf4",
+    "bnb_4bit_compute_dtype": torch.bfloat16
+    },
+  components_to_quantize=["transformer"]
 )
 
-# reduce memory requirements
-pipe.vae.enable_tiling()
-pipe.to("cuda")
+pipeline = HunyuanVideoPipeline.from_pretrained(
+    "hunyuanvideo-community/HunyuanVideo",
+    quantization_config=pipeline_quant_config,
+    torch_dtype=torch.bfloat16,
+)
 
-video = pipe(
-    prompt="A cat walks on the grass, realistic",
-    height=320,
-    width=512,
-    num_frames=61,
-    num_inference_steps=30,
-).frames[0]
+# model-offloading and tiling
+pipeline.enable_model_cpu_offload()
+pipeline.vae.enable_tiling()
+
+prompt = "A fluffy teddy bear sits on a bed of soft pillows surrounded by children's toys."
+video = pipeline(prompt=prompt, num_frames=61, num_inference_steps=30).frames[0]
 export_to_video(video, "output.mp4", fps=15)
 ```
 
-<div class="flex justify-center">
-  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/hunyuan-video-output.gif"/>
-</div>
-
 </hfoption>
 <hfoption id="LTX-Video">
 
-[LTX-Video (LTXV)](https://huggingface.co/Lightricks/LTX-Video) is a diffusion transformer (DiT) with a focus on speed. It generates 768x512 resolution videos at 24 frames per second (fps), enabling near real-time generation of high-quality videos. LTXV is relatively lightweight compared to other modern video generation models, making it possible to run on consumer GPUs.
+```py
+import torch
+from diffusers import LTXPipeline, AutoModel
+from diffusers.hooks import apply_group_offloading
+from diffusers.utils import export_to_video
+
+# fp8 layerwise weight-casting
+transformer = AutoModel.from_pretrained(
+    "Lightricks/LTX-Video",
+    subfolder="transformer",
+    torch_dtype=torch.bfloat16
+)
+transformer.enable_layerwise_casting(
+    storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16
+)
+
+pipeline = LTXPipeline.from_pretrained("Lightricks/LTX-Video", transformer=transformer, torch_dtype=torch.bfloat16)
+
+# group-offloading
+onload_device = torch.device("cuda")
+offload_device = torch.device("cpu")
+pipeline.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True)
+apply_group_offloading(pipeline.text_encoder, onload_device=onload_device, offload_type="block_level", num_blocks_per_group=2)
+apply_group_offloading(pipeline.vae, onload_device=onload_device, offload_type="leaf_level")
+
+prompt = """
+A woman with long brown hair and light skin smiles at another woman with long blonde hair. The woman with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek. The camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be real-life footage
+"""
+negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+
+video = pipeline(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    width=768,
+    height=512,
+    num_frames=161,
+    decode_timestep=0.03,
+    decode_noise_scale=0.025,
+    num_inference_steps=50,
+).frames[0]
+export_to_video(video, "output.mp4", fps=24)
+```
+
+</hfoption>
+</hfoptions>
+
+This guide will cover video generation basics such as which parameters to configure and how to reduce their memory usage.
+
+> [!TIP]
+> If you're interested in learning more about how to use a specific model, please refer to their pipeline API model card.
+
+## Pipeline parameters
+
+There are several parameters to configure in the pipeline that'll affect video generation quality or speed. Experimenting with different parameter values is important for discovering the appropriate quality and speed tradeoff.
+
+### num_frames
+
+A frame is a still image that is played in a sequence of other frames to create motion or a video. Control the number of frames generated per second with `num_frames`. Increasing `num_frames` increases perceived motion smoothness and visual coherence, making it especially important for videos with dynamic content. A higher `num_frames` value also increases video duration.
+
+Some video models require more specific `num_frames` values for inference. For example, [`HunyuanVideoPipeline`] recommends calculating the `num_frames` with `(4 * num_frames) +1`. Always check a pipelines API model card to see if there is a recommended value.
 
 ```py
 import torch
 from diffusers import LTXPipeline
 from diffusers.utils import export_to_video
 
-pipe = LTXPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16).to("cuda")
+pipeline = LTXPipeline.from_pretrained(
+    "Lightricks/LTX-Video", torch_dtype=torch.bfloat16
+).to("cuda")
 
-prompt = "A man walks towards a window, looks out, and then turns around. He has short, dark hair, dark skin, and is wearing a brown coat over a red and gray scarf. He walks from left to right towards a window, his gaze fixed on something outside. The camera follows him from behind at a medium distance. The room is brightly lit, with white walls and a large window covered by a white curtain. As he approaches the window, he turns his head slightly to the left, then back to the right. He then turns his entire body to the right, facing the window. The camera remains stationary as he stands in front of the window. The scene is captured in real-life footage."
-video = pipe(
+prompt = """
+A woman with long brown hair and light skin smiles at another woman with long blonde hair. The woman 
+with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek. The 
+camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and 
+natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be 
+real-life footage
+"""
+
+video = pipeline(
     prompt=prompt,
-    width=704,
-    height=480,
+    negative_prompt=negative_prompt,
+    width=768,
+    height=512,
     num_frames=161,
+    decode_timestep=0.03,
+    decode_noise_scale=0.025,
     num_inference_steps=50,
 ).frames[0]
 export_to_video(video, "output.mp4", fps=24)
 ```
 
-<div class="flex justify-center">
-  <img src="https://huggingface.co/Lightricks/LTX-Video/resolve/main/media/ltx-video_example_00014.gif"/>
-</div>
+### guidance_scale
 
-</hfoption>
-<hfoption id="Mochi-1">
-
-> [!TIP]
-> Mochi-1 is a 10B parameter model and requires a lot of memory. Refer to the Mochi [Quantization](../api/pipelines/mochi#quantization) guide to learn how to quantize the model. CogVideoX and LTX-Video are more lightweight options that can still generate high-quality videos.
-
-[Mochi-1](https://huggingface.co/genmo/mochi-1-preview) introduces the Asymmetric Diffusion Transformer (AsymmDiT) and Asymmetric Variational Autoencoder (AsymmVAE) to reduces memory requirements. AsymmVAE causally compresses videos 128x to improve memory efficiency, and AsymmDiT jointly attends to the compressed video tokens and user text tokens. This model is noted for generating videos with high-quality motion dynamics and strong prompt adherence.
+Guidance scale or "cfg" controls how closely the generated frames adhere to the input conditioning (text, image or both). Increasing `guidance_scale` generates frames that resemble the input conditions more closely and includes finer details, but risk introducing artifacts and reducing output diversity. Lower `guidance_scale` values encourages looser prompt adherence and increased output variety, but details may not be as great. If it's too low, it may ignore your prompt entirely and generate random noise.
 
 ```py
 import torch
-from diffusers import MochiPipeline
+from diffusers import CogVideoXPipeline, CogVideoXTransformer3DModel
 from diffusers.utils import export_to_video
 
-pipe = MochiPipeline.from_pretrained("genmo/mochi-1-preview", variant="bf16", torch_dtype=torch.bfloat16)
+pipeline = CogVideoXPipeline.from_pretrained(
+  "THUDM/CogVideoX-2b",
+  torch_dtype=torch.float16
+).to("cuda")
 
-# reduce memory requirements
-pipe.enable_model_cpu_offload()
-pipe.enable_vae_tiling()
+prompt = """
+A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over
+a plush, blue carpet that mimics the waves of the sea. The ship's hull is painted a rich brown, 
+with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an 
+oceanic expanse. Surrounding the ship are various other toys and children's items, hinting at 
+a playful environment. The scene captures the innocence and imagination of childhood, 
+with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting.
+"""
 
-prompt = "Close-up of a chameleon's eye, with its scaly skin changing color. Ultra high resolution 4k."
-video = pipe(prompt, num_frames=84).frames[0]
-export_to_video(video, "output.mp4", fps=30)
-```
-
-<div class="flex justify-center">
-  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/mochi-video-output.gif"/>
-</div>
-
-</hfoption>
-<hfoption id="StableVideoDiffusion">
-
-[StableVideoDiffusion (SVD)](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt) is based on the Stable Diffusion 2.1 model and it is trained on images, then low-resolution videos, and finally a smaller dataset of high-resolution videos. This model generates a short 2-4 second video from an initial image.
-
-```py
-import torch
-from diffusers import StableVideoDiffusionPipeline
-from diffusers.utils import load_image, export_to_video
-
-pipeline = StableVideoDiffusionPipeline.from_pretrained(
-    "stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16"
-)
-
-# reduce memory requirements
-pipeline.enable_model_cpu_offload()
-
-image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png")
-image = image.resize((1024, 576))
-
-generator = torch.manual_seed(42)
-frames = pipeline(image, decode_chunk_size=8, generator=generator).frames[0]
-export_to_video(frames, "generated.mp4", fps=7)
-```
-
-<div class="flex gap-4">
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">initial image</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/output_rocket.gif"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">generated video</figcaption>
-  </div>
-</div>
-
-</hfoption>
-<hfoption id="AnimateDiff">
-
-[AnimateDiff](https://huggingface.co/guoyww/animatediff) is an adapter model that inserts a motion module into a pretrained diffusion model to animate an image. The adapter is trained on video clips to learn motion which is used to condition the generation process to create a video. It is faster and easier to only train the adapter and it can be loaded into most diffusion models, effectively turning them into “video models”.
-
-Load a `MotionAdapter` and pass it to the [`AnimateDiffPipeline`].
-
-```py
-import torch
-from diffusers import AnimateDiffPipeline, DDIMScheduler, MotionAdapter
-from diffusers.utils import export_to_gif
-
-adapter = MotionAdapter.from_pretrained("guoyww/animatediff-motion-adapter-v1-5-2", torch_dtype=torch.float16)
-pipeline = AnimateDiffPipeline.from_pretrained("emilianJR/epiCRealism", motion_adapter=adapter, torch_dtype=torch.float16)
-scheduler = DDIMScheduler.from_pretrained(
-    "emilianJR/epiCRealism",
-    subfolder="scheduler",
-    clip_sample=False,
-    timestep_spacing="linspace",
-    beta_schedule="linear",
-    steps_offset=1,
-)
-pipeline.scheduler = scheduler
-
-# reduce memory requirements
-pipeline.enable_vae_slicing()
-pipeline.enable_model_cpu_offload()
-
-output = pipeline(
-    prompt="A space rocket with trails of smoke behind it launching into space from the desert, 4k, high resolution",
-    negative_prompt="bad quality, worse quality, low resolution",
-    num_frames=16,
-    guidance_scale=7.5,
-    num_inference_steps=50,
-    generator=torch.Generator("cpu").manual_seed(49),
-)
-frames = output.frames[0]
-export_to_gif(frames, "animation.gif")
-```
-
-<div class="flex justify-center">
-  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff.gif"/>
-</div>
-
-</hfoption>
-</hfoptions>
-
-## Configure model parameters
-
-There are a few important parameters you can configure in the pipeline that'll affect the video generation process and quality. Let's take a closer look at what these parameters do and how changing them affects the output.
-
-### Number of frames
-
-The `num_frames` parameter determines how many video frames are generated per second. A frame is an image that is played in a sequence of other frames to create motion or a video. This affects video length because the pipeline generates a certain number of frames per second (check a pipeline's API reference for the default value). To increase the video duration, you'll need to increase the `num_frames` parameter.
-
-```py
-import torch
-from diffusers import StableVideoDiffusionPipeline
-from diffusers.utils import load_image, export_to_video
-
-pipeline = StableVideoDiffusionPipeline.from_pretrained(
-    "stabilityai/stable-video-diffusion-img2vid", torch_dtype=torch.float16, variant="fp16"
-)
-pipeline.enable_model_cpu_offload()
-
-image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png")
-image = image.resize((1024, 576))
-
-generator = torch.manual_seed(42)
-frames = pipeline(image, decode_chunk_size=8, generator=generator, num_frames=25).frames[0]
-export_to_video(frames, "generated.mp4", fps=7)
-```
-
-<div class="flex gap-4">
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/num_frames_14.gif"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">num_frames=14</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/num_frames_25.gif"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">num_frames=25</figcaption>
-  </div>
-</div>
-
-### Guidance scale
-
-The `guidance_scale` parameter controls how closely aligned the generated video and text prompt or initial image is. A higher `guidance_scale` value means your generated video is more aligned with the text prompt or initial image, while a lower `guidance_scale` value means your generated video is less aligned which could give the model more "creativity" to interpret the conditioning input.
-
-<Tip>
-
-SVD uses the `min_guidance_scale` and `max_guidance_scale` parameters for applying guidance to the first and last frames respectively.
-
-</Tip>
-
-```py
-import torch
-from diffusers import I2VGenXLPipeline
-from diffusers.utils import export_to_gif, load_image
-
-pipeline = I2VGenXLPipeline.from_pretrained("ali-vilab/i2vgen-xl", torch_dtype=torch.float16, variant="fp16")
-pipeline.enable_model_cpu_offload()
-
-image_url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/i2vgen_xl_images/img_0009.png"
-image = load_image(image_url).convert("RGB")
-
-prompt = "Papers were floating in the air on a table in the library"
-negative_prompt = "Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms"
-generator = torch.manual_seed(0)
-
-frames = pipeline(
-    prompt=prompt,
-    image=image,
-    num_inference_steps=50,
-    negative_prompt=negative_prompt,
-    guidance_scale=1.0,
-    generator=generator
+video = pipeline(
+  prompt=prompt,
+  guidance_scale=6,
+  num_inference_steps=50
 ).frames[0]
-export_to_gif(frames, "i2v.gif")
+export_to_video(video, "output.mp4", fps=8)
 ```
 
-<div class="flex gap-4">
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/i2vgen-xl-example.gif"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">guidance_scale=9.0</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guidance_scale_1.0.gif"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">guidance_scale=1.0</figcaption>
-  </div>
-</div>
+### negative_prompt
 
-### Negative prompt
-
-A negative prompt deters the model from generating things you don’t want it to. This parameter is commonly used to improve overall generation quality by removing poor or bad features such as “low resolution” or “bad details”.
+A negative prompt is useful for excluding things you don't want to see in the generated video. It is commonly used to refine the quality and alignment of the generated video by pushing the model away from undesirable elements like "blurry, distorted, ugly". This can create cleaner and more focused videos.
 
 ```py
+# pip install ftfy
 import torch
-from diffusers import AnimateDiffPipeline, DDIMScheduler, MotionAdapter
-from diffusers.utils import export_to_gif
+from diffusers import WanPipeline
+from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
+from diffusers.utils import export_to_video
 
-adapter = MotionAdapter.from_pretrained("guoyww/animatediff-motion-adapter-v1-5-2", torch_dtype=torch.float16)
-
-pipeline = AnimateDiffPipeline.from_pretrained("emilianJR/epiCRealism", motion_adapter=adapter, torch_dtype=torch.float16)
-scheduler = DDIMScheduler.from_pretrained(
-    "emilianJR/epiCRealism",
-    subfolder="scheduler",
-    clip_sample=False,
-    timestep_spacing="linspace",
-    beta_schedule="linear",
-    steps_offset=1,
+vae = AutoencoderKLWan.from_pretrained(
+  "Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="vae", torch_dtype=torch.float32
 )
-pipeline.scheduler = scheduler
-pipeline.enable_vae_slicing()
+pipeline = WanPipeline.from_pretrained(
+  "Wan-AI/Wan2.1-T2V-14B-Diffusers", vae=vae, torch_dtype=torch.bfloat16
+)
+pipeline.scheduler = UniPCMultistepScheduler.from_config(
+  pipeline.scheduler.config, flow_shift=5.0
+)
+pipeline.to("cuda")
+
+pipeline.load_lora_weights("benjamin-paine/steamboat-willie-14b", adapter_name="steamboat-willie")
+pipeline.set_adapters("steamboat-willie")
+
 pipeline.enable_model_cpu_offload()
 
+# use "steamboat willie style" to trigger the LoRA
+prompt = """
+steamboat willie style, golden era animation, The camera rushes from far to near in a low-angle shot, 
+revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
+for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
+Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts 
+dynamic shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
+"""
+
 output = pipeline(
-    prompt="360 camera shot of a sushi roll in a restaurant",
-    negative_prompt="Distorted, discontinuous, ugly, blurry, low resolution, motionless, static",
-    num_frames=16,
-    guidance_scale=7.5,
-    num_inference_steps=50,
-    generator=torch.Generator("cpu").manual_seed(0),
+  prompt=prompt,
+  num_frames=81,
+  guidance_scale=5.0,
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
+```
+
+## Reduce memory usage
+
+Recent video models like [`HunyuanVideoPipeline`] and [`WanPipeline`], which have 10B+ parameters, require a lot of memory and it often exceeds the memory availabe on consumer hardware. Diffusers offers several techniques for reducing the memory requirements of these large models.
+
+> [!TIP]
+> Refer to the [Reduce memory usage](../optimization/memory) guide for more details about other memory saving techniques.
+
+One of these techniques is [group-offloading](../optimization/memory#group-offloading), which offloads groups of internal model layers (such as `torch.nn.Sequential`) to the CPU when it isn't being used. These layers are only loaded when they're needed for computation to avoid storing **all** the model components on the GPU. For a 14B parameter model like [`WanPipeline`], group-offloading can lower the required memory to ~13GB of VRAM.
+
+```py
+# pip install ftfy
+import torch
+import numpy as np
+from diffusers import AutoModel, WanPipeline
+from diffusers.hooks.group_offloading import apply_group_offloading
+from diffusers.utils import export_to_video, load_image
+from transformers import UMT5EncoderModel
+
+text_encoder = UMT5EncoderModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
+transformer = AutoModel.from_pretrained("Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+
+# group-offloading
+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
 )
-frames = output.frames[0]
-export_to_gif(frames, "animation.gif")
+transformer.enable_group_offload(
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="leaf_level",
+    use_stream=True
+)
+
+pipeline = WanPipeline.from_pretrained(
+    "Wan-AI/Wan2.1-T2V-14B-Diffusers",
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    torch_dtype=torch.bfloat16
+)
+pipeline.to("cuda")
+
+prompt = """
+The camera rushes from far to near in a low-angle shot, 
+revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
+for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
+Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic 
+shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
+"""
+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
+"""
+
+output = pipeline(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    num_frames=81,
+    guidance_scale=5.0,
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
 ```
 
-<div class="flex gap-4">
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff_no_neg.gif"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">no negative prompt</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff_neg.gif"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">negative prompt applied</figcaption>
-  </div>
-</div>
+Another option for reducing memory is to consider quantizing a model, which stores the model weights in a lower precision data type. However, quantization may impact video quality depending on the specific video model. Refer to the quantization [Overivew](../quantization/overview) to learn more about the different supported quantization backends.
 
-### Model-specific parameters
-
-There are some pipeline parameters that are unique to each model such as adjusting the motion in a video or adding noise to the initial image.
-
-<hfoptions id="special-parameters">
-<hfoption id="Stable Video Diffusion">
-
-Stable Video Diffusion provides additional micro-conditioning for the frame rate with the `fps` parameter and for motion with the `motion_bucket_id` parameter. Together, these parameters allow for adjusting the amount of motion in the generated video.
-
-There is also a `noise_aug_strength` parameter that increases the amount of noise added to the initial image. Varying this parameter affects how similar the generated video and initial image are. A higher `noise_aug_strength` also increases the amount of motion. To learn more, read the [Micro-conditioning](../using-diffusers/svd#micro-conditioning) guide.
-
-</hfoption>
-<hfoption id="Text2Video-Zero">
-
-Text2Video-Zero computes the amount of motion to apply to each frame from randomly sampled latents. You can use the `motion_field_strength_x` and `motion_field_strength_y` parameters to control the amount of motion to apply to the x and y-axes of the video. The parameters `t0` and `t1` are the timesteps to apply motion to the latents.
-
-</hfoption>
-</hfoptions>
-
-## Control video generation
-
-Video generation can be controlled similar to how text-to-image, image-to-image, and inpainting can be controlled with a [`ControlNetModel`]. The only difference is you need to use the [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor`] so each frame attends to the first frame.
-
-### Text2Video-Zero
-
-Text2Video-Zero video generation can be conditioned on pose and edge images for even greater control over a subject's motion in the generated video or to preserve the identity of a subject/object in the video. You can also use Text2Video-Zero with [InstructPix2Pix](../api/pipelines/pix2pix) for editing videos with text.
-
-<hfoptions id="t2v-zero">
-<hfoption id="pose control">
-
-Start by downloading a video and extracting the pose images from it.
+The example below uses [bitsandbytes](../quantization/bitsandbytes) to quantize a model.
 
 ```py
-from huggingface_hub import hf_hub_download
-from PIL import Image
-import imageio
+# pip install ftfy
 
-filename = "__assets__/poses_skeleton_gifs/dance1_corr.mp4"
-repo_id = "PAIR/Text2Video-Zero"
-video_path = hf_hub_download(repo_type="space", repo_id=repo_id, filename=filename)
+import torch
+from diffusers import WanPipeline
+from diffusers import AutoModel, WanPipeline
+from diffusers.quantizers import PipelineQuantizationConfig
+from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
+from transformers import UMT5EncoderModel
+from diffusers.utils import export_to_video
 
-reader = imageio.get_reader(video_path, "ffmpeg")
-frame_count = 8
-pose_images = [Image.fromarray(reader.get_data(i)) for i in range(frame_count)]
+# quantize transformer and text encoder weights with bitsandbytes
+pipeline_quant_config = PipelineQuantizationConfig(
+  quant_backend="bitsandbytes_4bit",
+  quant_kwargs={"load_in_4bit": True},
+  components_to_quantize=["transformer", "text_encoder"]
+)
+
+vae = AutoModel.from_pretrained(
+  "Wan-AI/Wan2.1-T2V-14B-Diffusers", subfolder="vae", torch_dtype=torch.float32
+)
+pipeline = WanPipeline.from_pretrained(
+  "Wan-AI/Wan2.1-T2V-14B-Diffusers", vae=vae, quantization_config=pipeline_quant_config, torch_dtype=torch.bfloat16
+)
+pipeline.scheduler = UniPCMultistepScheduler.from_config(
+  pipeline.scheduler.config, flow_shift=5.0
+)
+pipeline.to("cuda")
+
+pipeline.load_lora_weights("benjamin-paine/steamboat-willie-14b", adapter_name="steamboat-willie")
+pipeline.set_adapters("steamboat-willie")
+
+pipeline.enable_model_cpu_offload()
+
+# use "steamboat willie style" to trigger the LoRA
+prompt = """
+steamboat willie style, golden era animation, The camera rushes from far to near in a low-angle shot, 
+revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
+for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
+Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts 
+dynamic shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
+"""
+
+output = pipeline(
+  prompt=prompt,
+  num_frames=81,
+  guidance_scale=5.0,
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
 ```
 
-Load a [`ControlNetModel`] for pose estimation and a checkpoint into the [`StableDiffusionControlNetPipeline`]. Then you'll use the [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor`] for the UNet and ControlNet.
+## Inference speed
+
+[torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial_.html) can speedup inference by using optimized kernels. Compilation takes longer the first time, but once compiled, it is much faster. It is best to compile the pipeline once, and then use the pipeline multiple times without changing anything. A change, such as in the image size, triggers recompilation.
+
+The example below compiles the transformer in the pipeline and uses the `"max-autotune"` mode to maximize performance.
 
 ```py
 import torch
-from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
-from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero import CrossFrameAttnProcessor
+from diffusers import CogVideoXPipeline, CogVideoXTransformer3DModel
+from diffusers.utils import export_to_video
 
-model_id = "stable-diffusion-v1-5/stable-diffusion-v1-5"
-controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-openpose", torch_dtype=torch.float16)
-pipeline = StableDiffusionControlNetPipeline.from_pretrained(
-    model_id, controlnet=controlnet, torch_dtype=torch.float16
+pipeline = CogVideoXPipeline.from_pretrained(
+  "THUDM/CogVideoX-2b",
+  torch_dtype=torch.float16
 ).to("cuda")
 
-pipeline.unet.set_attn_processor(CrossFrameAttnProcessor(batch_size=2))
-pipeline.controlnet.set_attn_processor(CrossFrameAttnProcessor(batch_size=2))
-```
+# torch.compile
+pipeline.transformer.to(memory_format=torch.channels_last)
+pipeline.transformer = torch.compile(
+    pipeline.transformer, mode="max-autotune", fullgraph=True
+)
 
-Fix the latents for all the frames, and then pass your prompt and extracted pose images to the model to generate a video.
+prompt = """
+A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. 
+The ship's hull is painted a rich brown, with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an oceanic expanse. 
+Surrounding the ship are various other toys and children's items, hinting at a playful environment. The scene captures the innocence and imagination of childhood, 
+with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting.
+"""
 
-```py
-latents = torch.randn((1, 4, 64, 64), device="cuda", dtype=torch.float16).repeat(len(pose_images), 1, 1, 1)
-
-prompt = "Darth Vader dancing in a desert"
-result = pipeline(prompt=[prompt] * len(pose_images), image=pose_images, latents=latents).images
-imageio.mimsave("video.mp4", result, fps=4)
-```
-
-</hfoption>
-<hfoption id="edge control">
-
-Download a video and extract the edges from it.
-
-```py
-from huggingface_hub import hf_hub_download
-from PIL import Image
-import imageio
-
-filename = "__assets__/poses_skeleton_gifs/dance1_corr.mp4"
-repo_id = "PAIR/Text2Video-Zero"
-video_path = hf_hub_download(repo_type="space", repo_id=repo_id, filename=filename)
-
-reader = imageio.get_reader(video_path, "ffmpeg")
-frame_count = 8
-pose_images = [Image.fromarray(reader.get_data(i)) for i in range(frame_count)]
-```
-
-Load a [`ControlNetModel`] for canny edge and a checkpoint into the [`StableDiffusionControlNetPipeline`]. Then you'll use the [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor`] for the UNet and ControlNet.
-
-```py
-import torch
-from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
-from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero import CrossFrameAttnProcessor
-
-model_id = "stable-diffusion-v1-5/stable-diffusion-v1-5"
-controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
-pipeline = StableDiffusionControlNetPipeline.from_pretrained(
-    model_id, controlnet=controlnet, torch_dtype=torch.float16
-).to("cuda")
-
-pipeline.unet.set_attn_processor(CrossFrameAttnProcessor(batch_size=2))
-pipeline.controlnet.set_attn_processor(CrossFrameAttnProcessor(batch_size=2))
-```
-
-Fix the latents for all the frames, and then pass your prompt and extracted edge images to the model to generate a video.
-
-```py
-latents = torch.randn((1, 4, 64, 64), device="cuda", dtype=torch.float16).repeat(len(pose_images), 1, 1, 1)
-
-prompt = "Darth Vader dancing in a desert"
-result = pipeline(prompt=[prompt] * len(pose_images), image=pose_images, latents=latents).images
-imageio.mimsave("video.mp4", result, fps=4)
-```
-
-</hfoption>
-<hfoption id="InstructPix2Pix">
-
-InstructPix2Pix allows you to use text to describe the changes you want to make to the video. Start by downloading and reading a video.
-
-```py
-from huggingface_hub import hf_hub_download
-from PIL import Image
-import imageio
-
-filename = "__assets__/pix2pix video/camel.mp4"
-repo_id = "PAIR/Text2Video-Zero"
-video_path = hf_hub_download(repo_type="space", repo_id=repo_id, filename=filename)
-
-reader = imageio.get_reader(video_path, "ffmpeg")
-frame_count = 8
-video = [Image.fromarray(reader.get_data(i)) for i in range(frame_count)]
-```
-
-Load the [`StableDiffusionInstructPix2PixPipeline`] and set the [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor`] for the UNet.
-
-```py
-import torch
-from diffusers import StableDiffusionInstructPix2PixPipeline
-from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero import CrossFrameAttnProcessor
-
-pipeline = StableDiffusionInstructPix2PixPipeline.from_pretrained("timbrooks/instruct-pix2pix", torch_dtype=torch.float16).to("cuda")
-pipeline.unet.set_attn_processor(CrossFrameAttnProcessor(batch_size=3))
-```
-
-Pass a prompt describing the change you want to apply to the video.
-
-```py
-prompt = "make it Van Gogh Starry Night style"
-result = pipeline(prompt=[prompt] * len(video), image=video).images
-imageio.mimsave("edited_video.mp4", result, fps=4)
-```
-
-</hfoption>
-</hfoptions>
-
-## Optimize
-
-Video generation requires a lot of memory because you're generating many video frames at once. You can reduce your memory requirements at the expense of some inference speed. Try:
-
-1. offloading pipeline components that are no longer needed to the CPU
-2. feed-forward chunking runs the feed-forward layer in a loop instead of all at once
-3. break up the number of frames the VAE has to decode into chunks instead of decoding them all at once
-
-```diff
-- pipeline.enable_model_cpu_offload()
-- frames = pipeline(image, decode_chunk_size=8, generator=generator).frames[0]
-+ pipeline.enable_model_cpu_offload()
-+ pipeline.unet.enable_forward_chunking()
-+ frames = pipeline(image, decode_chunk_size=2, generator=generator, num_frames=25).frames[0]
-```
-
-If memory is not an issue and you want to optimize for speed, try wrapping the UNet with [`torch.compile`](../optimization/torch2.0#torchcompile).
-
-```diff
-- pipeline.enable_model_cpu_offload()
-+ pipeline.to("cuda")
-+ pipeline.unet = torch.compile(pipeline.unet, mode="reduce-overhead", fullgraph=True)
-```
-
-## 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) to learn more about supported quantization backends (bitsandbytes, torchao, gguf) and selecting a quantization backend that supports your use case.
+video = pipeline(
+  prompt=prompt,
+  guidance_scale=6,
+  num_inference_steps=50
+).frames[0]
+export_to_video(video, "output.mp4", fps=8)
+```

examples/advanced_diffusion_training/README.md

@@ -128,6 +128,7 @@ You can also load a dataset straight from by specifying it's name in `dataset_na
 Look [here](https://huggingface.co/blog/sdxl_lora_advanced_script#custom-captioning) for more info on creating/loading your own caption dataset.
 
 - **optimizer**: for this example, we'll use [prodigy](https://huggingface.co/blog/sdxl_lora_advanced_script#adaptive-optimizers) - an adaptive optimizer
+  - To use Prodigy, please make sure to install the prodigyopt library: `pip install prodigyopt`
 - **pivotal tuning**
 - **min SNR gamma**
 

examples/advanced_diffusion_training/README_flux.md

@@ -143,7 +143,8 @@ Now we'll simply specify the name of the dataset and caption column (in this cas
 You can also load a dataset straight from by specifying it's name in `dataset_name`.
 Look [here](https://huggingface.co/blog/sdxl_lora_advanced_script#custom-captioning) for more info on creating/loading your own caption dataset.
 
-- **optimizer**: for this example, we'll use [prodigy](https://huggingface.co/blog/sdxl_lora_advanced_script#adaptive-optimizers) - an adaptive optimizer
+- **optimizer**: for this example, we'll use [prodigy](https://huggingface.co/blog/sdxl_lora_advanced_script#adaptive-optimizers) - an adaptive optimizer 
+    - To use Prodigy, please make sure to install the prodigyopt library: `pip install prodigyopt`
 - **pivotal tuning**
 
 ### Example #1: Pivotal tuning

examples/cogvideo/train_cogvideox_image_to_video_lora.py

@@ -555,7 +555,7 @@ class VideoDataset(Dataset):
 
         if any(not path.is_file() for path in instance_videos):
             raise ValueError(
-                "Expected '--video_column' to be a path to a file in `--instance_data_root` containing line-separated paths to video data but found atleast one path that is not a valid file."
+                "Expected '--video_column' to be a path to a file in `--instance_data_root` containing line-separated paths to video data but found at least one path that is not a valid file."
             )
 
         return instance_prompts, instance_videos

examples/cogvideo/train_cogvideox_lora.py

@@ -539,7 +539,7 @@ class VideoDataset(Dataset):
 
         if any(not path.is_file() for path in instance_videos):
             raise ValueError(
-                "Expected '--video_column' to be a path to a file in `--instance_data_root` containing line-separated paths to video data but found atleast one path that is not a valid file."
+                "Expected '--video_column' to be a path to a file in `--instance_data_root` containing line-separated paths to video data but found at least one path that is not a valid file."
             )
 
         return instance_prompts, instance_videos

examples/controlnet/train_controlnet.py

@@ -178,11 +178,11 @@ def log_validation(
         else:
             logger.warning(f"image logging not implemented for {tracker.name}")
 
-        del pipeline
-        gc.collect()
-        torch.cuda.empty_cache()
+    del pipeline
+    gc.collect()
+    torch.cuda.empty_cache()
 
-        return image_logs
+    return image_logs
 
 
 def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):

examples/controlnet/train_controlnet_flux.py

@@ -192,9 +192,9 @@ def log_validation(
         else:
             logger.warning(f"image logging not implemented for {tracker.name}")
 
-        del pipeline
-        free_memory()
-        return image_logs
+    del pipeline
+    free_memory()
+    return image_logs
 
 
 def save_model_card(repo_id: str, image_logs=None, base_model=str, repo_folder=None):

examples/controlnet/train_controlnet_sd3.py

@@ -199,13 +199,13 @@ def log_validation(controlnet, args, accelerator, weight_dtype, step, is_final_v
         else:
             logger.warning(f"image logging not implemented for {tracker.name}")
 
-        del pipeline
-        free_memory()
+    del pipeline
+    free_memory()
 
-        if not is_final_validation:
-            controlnet.to(accelerator.device)
+    if not is_final_validation:
+        controlnet.to(accelerator.device)
 
-        return image_logs
+    return image_logs
 
 
 # Copied from dreambooth sd3 example

examples/controlnet/train_controlnet_sdxl.py

@@ -201,11 +201,11 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
         else:
             logger.warning(f"image logging not implemented for {tracker.name}")
 
-        del pipeline
-        gc.collect()
-        torch.cuda.empty_cache()
+    del pipeline
+    gc.collect()
+    torch.cuda.empty_cache()
 
-        return image_logs
+    return image_logs
 
 
 def import_model_class_from_model_name_or_path(

examples/dreambooth/README_flux.md

@@ -134,7 +134,7 @@ Note also that we use PEFT library as backend for LoRA training, make sure to ha
 Prodigy is an adaptive optimizer that dynamically adjusts the learning rate learned parameters based on past gradients, allowing for more efficient convergence. 
 By using prodigy we can "eliminate" the need for manual learning rate tuning. read more [here](https://huggingface.co/blog/sdxl_lora_advanced_script#adaptive-optimizers).
 
-to use prodigy, specify
+to use prodigy, first make sure to install the prodigyopt library: `pip install prodigyopt`, and then specify -
 ```bash
 --optimizer="prodigy"
 ```

examples/flux-control/train_control_flux.py

@@ -915,7 +915,7 @@ def main(args):
         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_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )

examples/flux-control/train_control_lora_flux.py

@@ -1060,7 +1060,7 @@ def main(args):
         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_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )

examples/research_projects/multi_subject_dreambooth_inpainting/README.md

@@ -73,7 +73,7 @@ accelerate launch train_multi_subject_dreambooth_inpaint.py \
 
 ## 3. Results
 
-A [![Weights & Biases](https://img.shields.io/badge/Weights%20&%20Biases-Report-blue)](https://wandb.ai/gzguevara/uncategorized/reports/Multi-Subject-Dreambooth-for-Inpainting--Vmlldzo2MzY5NDQ4?accessToken=y0nya2d7baguhbryxaikbfr1203amvn1jsmyl07vk122mrs7tnph037u1nqgse8t) is provided showing the training progress by every 50 steps. Note, the reported  weights & baises run was performed on a A100 GPU with the following stetting:
+A [![Weights & Biases](https://img.shields.io/badge/Weights%20&%20Biases-Report-blue)](https://wandb.ai/gzguevara/uncategorized/reports/Multi-Subject-Dreambooth-for-Inpainting--Vmlldzo2MzY5NDQ4?accessToken=y0nya2d7baguhbryxaikbfr1203amvn1jsmyl07vk122mrs7tnph037u1nqgse8t) is provided showing the training progress by every 50 steps. Note, the reported weights & biases run was performed on a A100 GPU with the following stetting:
 
 ```bash
 accelerate launch train_multi_subject_dreambooth_inpaint.py \

examples/textual_inversion/textual_inversion_sdxl.py

@@ -793,17 +793,22 @@ def main():
     )
 
     # 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,
     )
 
@@ -829,8 +834,14 @@ def main():
 
     # 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)
 

scripts/convert_wan_to_diffusers.py

@@ -1,6 +1,6 @@
 import argparse
 import pathlib
-from typing import Any, Dict
+from typing import Any, Dict, Tuple
 
 import torch
 from accelerate import init_empty_weights
@@ -14,6 +14,8 @@ from diffusers import (
     WanImageToVideoPipeline,
     WanPipeline,
     WanTransformer3DModel,
+    WanVACEPipeline,
+    WanVACETransformer3DModel,
 )
 
 
@@ -59,7 +61,52 @@ TRANSFORMER_KEYS_RENAME_DICT = {
     "attn2.norm_k_img": "attn2.norm_added_k",
 }
 
+VACE_TRANSFORMER_KEYS_RENAME_DICT = {
+    "time_embedding.0": "condition_embedder.time_embedder.linear_1",
+    "time_embedding.2": "condition_embedder.time_embedder.linear_2",
+    "text_embedding.0": "condition_embedder.text_embedder.linear_1",
+    "text_embedding.2": "condition_embedder.text_embedder.linear_2",
+    "time_projection.1": "condition_embedder.time_proj",
+    "head.modulation": "scale_shift_table",
+    "head.head": "proj_out",
+    "modulation": "scale_shift_table",
+    "ffn.0": "ffn.net.0.proj",
+    "ffn.2": "ffn.net.2",
+    # Hack to swap the layer names
+    # The original model calls the norms in following order: norm1, norm3, norm2
+    # We convert it to: norm1, norm2, norm3
+    "norm2": "norm__placeholder",
+    "norm3": "norm2",
+    "norm__placeholder": "norm3",
+    # # For the I2V model
+    # "img_emb.proj.0": "condition_embedder.image_embedder.norm1",
+    # "img_emb.proj.1": "condition_embedder.image_embedder.ff.net.0.proj",
+    # "img_emb.proj.3": "condition_embedder.image_embedder.ff.net.2",
+    # "img_emb.proj.4": "condition_embedder.image_embedder.norm2",
+    # # for the FLF2V model
+    # "img_emb.emb_pos": "condition_embedder.image_embedder.pos_embed",
+    # Add attention component mappings
+    "self_attn.q": "attn1.to_q",
+    "self_attn.k": "attn1.to_k",
+    "self_attn.v": "attn1.to_v",
+    "self_attn.o": "attn1.to_out.0",
+    "self_attn.norm_q": "attn1.norm_q",
+    "self_attn.norm_k": "attn1.norm_k",
+    "cross_attn.q": "attn2.to_q",
+    "cross_attn.k": "attn2.to_k",
+    "cross_attn.v": "attn2.to_v",
+    "cross_attn.o": "attn2.to_out.0",
+    "cross_attn.norm_q": "attn2.norm_q",
+    "cross_attn.norm_k": "attn2.norm_k",
+    "attn2.to_k_img": "attn2.add_k_proj",
+    "attn2.to_v_img": "attn2.add_v_proj",
+    "attn2.norm_k_img": "attn2.norm_added_k",
+    "before_proj": "proj_in",
+    "after_proj": "proj_out",
+}
+
 TRANSFORMER_SPECIAL_KEYS_REMAP = {}
+VACE_TRANSFORMER_SPECIAL_KEYS_REMAP = {}
 
 
 def update_state_dict_(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
@@ -74,7 +121,7 @@ def load_sharded_safetensors(dir: pathlib.Path):
     return state_dict
 
 
-def get_transformer_config(model_type: str) -> Dict[str, Any]:
+def get_transformer_config(model_type: str) -> Tuple[Dict[str, Any], ...]:
     if model_type == "Wan-T2V-1.3B":
         config = {
             "model_id": "StevenZhang/Wan2.1-T2V-1.3B-Diff",
@@ -94,6 +141,8 @@ def get_transformer_config(model_type: str) -> Dict[str, Any]:
                 "text_dim": 4096,
             },
         }
+        RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT
+        SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP
     elif model_type == "Wan-T2V-14B":
         config = {
             "model_id": "StevenZhang/Wan2.1-T2V-14B-Diff",
@@ -113,6 +162,8 @@ def get_transformer_config(model_type: str) -> Dict[str, Any]:
                 "text_dim": 4096,
             },
         }
+        RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT
+        SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP
     elif model_type == "Wan-I2V-14B-480p":
         config = {
             "model_id": "StevenZhang/Wan2.1-I2V-14B-480P-Diff",
@@ -133,6 +184,8 @@ def get_transformer_config(model_type: str) -> Dict[str, Any]:
                 "text_dim": 4096,
             },
         }
+        RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT
+        SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP
     elif model_type == "Wan-I2V-14B-720p":
         config = {
             "model_id": "StevenZhang/Wan2.1-I2V-14B-720P-Diff",
@@ -153,6 +206,8 @@ def get_transformer_config(model_type: str) -> Dict[str, Any]:
                 "text_dim": 4096,
             },
         }
+        RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT
+        SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP
     elif model_type == "Wan-FLF2V-14B-720P":
         config = {
             "model_id": "ypyp/Wan2.1-FLF2V-14B-720P",  # This is just a placeholder
@@ -175,11 +230,60 @@ def get_transformer_config(model_type: str) -> Dict[str, Any]:
                 "pos_embed_seq_len": 257 * 2,
             },
         }
-    return config
+        RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT
+        SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP
+    elif model_type == "Wan-VACE-1.3B":
+        config = {
+            "model_id": "Wan-AI/Wan2.1-VACE-1.3B",
+            "diffusers_config": {
+                "added_kv_proj_dim": None,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "in_channels": 16,
+                "num_attention_heads": 12,
+                "num_layers": 30,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "vace_layers": [0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28],
+                "vace_in_channels": 96,
+            },
+        }
+        RENAME_DICT = VACE_TRANSFORMER_KEYS_RENAME_DICT
+        SPECIAL_KEYS_REMAP = VACE_TRANSFORMER_SPECIAL_KEYS_REMAP
+    elif model_type == "Wan-VACE-14B":
+        config = {
+            "model_id": "Wan-AI/Wan2.1-VACE-14B",
+            "diffusers_config": {
+                "added_kv_proj_dim": None,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 16,
+                "num_attention_heads": 40,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "vace_layers": [0, 5, 10, 15, 20, 25, 30, 35],
+                "vace_in_channels": 96,
+            },
+        }
+        RENAME_DICT = VACE_TRANSFORMER_KEYS_RENAME_DICT
+        SPECIAL_KEYS_REMAP = VACE_TRANSFORMER_SPECIAL_KEYS_REMAP
+    return config, RENAME_DICT, SPECIAL_KEYS_REMAP
 
 
 def convert_transformer(model_type: str):
-    config = get_transformer_config(model_type)
+    config, RENAME_DICT, SPECIAL_KEYS_REMAP = get_transformer_config(model_type)
+
     diffusers_config = config["diffusers_config"]
     model_id = config["model_id"]
     model_dir = pathlib.Path(snapshot_download(model_id, repo_type="model"))
@@ -187,16 +291,19 @@ def convert_transformer(model_type: str):
     original_state_dict = load_sharded_safetensors(model_dir)
 
     with init_empty_weights():
-        transformer = WanTransformer3DModel.from_config(diffusers_config)
+        if "VACE" not in model_type:
+            transformer = WanTransformer3DModel.from_config(diffusers_config)
+        else:
+            transformer = WanVACETransformer3DModel.from_config(diffusers_config)
 
     for key in list(original_state_dict.keys()):
         new_key = key[:]
-        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+        for replace_key, rename_key in RENAME_DICT.items():
             new_key = new_key.replace(replace_key, rename_key)
         update_state_dict_(original_state_dict, key, new_key)
 
     for key in list(original_state_dict.keys()):
-        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+        for special_key, handler_fn_inplace in SPECIAL_KEYS_REMAP.items():
             if special_key not in key:
                 continue
             handler_fn_inplace(key, original_state_dict)
@@ -412,7 +519,7 @@ def get_args():
     parser = argparse.ArgumentParser()
     parser.add_argument("--model_type", type=str, default=None)
     parser.add_argument("--output_path", type=str, required=True)
-    parser.add_argument("--dtype", default="fp32")
+    parser.add_argument("--dtype", default="fp32", choices=["fp32", "fp16", "bf16", "none"])
     return parser.parse_args()
 
 
@@ -426,18 +533,20 @@ DTYPE_MAPPING = {
 if __name__ == "__main__":
     args = get_args()
 
-    transformer = None
-    dtype = DTYPE_MAPPING[args.dtype]
-
-    transformer = convert_transformer(args.model_type).to(dtype=dtype)
+    transformer = convert_transformer(args.model_type)
     vae = convert_vae()
-    text_encoder = UMT5EncoderModel.from_pretrained("google/umt5-xxl")
+    text_encoder = UMT5EncoderModel.from_pretrained("google/umt5-xxl", torch_dtype=torch.bfloat16)
     tokenizer = AutoTokenizer.from_pretrained("google/umt5-xxl")
     flow_shift = 16.0 if "FLF2V" in args.model_type else 3.0
     scheduler = UniPCMultistepScheduler(
         prediction_type="flow_prediction", use_flow_sigmas=True, num_train_timesteps=1000, flow_shift=flow_shift
     )
 
+    # If user has specified "none", we keep the original dtypes of the state dict without any conversion
+    if args.dtype != "none":
+        dtype = DTYPE_MAPPING[args.dtype]
+        transformer.to(dtype)
+
     if "I2V" in args.model_type or "FLF2V" in args.model_type:
         image_encoder = CLIPVisionModelWithProjection.from_pretrained(
             "laion/CLIP-ViT-H-14-laion2B-s32B-b79K", torch_dtype=torch.bfloat16
@@ -452,6 +561,14 @@ if __name__ == "__main__":
             image_encoder=image_encoder,
             image_processor=image_processor,
         )
+    elif "VACE" in args.model_type:
+        pipe = WanVACEPipeline(
+            transformer=transformer,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            vae=vae,
+            scheduler=scheduler,
+        )
     else:
         pipe = WanPipeline(
             transformer=transformer,

src/diffusers/__init__.py

@@ -215,6 +215,7 @@ else:
             "UVit2DModel",
             "VQModel",
             "WanTransformer3DModel",
+            "WanVACETransformer3DModel",
         ]
     )
     _import_structure["optimization"] = [
@@ -441,6 +442,7 @@ else:
             "SanaControlNetPipeline",
             "SanaPAGPipeline",
             "SanaPipeline",
+            "SanaSprintImg2ImgPipeline",
             "SanaSprintPipeline",
             "SemanticStableDiffusionPipeline",
             "ShapEImg2ImgPipeline",
@@ -526,6 +528,7 @@ else:
             "VQDiffusionPipeline",
             "WanImageToVideoPipeline",
             "WanPipeline",
+            "WanVACEPipeline",
             "WanVideoToVideoPipeline",
             "WuerstchenCombinedPipeline",
             "WuerstchenDecoderPipeline",
@@ -691,6 +694,7 @@ else:
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     from .configuration_utils import ConfigMixin
+    from .quantizers import PipelineQuantizationConfig
 
     try:
         if not is_bitsandbytes_available():
@@ -819,6 +823,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             UVit2DModel,
             VQModel,
             WanTransformer3DModel,
+            WanVACETransformer3DModel,
         )
         from .optimization import (
             get_constant_schedule,
@@ -1025,6 +1030,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             SanaControlNetPipeline,
             SanaPAGPipeline,
             SanaPipeline,
+            SanaSprintImg2ImgPipeline,
             SanaSprintPipeline,
             SemanticStableDiffusionPipeline,
             ShapEImg2ImgPipeline,
@@ -1109,6 +1115,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VQDiffusionPipeline,
             WanImageToVideoPipeline,
             WanPipeline,
+            WanVACEPipeline,
             WanVideoToVideoPipeline,
             WuerstchenCombinedPipeline,
             WuerstchenDecoderPipeline,

src/diffusers/hooks/faster_cache.py

@@ -146,7 +146,7 @@ class FasterCacheConfig:
     alpha_low_frequency: float = 1.1
     alpha_high_frequency: float = 1.1
 
-    # n as described in CFG-Cache explanation in the paper - dependant on the model
+    # n as described in CFG-Cache explanation in the paper - dependent on the model
     unconditional_batch_skip_range: int = 5
     unconditional_batch_timestep_skip_range: Tuple[int, int] = (-1, 641)
 

src/diffusers/hooks/group_offloading.py

@@ -113,6 +113,7 @@ class ModuleGroup:
         finally:
             pinned_dict = None
 
+    @torch.compiler.disable()
     def onload_(self):
         r"""Onloads the group of modules to the onload_device."""
         torch_accelerator_module = (
@@ -165,6 +166,7 @@ class ModuleGroup:
                     if self.record_stream:
                         buffer.data.record_stream(current_stream)
 
+    @torch.compiler.disable()
     def offload_(self):
         r"""Offloads the group of modules to the offload_device."""
 

src/diffusers/hooks/hooks.py

@@ -45,7 +45,7 @@ class ModelHook:
 
     def deinitalize_hook(self, module: torch.nn.Module) -> torch.nn.Module:
         r"""
-        Hook that is executed when a model is deinitalized.
+        Hook that is executed when a model is deinitialized.
 
         Args:
             module (`torch.nn.Module`):

src/diffusers/hooks/layerwise_casting.py

@@ -62,7 +62,7 @@ class LayerwiseCastingHook(ModelHook):
 
     def deinitalize_hook(self, module: torch.nn.Module):
         raise NotImplementedError(
-            "LayerwiseCastingHook does not support deinitalization. A model once enabled with layerwise casting will "
+            "LayerwiseCastingHook does not support deinitialization. A model once enabled with layerwise casting will "
             "have casted its weights to a lower precision dtype for storage. Casting this back to the original dtype "
             "will lead to precision loss, which might have an impact on the model's generation quality. The model should "
             "be re-initialized and loaded in the original dtype."

src/diffusers/loaders/lora_base.py

@@ -465,7 +465,7 @@ class LoraBaseMixin:
     """Utility class for handling LoRAs."""
 
     _lora_loadable_modules = []
-    num_fused_loras = 0
+    _merged_adapters = set()
 
     def load_lora_weights(self, **kwargs):
         raise NotImplementedError("`load_lora_weights()` is not implemented.")
@@ -592,6 +592,9 @@ class LoraBaseMixin:
         if len(components) == 0:
             raise ValueError("`components` cannot be an empty list.")
 
+        # Need to retrieve the names as `adapter_names` can be None. So we cannot directly use it
+        # in `self._merged_adapters = self._merged_adapters | merged_adapter_names`.
+        merged_adapter_names = set()
         for fuse_component in components:
             if fuse_component not in self._lora_loadable_modules:
                 raise ValueError(f"{fuse_component} is not found in {self._lora_loadable_modules=}.")
@@ -601,13 +604,19 @@ class LoraBaseMixin:
                 # check if diffusers model
                 if issubclass(model.__class__, ModelMixin):
                     model.fuse_lora(lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names)
+                    for module in model.modules():
+                        if isinstance(module, BaseTunerLayer):
+                            merged_adapter_names.update(set(module.merged_adapters))
                 # handle transformers models.
                 if issubclass(model.__class__, PreTrainedModel):
                     fuse_text_encoder_lora(
                         model, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
                     )
+                    for module in model.modules():
+                        if isinstance(module, BaseTunerLayer):
+                            merged_adapter_names.update(set(module.merged_adapters))
 
-        self.num_fused_loras += 1
+        self._merged_adapters = self._merged_adapters | merged_adapter_names
 
     def unfuse_lora(self, components: List[str] = [], **kwargs):
         r"""
@@ -661,9 +670,18 @@ class LoraBaseMixin:
                 if issubclass(model.__class__, (ModelMixin, PreTrainedModel)):
                     for module in model.modules():
                         if isinstance(module, BaseTunerLayer):
+                            for adapter in set(module.merged_adapters):
+                                if adapter and adapter in self._merged_adapters:
+                                    self._merged_adapters = self._merged_adapters - {adapter}
                             module.unmerge()
 
-        self.num_fused_loras -= 1
+    @property
+    def num_fused_loras(self):
+        return len(self._merged_adapters)
+
+    @property
+    def fused_loras(self):
+        return self._merged_adapters
 
     def set_adapters(
         self,

src/diffusers/loaders/peft.py

@@ -58,6 +58,7 @@ _SET_ADAPTER_SCALE_FN_MAPPING = {
     "CogView4Transformer2DModel": lambda model_cls, weights: weights,
     "HiDreamImageTransformer2DModel": lambda model_cls, weights: weights,
     "HunyuanVideoFramepackTransformer3DModel": lambda model_cls, weights: weights,
+    "WanVACETransformer3DModel": lambda model_cls, weights: weights,
 }
 
 
@@ -251,7 +252,7 @@ class PeftAdapterMixin:
 
             rank = {}
             for key, val in state_dict.items():
-                # Cannot figure out rank from lora layers that don't have atleast 2 dimensions.
+                # Cannot figure out rank from lora layers that don't have at least 2 dimensions.
                 # Bias layers in LoRA only have a single dimension
                 if "lora_B" in key and val.ndim > 1:
                     # Check out https://github.com/huggingface/peft/pull/2419 for the `^` symbol.

src/diffusers/models/__init__.py

@@ -89,6 +89,7 @@ if is_torch_available():
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
     _import_structure["transformers.transformer_wan"] = ["WanTransformer3DModel"]
+    _import_structure["transformers.transformer_wan_vace"] = ["WanVACETransformer3DModel"]
     _import_structure["unets.unet_1d"] = ["UNet1DModel"]
     _import_structure["unets.unet_2d"] = ["UNet2DModel"]
     _import_structure["unets.unet_2d_condition"] = ["UNet2DConditionModel"]
@@ -178,6 +179,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             Transformer2DModel,
             TransformerTemporalModel,
             WanTransformer3DModel,
+            WanVACETransformer3DModel,
         )
         from .unets import (
             I2VGenXLUNet,

src/diffusers/models/attention.py

@@ -11,36 +11,814 @@
 # 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.
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Callable, Dict, Optional, Tuple, Union
 
 import torch
 import torch.nn.functional as F
 from torch import nn
 
-from ..utils import deprecate, logging
+
+# Import xformers only if it's available
+try:
+    import xformers
+    import xformers.ops
+except ImportError:
+    xformers = None
+
+from ..utils import logging
+from ..utils.import_utils import (
+    is_torch_npu_available,
+    is_torch_xla_available,
+    is_xformers_available,
+)
 from ..utils.torch_utils import maybe_allow_in_graph
-from .activations import GEGLU, GELU, ApproximateGELU, FP32SiLU, LinearActivation, SwiGLU
-from .attention_processor import Attention, JointAttnProcessor2_0
-from .embeddings import SinusoidalPositionalEmbedding
-from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm, SD35AdaLayerNormZeroX
+from .attention_processor import (
+    AttentionProcessor,
+    AttnProcessor,
+)
+from .normalization import RMSNorm
 
 
 logger = logging.get_logger(__name__)
 
 
-def _chunked_feed_forward(ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int):
-    # "feed_forward_chunk_size" can be used to save memory
-    if hidden_states.shape[chunk_dim] % chunk_size != 0:
-        raise ValueError(
-            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
-        )
+class AttentionMixin:
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
 
-    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
-    ff_output = torch.cat(
-        [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
-        dim=chunk_dim,
-    )
-    return ff_output
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        """
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        for module in self.modules():
+            if isinstance(module, AttentionModuleMixin):
+                module.fuse_projections(fuse=True)
+
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        for _, attn_processor in self.attn_processors.items():
+            attn_processor.fused_projections = False
+
+
+class AttentionModuleMixin:
+    _default_processor_cls = None
+    _available_processors = []
+    fused_projections = False
+
+    def set_processor(self, processor: "AttnProcessor") -> None:
+        """
+        Set the attention processor to use.
+
+        Args:
+            processor (`AttnProcessor`):
+                The attention processor to use.
+        """
+        # if current processor is in `self._modules` and if passed `processor` is not, we need to
+        # pop `processor` from `self._modules`
+        if (
+            hasattr(self, "processor")
+            and isinstance(self.processor, torch.nn.Module)
+            and not isinstance(processor, torch.nn.Module)
+        ):
+            logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}")
+            self._modules.pop("processor")
+
+        self.processor = processor
+
+    def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor":
+        """
+        Get the attention processor in use.
+
+        Args:
+            return_deprecated_lora (`bool`, *optional*, defaults to `False`):
+                Set to `True` to return the deprecated LoRA attention processor.
+
+        Returns:
+            "AttentionProcessor": The attention processor in use.
+        """
+        if not return_deprecated_lora:
+            return self.processor
+
+    def set_attention_backend(self, backend: str):
+        from .attention_dispatch import AttentionBackendName
+
+        available_backends = {x.value for x in AttentionBackendName.__members__.values()}
+        if backend not in available_backends:
+            raise ValueError(f"`{backend=}` must be one of the following: " + ", ".join(available_backends))
+
+        backend = AttentionBackendName(backend.lower())
+        self.processor._attention_backend = backend
+
+    def set_use_npu_flash_attention(self, use_npu_flash_attention: bool) -> None:
+        """
+        Set whether to use NPU flash attention from `torch_npu` or not.
+
+        Args:
+            use_npu_flash_attention (`bool`): Whether to use NPU flash attention or not.
+        """
+
+        if use_npu_flash_attention:
+            if not is_torch_npu_available():
+                raise ImportError("torch_npu is not available")
+
+        self.set_attention_backend("_native_npu")
+
+    def set_use_xla_flash_attention(
+        self,
+        use_xla_flash_attention: bool,
+        partition_spec: Optional[Tuple[Optional[str], ...]] = None,
+        is_flux=False,
+    ) -> None:
+        """
+        Set whether to use XLA flash attention from `torch_xla` or not.
+
+        Args:
+            use_xla_flash_attention (`bool`):
+                Whether to use pallas flash attention kernel from `torch_xla` or not.
+            partition_spec (`Tuple[]`, *optional*):
+                Specify the partition specification if using SPMD. Otherwise None.
+            is_flux (`bool`, *optional*, defaults to `False`):
+                Whether the model is a Flux model.
+        """
+        if use_xla_flash_attention:
+            if not is_torch_xla_available():
+                raise ImportError("torch_xla is not available")
+
+        self.set_attention_backend("_native_xla")
+
+    def set_use_memory_efficient_attention_xformers(
+        self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None
+    ) -> None:
+        """
+        Set whether to use memory efficient attention from `xformers` or not.
+
+        Args:
+            use_memory_efficient_attention_xformers (`bool`):
+                Whether to use memory efficient attention from `xformers` or not.
+            attention_op (`Callable`, *optional*):
+                The attention operation to use. Defaults to `None` which uses the default attention operation from
+                `xformers`.
+        """
+        if use_memory_efficient_attention_xformers:
+            if not is_xformers_available():
+                raise ModuleNotFoundError(
+                    "Refer to https://github.com/facebookresearch/xformers for more information on how to install xformers",
+                    name="xformers",
+                )
+            elif not torch.cuda.is_available():
+                raise ValueError(
+                    "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
+                    " only available for GPU "
+                )
+            else:
+                try:
+                    # Make sure we can run the memory efficient attention
+                    if xformers is not None:
+                        dtype = None
+                        if attention_op is not None:
+                            op_fw, op_bw = attention_op
+                            dtype, *_ = op_fw.SUPPORTED_DTYPES
+                        q = torch.randn((1, 2, 40), device="cuda", dtype=dtype)
+                        _ = xformers.ops.memory_efficient_attention(q, q, q)
+                except Exception as e:
+                    raise e
+
+                self.set_attention_backend("xformers")
+
+    @torch.no_grad()
+    def fuse_projections(self):
+        """
+        Fuse the query, key, and value projections into a single projection for efficiency.
+        """
+        # Skip if already fused
+        if getattr(self, "fused_projections", False):
+            return
+
+        device = self.to_q.weight.data.device
+        dtype = self.to_q.weight.data.dtype
+
+        if hasattr(self, "is_cross_attention") and self.is_cross_attention:
+            # Fuse cross-attention key-value projections
+            concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            self.to_kv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
+            self.to_kv.weight.copy_(concatenated_weights)
+            if hasattr(self, "use_bias") and self.use_bias:
+                concatenated_bias = torch.cat([self.to_k.bias.data, self.to_v.bias.data])
+                self.to_kv.bias.copy_(concatenated_bias)
+        else:
+            # Fuse self-attention projections
+            concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            self.to_qkv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
+            self.to_qkv.weight.copy_(concatenated_weights)
+            if hasattr(self, "use_bias") and self.use_bias:
+                concatenated_bias = torch.cat([self.to_q.bias.data, self.to_k.bias.data, self.to_v.bias.data])
+                self.to_qkv.bias.copy_(concatenated_bias)
+
+        # Handle added projections for models like SD3, Flux, etc.
+        if (
+            getattr(self, "add_q_proj", None) is not None
+            and getattr(self, "add_k_proj", None) is not None
+            and getattr(self, "add_v_proj", None) is not None
+        ):
+            concatenated_weights = torch.cat(
+                [self.add_q_proj.weight.data, self.add_k_proj.weight.data, self.add_v_proj.weight.data]
+            )
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            self.to_added_qkv = nn.Linear(
+                in_features, out_features, bias=self.added_proj_bias, device=device, dtype=dtype
+            )
+            self.to_added_qkv.weight.copy_(concatenated_weights)
+            if self.added_proj_bias:
+                concatenated_bias = torch.cat(
+                    [self.add_q_proj.bias.data, self.add_k_proj.bias.data, self.add_v_proj.bias.data]
+                )
+                self.to_added_qkv.bias.copy_(concatenated_bias)
+
+        self.fused_projections = True
+
+    @torch.no_grad()
+    def unfuse_projections(self):
+        """
+        Unfuse the query, key, and value projections back to separate projections.
+        """
+        # Skip if not fused
+        if not getattr(self, "fused_projections", False):
+            return
+
+        # Remove fused projection layers
+        if hasattr(self, "to_qkv"):
+            delattr(self, "to_qkv")
+
+        if hasattr(self, "to_kv"):
+            delattr(self, "to_kv")
+
+        if hasattr(self, "to_added_qkv"):
+            delattr(self, "to_added_qkv")
+
+        self.fused_projections = False
+
+    def set_attention_slice(self, slice_size: int) -> None:
+        """
+        Set the slice size for attention computation.
+
+        Args:
+            slice_size (`int`):
+                The slice size for attention computation.
+        """
+        if hasattr(self, "sliceable_head_dim") and slice_size is not None and slice_size > self.sliceable_head_dim:
+            raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")
+
+        processor = None
+
+        # Try to get a compatible processor for sliced attention
+        if slice_size is not None:
+            processor = self._get_compatible_processor("sliced")
+
+        # If no processor was found or slice_size is None, use default processor
+        if processor is None:
+            processor = self.default_processor_cls()
+
+        self.set_processor(processor)
+
+    def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor:
+        """
+        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`.
+
+        Args:
+            tensor (`torch.Tensor`): The tensor to reshape.
+
+        Returns:
+            `torch.Tensor`: The reshaped tensor.
+        """
+        head_size = self.heads
+        batch_size, seq_len, dim = tensor.shape
+        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
+        return tensor
+
+    def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor:
+        """
+        Reshape the tensor for multi-head attention processing.
+
+        Args:
+            tensor (`torch.Tensor`): The tensor to reshape.
+            out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor.
+
+        Returns:
+            `torch.Tensor`: The reshaped tensor.
+        """
+        head_size = self.heads
+        if tensor.ndim == 3:
+            batch_size, seq_len, dim = tensor.shape
+            extra_dim = 1
+        else:
+            batch_size, extra_dim, seq_len, dim = tensor.shape
+        tensor = tensor.reshape(batch_size, seq_len * extra_dim, head_size, dim // head_size)
+        tensor = tensor.permute(0, 2, 1, 3)
+
+        if out_dim == 3:
+            tensor = tensor.reshape(batch_size * head_size, seq_len * extra_dim, dim // head_size)
+
+        return tensor
+
+    def get_attention_scores(
+        self, query: torch.Tensor, key: torch.Tensor, attention_mask: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        """
+        Compute the attention scores.
+
+        Args:
+            query (`torch.Tensor`): The query tensor.
+            key (`torch.Tensor`): The key tensor.
+            attention_mask (`torch.Tensor`, *optional*): The attention mask to use.
+
+        Returns:
+            `torch.Tensor`: The attention probabilities/scores.
+        """
+        dtype = query.dtype
+        if self.upcast_attention:
+            query = query.float()
+            key = key.float()
+
+        if attention_mask is None:
+            baddbmm_input = torch.empty(
+                query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device
+            )
+            beta = 0
+        else:
+            baddbmm_input = attention_mask
+            beta = 1
+
+        attention_scores = torch.baddbmm(
+            baddbmm_input,
+            query,
+            key.transpose(-1, -2),
+            beta=beta,
+            alpha=self.scale,
+        )
+        del baddbmm_input
+
+        if self.upcast_softmax:
+            attention_scores = attention_scores.float()
+
+        attention_probs = attention_scores.softmax(dim=-1)
+        del attention_scores
+
+        attention_probs = attention_probs.to(dtype)
+
+        return attention_probs
+
+    def prepare_attention_mask(
+        self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3
+    ) -> torch.Tensor:
+        """
+        Prepare the attention mask for the attention computation.
+
+        Args:
+            attention_mask (`torch.Tensor`): The attention mask to prepare.
+            target_length (`int`): The target length of the attention mask.
+            batch_size (`int`): The batch size for repeating the attention mask.
+            out_dim (`int`, *optional*, defaults to `3`): Output dimension.
+
+        Returns:
+            `torch.Tensor`: The prepared attention mask.
+        """
+        head_size = self.heads
+        if attention_mask is None:
+            return attention_mask
+
+        current_length: int = attention_mask.shape[-1]
+        if current_length != target_length:
+            if attention_mask.device.type == "mps":
+                # HACK: MPS: Does not support padding by greater than dimension of input tensor.
+                # Instead, we can manually construct the padding tensor.
+                padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length)
+                padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device)
+                attention_mask = torch.cat([attention_mask, padding], dim=2)
+            else:
+                # TODO: for pipelines such as stable-diffusion, padding cross-attn mask:
+                #       we want to instead pad by (0, remaining_length), where remaining_length is:
+                #       remaining_length: int = target_length - current_length
+                # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding
+                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
+
+        if out_dim == 3:
+            if attention_mask.shape[0] < batch_size * head_size:
+                attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
+        elif out_dim == 4:
+            attention_mask = attention_mask.unsqueeze(1)
+            attention_mask = attention_mask.repeat_interleave(head_size, dim=1)
+
+        return attention_mask
+
+    def norm_encoder_hidden_states(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
+        """
+        Normalize the encoder hidden states.
+
+        Args:
+            encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder.
+
+        Returns:
+            `torch.Tensor`: The normalized encoder hidden states.
+        """
+        assert self.norm_cross is not None, "self.norm_cross must be defined to call self.norm_encoder_hidden_states"
+        if isinstance(self.norm_cross, nn.LayerNorm):
+            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+        elif isinstance(self.norm_cross, nn.GroupNorm):
+            # Group norm norms along the channels dimension and expects
+            # input to be in the shape of (N, C, *). In this case, we want
+            # to norm along the hidden dimension, so we need to move
+            # (batch_size, sequence_length, hidden_size) ->
+            # (batch_size, hidden_size, sequence_length)
+            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+        else:
+            assert False
+
+        return encoder_hidden_states
+
+
+@maybe_allow_in_graph
+class Attention(nn.Module, AttentionModuleMixin):
+    r"""
+    A cross attention layer.
+
+    Parameters:
+        query_dim (`int`):
+            The number of channels in the query.
+        cross_attention_dim (`int`, *optional*):
+            The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
+        heads (`int`,  *optional*, defaults to 8):
+            The number of heads to use for multi-head attention.
+        kv_heads (`int`,  *optional*, defaults to `None`):
+            The number of key and value heads to use for multi-head attention. Defaults to `heads`. If
+            `kv_heads=heads`, the model will use Multi Head Attention (MHA), if `kv_heads=1` the model will use Multi
+            Query Attention (MQA) otherwise GQA is used.
+        dim_head (`int`,  *optional*, defaults to 64):
+            The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability to use.
+        bias (`bool`, *optional*, defaults to False):
+            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
+        upcast_attention (`bool`, *optional*, defaults to False):
+            Set to `True` to upcast the attention computation to `float32`.
+        upcast_softmax (`bool`, *optional*, defaults to False):
+            Set to `True` to upcast the softmax computation to `float32`.
+        cross_attention_norm (`str`, *optional*, defaults to `None`):
+            The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.
+        cross_attention_norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use for the group norm in the cross attention.
+        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the added key and value projections. If `None`, no projection is used.
+        norm_num_groups (`int`, *optional*, defaults to `None`):
+            The number of groups to use for the group norm in the attention.
+        spatial_norm_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the spatial normalization.
+        out_bias (`bool`, *optional*, defaults to `True`):
+            Set to `True` to use a bias in the output linear layer.
+        scale_qk (`bool`, *optional*, defaults to `True`):
+            Set to `True` to scale the query and key by `1 / sqrt(dim_head)`.
+        only_cross_attention (`bool`, *optional*, defaults to `False`):
+            Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if
+            `added_kv_proj_dim` is not `None`.
+        eps (`float`, *optional*, defaults to 1e-5):
+            An additional value added to the denominator in group normalization that is used for numerical stability.
+        rescale_output_factor (`float`, *optional*, defaults to 1.0):
+            A factor to rescale the output by dividing it with this value.
+        residual_connection (`bool`, *optional*, defaults to `False`):
+            Set to `True` to add the residual connection to the output.
+        _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`):
+            Set to `True` if the attention block is loaded from a deprecated state dict.
+        processor (`AttnProcessor`, *optional*, defaults to `None`):
+            The attention processor to use. If `None`, defaults to `AttnProcessorSDPA` if `torch 2.x` is used and
+            `AttnProcessor` otherwise.
+    """
+
+    def __init__(
+        self,
+        query_dim: int,
+        cross_attention_dim: Optional[int] = None,
+        heads: int = 8,
+        kv_heads: Optional[int] = None,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias: bool = False,
+        upcast_attention: bool = False,
+        upcast_softmax: bool = False,
+        cross_attention_norm: Optional[str] = None,
+        cross_attention_norm_num_groups: int = 32,
+        qk_norm: Optional[str] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        added_proj_bias: Optional[bool] = True,
+        norm_num_groups: Optional[int] = None,
+        spatial_norm_dim: Optional[int] = None,
+        out_bias: bool = True,
+        scale_qk: bool = True,
+        only_cross_attention: bool = False,
+        eps: float = 1e-5,
+        rescale_output_factor: float = 1.0,
+        residual_connection: bool = False,
+        _from_deprecated_attn_block: bool = False,
+        processor: Optional["AttnProcessor"] = None,
+        out_dim: int = None,
+        out_context_dim: int = None,
+        context_pre_only=None,
+        pre_only=False,
+        elementwise_affine: bool = True,
+        is_causal: bool = False,
+    ):
+        super().__init__()
+
+        # To prevent circular import.
+        from .normalization import FP32LayerNorm, LpNorm
+
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads
+        self.query_dim = query_dim
+        self.use_bias = bias
+        self.is_cross_attention = cross_attention_dim is not None
+        self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+        self.rescale_output_factor = rescale_output_factor
+        self.residual_connection = residual_connection
+        self.dropout = dropout
+        self.fused_projections = False
+        self.out_dim = out_dim if out_dim is not None else query_dim
+        self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim
+        self.context_pre_only = context_pre_only
+        self.pre_only = pre_only
+        self.is_causal = is_causal
+
+        # we make use of this private variable to know whether this class is loaded
+        # with an deprecated state dict so that we can convert it on the fly
+        self._from_deprecated_attn_block = _from_deprecated_attn_block
+
+        self.scale_qk = scale_qk
+        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        # for slice_size > 0 the attention score computation
+        # is split across the batch axis to save memory
+        # You can set slice_size with `set_attention_slice`
+        self.sliceable_head_dim = heads
+
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.only_cross_attention = only_cross_attention
+
+        if self.added_kv_proj_dim is None and self.only_cross_attention:
+            raise ValueError(
+                "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
+            )
+
+        if norm_num_groups is not None:
+            self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True)
+        else:
+            self.group_norm = None
+
+        if spatial_norm_dim is not None:
+            self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim)
+        else:
+            self.spatial_norm = None
+
+        if qk_norm is None:
+            self.norm_q = None
+            self.norm_k = None
+        elif qk_norm == "layer_norm":
+            self.norm_q = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+            self.norm_k = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+        elif qk_norm == "fp32_layer_norm":
+            self.norm_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
+            self.norm_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
+        elif qk_norm == "layer_norm_across_heads":
+            # Lumina applies qk norm across all heads
+            self.norm_q = nn.LayerNorm(dim_head * heads, eps=eps)
+            self.norm_k = nn.LayerNorm(dim_head * kv_heads, eps=eps)
+        elif qk_norm == "rms_norm":
+            self.norm_q = RMSNorm(dim_head, eps=eps)
+            self.norm_k = RMSNorm(dim_head, eps=eps)
+        elif qk_norm == "rms_norm_across_heads":
+            # LTX applies qk norm across all heads
+            self.norm_q = RMSNorm(dim_head * heads, eps=eps)
+            self.norm_k = RMSNorm(dim_head * kv_heads, eps=eps)
+        elif qk_norm == "l2":
+            self.norm_q = LpNorm(p=2, dim=-1, eps=eps)
+            self.norm_k = LpNorm(p=2, dim=-1, eps=eps)
+        else:
+            raise ValueError(
+                f"unknown qk_norm: {qk_norm}. Should be one of None, 'layer_norm', 'fp32_layer_norm', 'layer_norm_across_heads', 'rms_norm', 'rms_norm_across_heads', 'l2'."
+            )
+
+        if cross_attention_norm is None:
+            self.norm_cross = None
+        elif cross_attention_norm == "layer_norm":
+            self.norm_cross = nn.LayerNorm(self.cross_attention_dim)
+        elif cross_attention_norm == "group_norm":
+            if self.added_kv_proj_dim is not None:
+                # The given `encoder_hidden_states` are initially of shape
+                # (batch_size, seq_len, added_kv_proj_dim) before being projected
+                # to (batch_size, seq_len, cross_attention_dim). The norm is applied
+                # before the projection, so we need to use `added_kv_proj_dim` as
+                # the number of channels for the group norm.
+                norm_cross_num_channels = added_kv_proj_dim
+            else:
+                norm_cross_num_channels = self.cross_attention_dim
+
+            self.norm_cross = nn.GroupNorm(
+                num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True
+            )
+        else:
+            raise ValueError(
+                f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'"
+            )
+
+        self.to_q = nn.Linear(query_dim, self.inner_dim, bias=bias)
+
+        if not self.only_cross_attention:
+            # only relevant for the `AddedKVProcessor` classes
+            self.to_k = nn.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias)
+            self.to_v = nn.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias)
+        else:
+            self.to_k = None
+            self.to_v = None
+
+        self.added_proj_bias = added_proj_bias
+        if self.added_kv_proj_dim is not None:
+            self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias)
+            self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias)
+            if self.context_pre_only is not None:
+                self.add_q_proj = nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+        else:
+            self.add_q_proj = None
+            self.add_k_proj = None
+            self.add_v_proj = None
+
+        if not self.pre_only:
+            self.to_out = nn.ModuleList([])
+            self.to_out.append(nn.Linear(self.inner_dim, self.out_dim, bias=out_bias))
+            self.to_out.append(nn.Dropout(dropout))
+        else:
+            self.to_out = None
+
+        if self.context_pre_only is not None and not self.context_pre_only:
+            self.to_add_out = nn.Linear(self.inner_dim, self.out_context_dim, bias=out_bias)
+        else:
+            self.to_add_out = None
+
+        if qk_norm is not None and added_kv_proj_dim is not None:
+            if qk_norm == "layer_norm":
+                self.norm_added_q = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+                self.norm_added_k = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+            elif qk_norm == "fp32_layer_norm":
+                self.norm_added_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
+                self.norm_added_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
+            elif qk_norm == "rms_norm":
+                self.norm_added_q = RMSNorm(dim_head, eps=eps)
+                self.norm_added_k = RMSNorm(dim_head, eps=eps)
+            elif qk_norm == "rms_norm_across_heads":
+                # Wan applies qk norm across all heads
+                # Wan also doesn't apply a q norm
+                self.norm_added_q = None
+                self.norm_added_k = RMSNorm(dim_head * kv_heads, eps=eps)
+            else:
+                raise ValueError(
+                    f"unknown qk_norm: {qk_norm}. Should be one of `None,'layer_norm','fp32_layer_norm','rms_norm'`"
+                )
+        else:
+            self.norm_added_q = None
+            self.norm_added_k = None
+
+        # set attention processor
+        # We use the AttnProcessorSDPA by default when torch 2.x is used which uses
+        # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+        # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+        if processor is None:
+            processor = self._default_processor_cls
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **cross_attention_kwargs,
+    ) -> torch.Tensor:
+        r"""
+        The forward method of the `Attention` class.
+
+        Args:
+            hidden_states (`torch.Tensor`):
+                The hidden states of the query.
+            encoder_hidden_states (`torch.Tensor`, *optional*):
+                The hidden states of the encoder.
+            attention_mask (`torch.Tensor`, *optional*):
+                The attention mask to use. If `None`, no mask is applied.
+            **cross_attention_kwargs:
+                Additional keyword arguments to pass along to the cross attention.
+
+        Returns:
+            `torch.Tensor`: The output of the attention layer.
+        """
+        # The `Attention` class can call different attention processors / attention functions
+        # here we simply pass along all tensors to the selected processor class
+        # For standard processors that are defined here, `**cross_attention_kwargs` is empty
+
+        attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
+        quiet_attn_parameters = {"ip_adapter_masks", "ip_hidden_states"}
+        unused_kwargs = [
+            k for k, _ in cross_attention_kwargs.items() if k not in attn_parameters and k not in quiet_attn_parameters
+        ]
+        if len(unused_kwargs) > 0:
+            logger.warning(
+                f"cross_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."
+            )
+        cross_attention_kwargs = {k: w for k, w in cross_attention_kwargs.items() if k in attn_parameters}
+
+        return self.processor(
+            self,
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
 
 
 @maybe_allow_in_graph

src/diffusers/models/autoencoders/autoencoder_kl.py

@@ -63,8 +63,8 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalModelMixin, PeftAdapter
             Synthesis with Latent Diffusion Models](https://huggingface.co/papers/2112.10752) paper.
         force_upcast (`bool`, *optional*, default to `True`):
             If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
-            can be fine-tuned / trained to a lower range without loosing too much precision in which case
-            `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+            can be fine-tuned / trained to a lower range without losing too much precision in which case `force_upcast`
+            can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
         mid_block_add_attention (`bool`, *optional*, default to `True`):
             If enabled, the mid_block of the Encoder and Decoder will have attention blocks. If set to false, the
             mid_block will only have resnet blocks

src/diffusers/models/autoencoders/autoencoder_kl_allegro.py

@@ -715,8 +715,8 @@ class AutoencoderKLAllegro(ModelMixin, ConfigMixin):
             Synthesis with Latent Diffusion Models](https://huggingface.co/papers/2112.10752) paper.
         force_upcast (`bool`, default to `True`):
             If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
-            can be fine-tuned / trained to a lower range without loosing too much precision in which case
-            `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+            can be fine-tuned / trained to a lower range without losing too much precision in which case `force_upcast`
+            can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
     """
 
     _supports_gradient_checkpointing = True

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -983,8 +983,8 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             Synthesis with Latent Diffusion Models](https://huggingface.co/papers/2112.10752) paper.
         force_upcast (`bool`, *optional*, default to `True`):
             If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
-            can be fine-tuned / trained to a lower range without loosing too much precision in which case
-            `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+            can be fine-tuned / trained to a lower range without losing too much precision in which case `force_upcast`
+            can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
     """
 
     _supports_gradient_checkpointing = True

src/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py

@@ -161,8 +161,8 @@ class AutoencoderKLTemporalDecoder(ModelMixin, ConfigMixin):
             Synthesis with Latent Diffusion Models](https://huggingface.co/papers/2112.10752) paper.
         force_upcast (`bool`, *optional*, default to `True`):
             If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
-            can be fine-tuned / trained to a lower range without loosing too much precision in which case
-            `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+            can be fine-tuned / trained to a lower range without losing too much precision in which case `force_upcast`
+            can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
     """
 
     _supports_gradient_checkpointing = True

src/diffusers/models/transformers/__init__.py

@@ -32,3 +32,4 @@ if is_torch_available():
     from .transformer_sd3 import SD3Transformer2DModel
     from .transformer_temporal import TransformerTemporalModel
     from .transformer_wan import WanTransformer3DModel
+    from .transformer_wan_vace import WanVACETransformer3DModel

src/diffusers/models/transformers/transformer_flux.py

@@ -13,25 +13,19 @@
 # limitations under the License.
 
 
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
-from ...utils.import_utils import is_torch_npu_available
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..attention import FeedForward
-from ..attention_processor import (
-    Attention,
-    AttentionProcessor,
-    FluxAttnProcessor2_0,
-    FluxAttnProcessor2_0_NPU,
-    FusedFluxAttnProcessor2_0,
-)
+from ..attention import Attention, AttentionMixin, FeedForward
+from ..attention_dispatch import dispatch_attention_fn
 from ..cache_utils import CacheMixin
 from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
 from ..modeling_outputs import Transformer2DModelOutput
@@ -42,6 +36,270 @@ from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNo
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
+class FluxAttnProcessor:
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(f"{self.__class__.__name__} requires PyTorch 2.0, please upgrade PyTorch to 2.0.")
+
+    def _get_projections(self, attn, hidden_states, encoder_hidden_states=None):
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        encoder_projections = None
+        if encoder_hidden_states is not None and hasattr(attn, "add_q_proj"):
+            encoder_query = attn.add_q_proj(encoder_hidden_states)
+            encoder_key = attn.add_k_proj(encoder_hidden_states)
+            encoder_value = attn.add_v_proj(encoder_hidden_states)
+            encoder_projections = (encoder_query, encoder_key, encoder_value)
+
+        return query, key, value, encoder_projections
+
+    def _get_fused_projections(self, attn, hidden_states, encoder_hidden_states=None):
+        qkv = attn.to_qkv(hidden_states)
+        split_size = qkv.shape[-1] // 3
+        query, key, value = torch.split(qkv, split_size, dim=-1)
+
+        encoder_projections = None
+        if encoder_hidden_states is not None and hasattr(attn, "to_added_qkv"):
+            encoder_qkv = attn.to_added_qkv(encoder_hidden_states)
+            split_size = encoder_qkv.shape[-1] // 3
+            encoder_query, encoder_key, encoder_value = torch.split(encoder_qkv, split_size, dim=-1)
+            encoder_projections = (encoder_query, encoder_key, encoder_value)
+
+        return query, key, value, encoder_projections
+
+    def get_qkv_projections(self, attn, hidden_states, encoder_hidden_states=None):
+        """Public method to get projections based on whether we're using fused mode or not."""
+        if attn.is_fused and hasattr(attn, "to_qkv"):
+            return self._get_fused_projections(attn, hidden_states, encoder_hidden_states)
+
+        return self._get_projections(attn, hidden_states, encoder_hidden_states)
+
+    def __call__(
+        self,
+        attn: "FluxAttention",
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+
+        query, key, value, encoder_projections = self.get_qkv_projections(attn, hidden_states, encoder_hidden_states)
+
+        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)
+
+        if encoder_projections is not None:
+            encoder_query, encoder_key, encoder_value = encoder_projections
+            encoder_query = encoder_query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            encoder_key = encoder_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            encoder_value = encoder_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            if attn.norm_added_q is not None:
+                encoder_query = attn.norm_added_q(encoder_query)
+            if attn.norm_added_k is not None:
+                encoder_key = attn.norm_added_k(encoder_key)
+
+            # Concatenate for joint attention
+            query = torch.cat([encoder_query, query], dim=2)
+            key = torch.cat([encoder_key, key], dim=2)
+            value = torch.cat([encoder_value, value], dim=2)
+
+        if image_rotary_emb is not None:
+            from ..embeddings import apply_rotary_emb
+
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = (
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+                hidden_states[:, encoder_hidden_states.shape[1] :],
+            )
+
+            hidden_states = attn.to_out[0](hidden_states)
+            hidden_states = attn.to_out[1](hidden_states)
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states
+
+
+class FluxIPAdapterAttnProcessor(torch.nn.Module):
+    """Flux Attention processor for IP-Adapter."""
+
+    def __init__(
+        self, hidden_size: int, cross_attention_dim: int, num_tokens=(4,), scale=1.0, device=None, dtype=None
+    ):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                f"{self.__class__.__name__} requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+
+        if not isinstance(num_tokens, (tuple, list)):
+            num_tokens = [num_tokens]
+
+        if not isinstance(scale, list):
+            scale = [scale] * len(num_tokens)
+        if len(scale) != len(num_tokens):
+            raise ValueError("`scale` should be a list of integers with the same length as `num_tokens`.")
+        self.scale = scale
+
+        self.to_k_ip = nn.ModuleList(
+            [
+                nn.Linear(cross_attention_dim, hidden_size, bias=True, device=device, dtype=dtype)
+                for _ in range(len(num_tokens))
+            ]
+        )
+        self.to_v_ip = nn.ModuleList(
+            [
+                nn.Linear(cross_attention_dim, hidden_size, bias=True, device=device, dtype=dtype)
+                for _ in range(len(num_tokens))
+            ]
+        )
+
+    def __call__(
+        self,
+        attn: "FluxAttention",
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+        ip_hidden_states: Optional[List[torch.Tensor]] = None,
+        ip_adapter_masks: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+
+        # `sample` projections.
+        hidden_states_query_proj = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        hidden_states_query_proj = hidden_states_query_proj.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:
+            hidden_states_query_proj = attn.norm_q(hidden_states_query_proj)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
+        if encoder_hidden_states is not None:
+            # `context` projections.
+            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+
+            if attn.norm_added_q is not None:
+                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
+            if attn.norm_added_k is not None:
+                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
+
+            # attention
+            query = torch.cat([encoder_hidden_states_query_proj, hidden_states_query_proj], dim=2)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
+
+        if image_rotary_emb is not None:
+            from .embeddings import apply_rotary_emb
+
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = (
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+                hidden_states[:, encoder_hidden_states.shape[1] :],
+            )
+
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            hidden_states = attn.to_out[1](hidden_states)
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+            # IP-adapter
+            ip_query = hidden_states_query_proj
+            ip_attn_output = torch.zeros_like(hidden_states)
+
+            for current_ip_hidden_states, scale, to_k_ip, to_v_ip in zip(
+                ip_hidden_states, self.scale, self.to_k_ip, self.to_v_ip
+            ):
+                ip_key = to_k_ip(current_ip_hidden_states)
+                ip_value = to_v_ip(current_ip_hidden_states)
+
+                ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+                ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+                # the output of sdp = (batch, num_heads, seq_len, head_dim)
+                # TODO: add support for attn.scale when we move to Torch 2.1
+                current_ip_hidden_states = F.scaled_dot_product_attention(
+                    ip_query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+                )
+                current_ip_hidden_states = current_ip_hidden_states.transpose(1, 2).reshape(
+                    batch_size, -1, attn.heads * head_dim
+                )
+                current_ip_hidden_states = current_ip_hidden_states.to(ip_query.dtype)
+                ip_attn_output += scale * current_ip_hidden_states
+
+            return hidden_states, encoder_hidden_states, ip_attn_output
+        else:
+            return hidden_states
+
+
+@maybe_allow_in_graph
+class FluxAttention(Attention):
+    _default_processor_cls = FluxAttnProcessor
+    _available_processors = [
+        FluxAttnProcessor,
+        FluxIPAdapterAttnProcessor,
+    ]
+
+
 @maybe_allow_in_graph
 class FluxSingleTransformerBlock(nn.Module):
     def __init__(self, dim: int, num_attention_heads: int, attention_head_dim: int, mlp_ratio: float = 4.0):
@@ -53,24 +311,13 @@ class FluxSingleTransformerBlock(nn.Module):
         self.act_mlp = nn.GELU(approximate="tanh")
         self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)
 
-        if is_torch_npu_available():
-            deprecation_message = (
-                "Defaulting to FluxAttnProcessor2_0_NPU for NPU devices will be removed. Attention processors "
-                "should be set explicitly using the `set_attn_processor` method."
-            )
-            deprecate("npu_processor", "0.34.0", deprecation_message)
-            processor = FluxAttnProcessor2_0_NPU()
-        else:
-            processor = FluxAttnProcessor2_0()
-
-        self.attn = Attention(
+        self.attn = FluxAttention(
             query_dim=dim,
-            cross_attention_dim=None,
+            out_dim=dim,
             dim_head=attention_head_dim,
             heads=num_attention_heads,
-            out_dim=dim,
+            dropout=0.0,
             bias=True,
-            processor=processor,
             qk_norm="rms_norm",
             eps=1e-6,
             pre_only=True,
@@ -113,18 +360,16 @@ class FluxTransformerBlock(nn.Module):
         self.norm1 = AdaLayerNormZero(dim)
         self.norm1_context = AdaLayerNormZero(dim)
 
-        self.attn = Attention(
+        self.attn = FluxAttention(
             query_dim=dim,
             cross_attention_dim=None,
-            added_kv_proj_dim=dim,
             dim_head=attention_head_dim,
             heads=num_attention_heads,
-            out_dim=dim,
-            context_pre_only=False,
-            bias=True,
-            processor=FluxAttnProcessor2_0(),
             qk_norm=qk_norm,
             eps=eps,
+            dropout=0.0,
+            bias=True,
+            added_kv_proj_dim=dim,
         )
 
         self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
@@ -191,7 +436,13 @@ class FluxTransformerBlock(nn.Module):
 
 
 class FluxTransformer2DModel(
-    ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, FluxTransformer2DLoadersMixin, CacheMixin
+    ModelMixin,
+    ConfigMixin,
+    PeftAdapterMixin,
+    FromOriginalModelMixin,
+    FluxTransformer2DLoadersMixin,
+    CacheMixin,
+    AttentionMixin,
 ):
     """
     The Transformer model introduced in Flux.
@@ -241,7 +492,7 @@ class FluxTransformer2DModel(
         joint_attention_dim: int = 4096,
         pooled_projection_dim: int = 768,
         guidance_embeds: bool = False,
-        axes_dims_rope: Tuple[int] = (16, 56, 56),
+        axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
     ):
         super().__init__()
         self.out_channels = out_channels or in_channels
@@ -286,105 +537,9 @@ class FluxTransformer2DModel(
 
         self.gradient_checkpointing = False
 
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
+    # Using inherited methods from AttentionMixin
 
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
-            if hasattr(module, "set_processor"):
-                if not isinstance(processor, dict):
-                    module.set_processor(processor)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
-    def fuse_qkv_projections(self):
-        """
-        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
-        are fused. For cross-attention modules, key and value projection matrices are fused.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-        """
-        self.original_attn_processors = None
-
-        for _, attn_processor in self.attn_processors.items():
-            if "Added" in str(attn_processor.__class__.__name__):
-                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
-
-        self.original_attn_processors = self.attn_processors
-
-        for module in self.modules():
-            if isinstance(module, Attention):
-                module.fuse_projections(fuse=True)
-
-        self.set_attn_processor(FusedFluxAttnProcessor2_0())
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
-    def unfuse_qkv_projections(self):
-        """Disables the fused QKV projection if enabled.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-
-        """
-        if self.original_attn_processors is not None:
-            self.set_attn_processor(self.original_attn_processors)
+    # Using inherited methods from AttentionMixin
 
     def forward(
         self,
@@ -447,8 +602,6 @@ class FluxTransformer2DModel(
         timestep = timestep.to(hidden_states.dtype) * 1000
         if guidance is not None:
             guidance = guidance.to(hidden_states.dtype) * 1000
-        else:
-            guidance = None
 
         temb = (
             self.time_text_embed(timestep, pooled_projections)

src/diffusers/models/transformers/transformer_wan.py

@@ -72,7 +72,8 @@ class WanAttnProcessor2_0:
         if rotary_emb is not None:
 
             def apply_rotary_emb(hidden_states: torch.Tensor, freqs: torch.Tensor):
-                x_rotated = torch.view_as_complex(hidden_states.to(torch.float64).unflatten(3, (-1, 2)))
+                dtype = torch.float32 if hidden_states.device.type == "mps" else torch.float64
+                x_rotated = torch.view_as_complex(hidden_states.to(dtype).unflatten(3, (-1, 2)))
                 x_out = torch.view_as_real(x_rotated * freqs).flatten(3, 4)
                 return x_out.type_as(hidden_states)
 
@@ -190,9 +191,10 @@ class WanRotaryPosEmbed(nn.Module):
         t_dim = attention_head_dim - h_dim - w_dim
 
         freqs = []
+        freqs_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64
         for dim in [t_dim, h_dim, w_dim]:
             freq = get_1d_rotary_pos_embed(
-                dim, max_seq_len, theta, use_real=False, repeat_interleave_real=False, freqs_dtype=torch.float64
+                dim, max_seq_len, theta, use_real=False, repeat_interleave_real=False, freqs_dtype=freqs_dtype
             )
             freqs.append(freq)
         self.freqs = torch.cat(freqs, dim=1)

src/diffusers/models/transformers/transformer_wan_vace.py

@@ -0,0 +1,393 @@
+# 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 math
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from ..attention import FeedForward
+from ..attention_processor import Attention
+from ..cache_utils import CacheMixin
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import FP32LayerNorm
+from .transformer_wan import WanAttnProcessor2_0, WanRotaryPosEmbed, WanTimeTextImageEmbedding, WanTransformerBlock
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class WanVACETransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        ffn_dim: int,
+        num_heads: int,
+        qk_norm: str = "rms_norm_across_heads",
+        cross_attn_norm: bool = False,
+        eps: float = 1e-6,
+        added_kv_proj_dim: Optional[int] = None,
+        apply_input_projection: bool = False,
+        apply_output_projection: bool = False,
+    ):
+        super().__init__()
+
+        # 1. Input projection
+        self.proj_in = None
+        if apply_input_projection:
+            self.proj_in = nn.Linear(dim, dim)
+
+        # 2. Self-attention
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            processor=WanAttnProcessor2_0(),
+        )
+
+        # 3. Cross-attention
+        self.attn2 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            added_kv_proj_dim=added_kv_proj_dim,
+            added_proj_bias=True,
+            processor=WanAttnProcessor2_0(),
+        )
+        self.norm2 = FP32LayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()
+
+        # 4. Feed-forward
+        self.ffn = FeedForward(dim, inner_dim=ffn_dim, activation_fn="gelu-approximate")
+        self.norm3 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+
+        # 5. Output projection
+        self.proj_out = None
+        if apply_output_projection:
+            self.proj_out = nn.Linear(dim, dim)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        control_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        rotary_emb: torch.Tensor,
+    ) -> torch.Tensor:
+        if self.proj_in is not None:
+            control_hidden_states = self.proj_in(control_hidden_states)
+            control_hidden_states = control_hidden_states + hidden_states
+
+        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(control_hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(
+            control_hidden_states
+        )
+        attn_output = self.attn1(hidden_states=norm_hidden_states, rotary_emb=rotary_emb)
+        control_hidden_states = (control_hidden_states.float() + attn_output * gate_msa).type_as(control_hidden_states)
+
+        # 2. Cross-attention
+        norm_hidden_states = self.norm2(control_hidden_states.float()).type_as(control_hidden_states)
+        attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+        control_hidden_states = control_hidden_states + attn_output
+
+        # 3. Feed-forward
+        norm_hidden_states = (self.norm3(control_hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(
+            control_hidden_states
+        )
+        ff_output = self.ffn(norm_hidden_states)
+        control_hidden_states = (control_hidden_states.float() + ff_output.float() * c_gate_msa).type_as(
+            control_hidden_states
+        )
+
+        conditioning_states = None
+        if self.proj_out is not None:
+            conditioning_states = self.proj_out(control_hidden_states)
+
+        return conditioning_states, control_hidden_states
+
+
+class WanVACETransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
+    r"""
+    A Transformer model for video-like data used in the Wan model.
+
+    Args:
+        patch_size (`Tuple[int]`, defaults to `(1, 2, 2)`):
+            3D patch dimensions for video embedding (t_patch, h_patch, w_patch).
+        num_attention_heads (`int`, defaults to `40`):
+            Fixed length for text embeddings.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        text_dim (`int`, defaults to `512`):
+            Input dimension for text embeddings.
+        freq_dim (`int`, defaults to `256`):
+            Dimension for sinusoidal time embeddings.
+        ffn_dim (`int`, defaults to `13824`):
+            Intermediate dimension in feed-forward network.
+        num_layers (`int`, defaults to `40`):
+            The number of layers of transformer blocks to use.
+        window_size (`Tuple[int]`, defaults to `(-1, -1)`):
+            Window size for local attention (-1 indicates global attention).
+        cross_attn_norm (`bool`, defaults to `True`):
+            Enable cross-attention normalization.
+        qk_norm (`bool`, defaults to `True`):
+            Enable query/key normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        add_img_emb (`bool`, defaults to `False`):
+            Whether to use img_emb.
+        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the added key and value projections. If `None`, no projection is used.
+    """
+
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["patch_embedding", "vace_patch_embedding", "condition_embedder", "norm"]
+    _no_split_modules = ["WanTransformerBlock", "WanVACETransformerBlock"]
+    _keep_in_fp32_modules = ["time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
+    _keys_to_ignore_on_load_unexpected = ["norm_added_q"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: Tuple[int] = (1, 2, 2),
+        num_attention_heads: int = 40,
+        attention_head_dim: int = 128,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        text_dim: int = 4096,
+        freq_dim: int = 256,
+        ffn_dim: int = 13824,
+        num_layers: int = 40,
+        cross_attn_norm: bool = True,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        eps: float = 1e-6,
+        image_dim: Optional[int] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        rope_max_seq_len: int = 1024,
+        pos_embed_seq_len: Optional[int] = None,
+        vace_layers: List[int] = [0, 5, 10, 15, 20, 25, 30, 35],
+        vace_in_channels: int = 96,
+    ) -> None:
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels or in_channels
+
+        if max(vace_layers) >= num_layers:
+            raise ValueError(f"VACE layers {vace_layers} exceed the number of transformer layers {num_layers}.")
+        if 0 not in vace_layers:
+            raise ValueError("VACE layers must include layer 0.")
+
+        # 1. Patch & position embedding
+        self.rope = WanRotaryPosEmbed(attention_head_dim, patch_size, rope_max_seq_len)
+        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+        self.vace_patch_embedding = nn.Conv3d(vace_in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+
+        # 2. Condition embeddings
+        # image_embedding_dim=1280 for I2V model
+        self.condition_embedder = WanTimeTextImageEmbedding(
+            dim=inner_dim,
+            time_freq_dim=freq_dim,
+            time_proj_dim=inner_dim * 6,
+            text_embed_dim=text_dim,
+            image_embed_dim=image_dim,
+            pos_embed_seq_len=pos_embed_seq_len,
+        )
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList(
+            [
+                WanTransformerBlock(
+                    inner_dim, ffn_dim, num_attention_heads, qk_norm, cross_attn_norm, eps, added_kv_proj_dim
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        self.vace_blocks = nn.ModuleList(
+            [
+                WanVACETransformerBlock(
+                    inner_dim,
+                    ffn_dim,
+                    num_attention_heads,
+                    qk_norm,
+                    cross_attn_norm,
+                    eps,
+                    added_kv_proj_dim,
+                    apply_input_projection=i == 0,  # Layer 0 always has input projection and is in vace_layers
+                    apply_output_projection=True,
+                )
+                for i in range(len(vace_layers))
+            ]
+        )
+
+        # 4. Output norm & projection
+        self.norm_out = FP32LayerNorm(inner_dim, eps, elementwise_affine=False)
+        self.proj_out = nn.Linear(inner_dim, out_channels * math.prod(patch_size))
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        control_hidden_states: torch.Tensor = None,
+        control_hidden_states_scale: torch.Tensor = None,
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        if control_hidden_states_scale is None:
+            control_hidden_states_scale = control_hidden_states.new_ones(len(self.config.vace_layers))
+        control_hidden_states_scale = torch.unbind(control_hidden_states_scale)
+        if len(control_hidden_states_scale) != len(self.config.vace_layers):
+            raise ValueError(
+                f"Length of `control_hidden_states_scale` {len(control_hidden_states_scale)} should be "
+                f"equal to {len(self.config.vace_layers)}."
+            )
+
+        # 1. Rotary position embedding
+        rotary_emb = self.rope(hidden_states)
+
+        # 2. Patch embedding
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        control_hidden_states = self.vace_patch_embedding(control_hidden_states)
+        control_hidden_states = control_hidden_states.flatten(2).transpose(1, 2)
+        control_hidden_states_padding = control_hidden_states.new_zeros(
+            batch_size, hidden_states.size(1) - control_hidden_states.size(1), control_hidden_states.size(2)
+        )
+        control_hidden_states = torch.cat([control_hidden_states, control_hidden_states_padding], dim=1)
+
+        # 3. Time embedding
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, encoder_hidden_states, encoder_hidden_states_image
+        )
+        timestep_proj = timestep_proj.unflatten(1, (6, -1))
+
+        # 4. Image embedding
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
+
+        # 5. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            # Prepare VACE hints
+            control_hidden_states_list = []
+            for i, block in enumerate(self.vace_blocks):
+                conditioning_states, control_hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, control_hidden_states, timestep_proj, rotary_emb
+                )
+                control_hidden_states_list.append((conditioning_states, control_hidden_states_scale[i]))
+            control_hidden_states_list = control_hidden_states_list[::-1]
+
+            for i, block in enumerate(self.blocks):
+                hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, timestep_proj, rotary_emb
+                )
+                if i in self.config.vace_layers:
+                    control_hint, scale = control_hidden_states_list.pop()
+                    hidden_states = hidden_states + control_hint * scale
+        else:
+            # Prepare VACE hints
+            control_hidden_states_list = []
+            for i, block in enumerate(self.vace_blocks):
+                conditioning_states, control_hidden_states = block(
+                    hidden_states, encoder_hidden_states, control_hidden_states, timestep_proj, rotary_emb
+                )
+                control_hidden_states_list.append((conditioning_states, control_hidden_states_scale[i]))
+            control_hidden_states_list = control_hidden_states_list[::-1]
+
+            for i, block in enumerate(self.blocks):
+                hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, rotary_emb)
+                if i in self.config.vace_layers:
+                    control_hint, scale = control_hidden_states_list.pop()
+                    hidden_states = hidden_states + control_hint * scale
+
+        # 6. Output norm, projection & unpatchify
+        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
+
+        # Move the shift and scale tensors to the same device as hidden_states.
+        # When using multi-GPU inference via accelerate these will be on the
+        # first device rather than the last device, which hidden_states ends up
+        # on.
+        shift = shift.to(hidden_states.device)
+        scale = scale.to(hidden_states.device)
+
+        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

src/diffusers/pipelines/__init__.py

@@ -290,7 +290,12 @@ else:
     _import_structure["paint_by_example"] = ["PaintByExamplePipeline"]
     _import_structure["pia"] = ["PIAPipeline"]
     _import_structure["pixart_alpha"] = ["PixArtAlphaPipeline", "PixArtSigmaPipeline"]
-    _import_structure["sana"] = ["SanaPipeline", "SanaSprintPipeline", "SanaControlNetPipeline"]
+    _import_structure["sana"] = [
+        "SanaPipeline",
+        "SanaSprintPipeline",
+        "SanaControlNetPipeline",
+        "SanaSprintImg2ImgPipeline",
+    ]
     _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
     _import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"]
     _import_structure["stable_audio"] = [
@@ -366,7 +371,7 @@ else:
         "WuerstchenDecoderPipeline",
         "WuerstchenPriorPipeline",
     ]
-    _import_structure["wan"] = ["WanPipeline", "WanImageToVideoPipeline", "WanVideoToVideoPipeline"]
+    _import_structure["wan"] = ["WanPipeline", "WanImageToVideoPipeline", "WanVideoToVideoPipeline", "WanVACEPipeline"]
 try:
     if not is_onnx_available():
         raise OptionalDependencyNotAvailable()
@@ -675,7 +680,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .paint_by_example import PaintByExamplePipeline
         from .pia import PIAPipeline
         from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
-        from .sana import SanaControlNetPipeline, SanaPipeline, SanaSprintPipeline
+        from .sana import SanaControlNetPipeline, SanaPipeline, SanaSprintImg2ImgPipeline, SanaSprintPipeline
         from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
         from .stable_audio import StableAudioPipeline, StableAudioProjectionModel
@@ -734,7 +739,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             UniDiffuserTextDecoder,
         )
         from .visualcloze import VisualClozeGenerationPipeline, VisualClozePipeline
-        from .wan import WanImageToVideoPipeline, WanPipeline, WanVideoToVideoPipeline
+        from .wan import WanImageToVideoPipeline, WanPipeline, WanVACEPipeline, WanVideoToVideoPipeline
         from .wuerstchen import (
             WuerstchenCombinedPipeline,
             WuerstchenDecoderPipeline,

src/diffusers/pipelines/amused/pipeline_amused.py

@@ -21,7 +21,7 @@ from ...image_processor import VaeImageProcessor
 from ...models import UVit2DModel, VQModel
 from ...schedulers import AmusedScheduler
 from ...utils import is_torch_xla_available, replace_example_docstring
-from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput
 
 
 if is_torch_xla_available():
@@ -47,7 +47,8 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class AmusedPipeline(DiffusionPipeline):
+class AmusedPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
+    _last_supported_version = "0.33.1"
     image_processor: VaeImageProcessor
     vqvae: VQModel
     tokenizer: CLIPTokenizer

src/diffusers/pipelines/amused/pipeline_amused_img2img.py

@@ -21,7 +21,7 @@ from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...models import UVit2DModel, VQModel
 from ...schedulers import AmusedScheduler
 from ...utils import is_torch_xla_available, replace_example_docstring
-from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput
 
 
 if is_torch_xla_available():
@@ -57,7 +57,8 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class AmusedImg2ImgPipeline(DiffusionPipeline):
+class AmusedImg2ImgPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
+    _last_supported_version = "0.33.1"
     image_processor: VaeImageProcessor
     vqvae: VQModel
     tokenizer: CLIPTokenizer

src/diffusers/pipelines/amused/pipeline_amused_inpaint.py

@@ -22,7 +22,7 @@ from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...models import UVit2DModel, VQModel
 from ...schedulers import AmusedScheduler
 from ...utils import is_torch_xla_available, replace_example_docstring
-from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput
 
 
 if is_torch_xla_available():
@@ -65,7 +65,8 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class AmusedInpaintPipeline(DiffusionPipeline):
+class AmusedInpaintPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
+    _last_supported_version = "0.33.1"
     image_processor: VaeImageProcessor
     vqvae: VQModel
     tokenizer: CLIPTokenizer

src/diffusers/pipelines/audioldm/pipeline_audioldm.py

@@ -24,7 +24,7 @@ from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import is_torch_xla_available, logging, replace_example_docstring
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import AudioPipelineOutput, DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 
 
 if is_torch_xla_available():
@@ -57,7 +57,7 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class AudioLDMPipeline(DiffusionPipeline, StableDiffusionMixin):
+class AudioLDMPipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin):
     r"""
     Pipeline for text-to-audio generation using AudioLDM.
 
@@ -81,6 +81,7 @@ class AudioLDMPipeline(DiffusionPipeline, StableDiffusionMixin):
             Vocoder of class `SpeechT5HifiGan`.
     """
 
+    _last_supported_version = "0.33.1"
     model_cpu_offload_seq = "text_encoder->unet->vae"
 
     def __init__(

src/diffusers/pipelines/blip_diffusion/pipeline_blip_diffusion.py

@@ -25,7 +25,7 @@ from ...utils import (
     replace_example_docstring,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput
 from .blip_image_processing import BlipImageProcessor
 from .modeling_blip2 import Blip2QFormerModel
 from .modeling_ctx_clip import ContextCLIPTextModel
@@ -81,7 +81,7 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class BlipDiffusionPipeline(DiffusionPipeline):
+class BlipDiffusionPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
     """
     Pipeline for Zero-Shot Subject Driven Generation using Blip Diffusion.
 
@@ -107,6 +107,7 @@ class BlipDiffusionPipeline(DiffusionPipeline):
             Position of the context token in the text encoder.
     """
 
+    _last_supported_version = "0.33.1"
     model_cpu_offload_seq = "qformer->text_encoder->unet->vae"
 
     def __init__(

src/diffusers/pipelines/consisid/consisid_utils.py

@@ -166,7 +166,7 @@ def process_face_embeddings(
         raise RuntimeError("facexlib align face fail")
     align_face = face_helper_1.cropped_faces[0]  # (512, 512, 3)  # RGB
 
-    # incase insightface didn't detect face
+    # in case insightface didn't detect face
     if id_ante_embedding is None:
         logger.warning("Failed to detect face using insightface. Extracting embedding with align face")
         id_ante_embedding = face_helper_2.get_feat(align_face)

src/diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs.py

@@ -37,7 +37,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
@@ -98,6 +98,7 @@ EXAMPLE_DOC_STRING = """
 
 
 class StableDiffusionControlNetXSPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
     TextualInversionLoaderMixin,
@@ -138,6 +139,7 @@ class StableDiffusionControlNetXSPipeline(
             A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
     """
 
+    _last_supported_version = "0.33.1"
     model_cpu_offload_seq = "text_encoder->unet->vae"
     _optional_components = ["safety_checker", "feature_extractor"]
     _exclude_from_cpu_offload = ["safety_checker"]

src/diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs_sd_xl.py

@@ -46,7 +46,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
-from ..pipeline_utils import DiffusionPipeline
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline
 from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 
 
@@ -114,6 +114,7 @@ EXAMPLE_DOC_STRING = """
 
 
 class StableDiffusionXLControlNetXSPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     TextualInversionLoaderMixin,
     StableDiffusionXLLoraLoaderMixin,
@@ -158,6 +159,7 @@ class StableDiffusionXLControlNetXSPipeline(
             watermarker is used.
     """
 
+    _last_supported_version = "0.33.1"
     model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
     _optional_components = [
         "tokenizer",

src/diffusers/pipelines/dance_diffusion/pipeline_dance_diffusion.py

@@ -21,7 +21,7 @@ from ...models import UNet1DModel
 from ...schedulers import SchedulerMixin
 from ...utils import is_torch_xla_available, logging
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline
+from ..pipeline_utils import AudioPipelineOutput, DeprecatedPipelineMixin, DiffusionPipeline
 
 
 if is_torch_xla_available():
@@ -34,7 +34,7 @@ else:
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class DanceDiffusionPipeline(DiffusionPipeline):
+class DanceDiffusionPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
     r"""
     Pipeline for audio generation.
 
@@ -49,6 +49,7 @@ class DanceDiffusionPipeline(DiffusionPipeline):
             [`IPNDMScheduler`].
     """
 
+    _last_supported_version = "0.33.1"
     model_cpu_offload_seq = "unet"
 
     def __init__(self, unet: UNet1DModel, scheduler: SchedulerMixin):

src/diffusers/pipelines/flux/pipeline_flux.py

@@ -700,7 +700,7 @@ class FluxPipeline(
             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`.
+                tensor will be 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.

src/diffusers/pipelines/i2vgen_xl/pipeline_i2vgen_xl.py

@@ -33,7 +33,7 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 
 
 if is_torch_xla_available():
@@ -97,9 +97,11 @@ class I2VGenXLPipelineOutput(BaseOutput):
 
 
 class I2VGenXLPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
 ):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for image-to-video generation as proposed in [I2VGenXL](https://i2vgen-xl.github.io/).
 

src/diffusers/pipelines/musicldm/pipeline_musicldm.py

@@ -36,7 +36,7 @@ from ...utils import (
     replace_example_docstring,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import AudioPipelineOutput, DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 
 
 if is_librosa_available():
@@ -76,7 +76,8 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class MusicLDMPipeline(DiffusionPipeline, StableDiffusionMixin):
+class MusicLDMPipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for text-to-audio generation using MusicLDM.
 

src/diffusers/pipelines/paint_by_example/pipeline_paint_by_example.py

@@ -25,7 +25,7 @@ from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
 from ...utils import deprecate, is_torch_xla_available, logging
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from .image_encoder import PaintByExampleImageEncoder
@@ -155,7 +155,8 @@ def prepare_mask_and_masked_image(image, mask):
     return mask, masked_image
 
 
-class PaintByExamplePipeline(DiffusionPipeline, StableDiffusionMixin):
+class PaintByExamplePipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin):
+    _last_supported_version = "0.33.1"
     r"""
     <Tip warning={true}>
 

src/diffusers/pipelines/pia/pipeline_pia.py

@@ -46,7 +46,7 @@ from ...utils import (
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
 from ..free_init_utils import FreeInitMixin
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 
 
 if is_torch_xla_available():
@@ -132,6 +132,7 @@ class PIAPipelineOutput(BaseOutput):
 
 
 class PIAPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
     TextualInversionLoaderMixin,
@@ -140,6 +141,7 @@ class PIAPipeline(
     FromSingleFileMixin,
     FreeInitMixin,
 ):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for text-to-video generation.
 

src/diffusers/pipelines/pipeline_utils.py

@@ -139,6 +139,43 @@ class AudioPipelineOutput(BaseOutput):
     audios: np.ndarray
 
 
+class DeprecatedPipelineMixin:
+    """
+    A mixin that can be used to mark a pipeline as deprecated.
+
+    Pipelines inheriting from this mixin will raise a warning when instantiated, indicating that they are deprecated
+    and won't receive updates past the specified version. Tests will be skipped for pipelines that inherit from this
+    mixin.
+
+    Example usage:
+    ```python
+    class MyDeprecatedPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
+        _last_supported_version = "0.20.0"
+
+        def __init__(self, *args, **kwargs):
+            super().__init__(*args, **kwargs)
+    ```
+    """
+
+    # Override this in the inheriting class to specify the last version that will support this pipeline
+    _last_supported_version = None
+
+    def __init__(self, *args, **kwargs):
+        # Get the class name for the warning message
+        class_name = self.__class__.__name__
+
+        # Get the last supported version or use the current version if not specified
+        version_info = getattr(self.__class__, "_last_supported_version", __version__)
+
+        # Raise a warning that this pipeline is deprecated
+        logger.warning(
+            f"The {class_name} has been deprecated and will not receive bug fixes or feature updates after Diffusers version {version_info}. "
+        )
+
+        # Call the parent class's __init__ method
+        super().__init__(*args, **kwargs)
+
+
 class DiffusionPipeline(ConfigMixin, PushToHubMixin):
     r"""
     Base class for all pipelines.

src/diffusers/pipelines/sana/__init__.py

@@ -25,6 +25,7 @@ else:
     _import_structure["pipeline_sana"] = ["SanaPipeline"]
     _import_structure["pipeline_sana_controlnet"] = ["SanaControlNetPipeline"]
     _import_structure["pipeline_sana_sprint"] = ["SanaSprintPipeline"]
+    _import_structure["pipeline_sana_sprint_img2img"] = ["SanaSprintImg2ImgPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -37,6 +38,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .pipeline_sana import SanaPipeline
         from .pipeline_sana_controlnet import SanaControlNetPipeline
         from .pipeline_sana_sprint import SanaSprintPipeline
+        from .pipeline_sana_sprint_img2img import SanaSprintImg2ImgPipeline
 else:
     import sys
 

src/diffusers/pipelines/sana/pipeline_sana_sprint_img2img.py

@@ -0,0 +1,975 @@
+# Copyright 2024 PixArt-Sigma Authors 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 inspect
+import re
+import urllib.parse as ul
+import warnings
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PipelineImageInput, PixArtImageProcessor
+from ...loaders import SanaLoraLoaderMixin
+from ...models import AutoencoderDC, SanaTransformer2DModel
+from ...schedulers import DPMSolverMultistepScheduler
+from ...utils import (
+    BACKENDS_MAPPING,
+    USE_PEFT_BACKEND,
+    is_bs4_available,
+    is_ftfy_available,
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from ..pixart_alpha.pipeline_pixart_alpha import ASPECT_RATIO_1024_BIN
+from .pipeline_output import SanaPipelineOutput
+
+
+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
+
+if is_bs4_available():
+    from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+    import ftfy
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import SanaSprintImg2ImgPipeline
+        >>> from diffusers.utils.loading_utils import load_image
+
+        >>> pipe = SanaSprintImg2ImgPipeline.from_pretrained(
+        ...     "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers", torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/penguin.png"
+        ... )
+
+
+        >>> image = pipe(prompt="a cute pink bear", image=image, strength=0.5, height=832, width=480).images[0]
+        >>> image[0].save("output.png")
+        ```
+"""
+
+
+# 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 SanaSprintImg2ImgPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-image generation using [SANA-Sprint](https://huggingface.co/papers/2503.09641).
+    """
+
+    # fmt: off
+    bad_punct_regex = re.compile(
+        r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}")
+    # fmt: on
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: Union[GemmaTokenizer, GemmaTokenizerFast],
+        text_encoder: Gemma2PreTrainedModel,
+        vae: AutoencoderDC,
+        transformer: SanaTransformer2DModel,
+        scheduler: DPMSolverMultistepScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+
+        self.vae_scale_factor = (
+            2 ** (len(self.vae.config.encoder_block_out_channels) - 1)
+            if hasattr(self, "vae") and self.vae is not None
+            else 32
+        )
+        self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.enable_vae_slicing
+    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()
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.disable_vae_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()
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.enable_vae_tiling
+    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()
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._get_gemma_prompt_embeds
+    def _get_gemma_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        device: torch.device,
+        dtype: torch.dtype,
+        clean_caption: bool = False,
+        max_sequence_length: int = 300,
+        complex_human_instruction: Optional[List[str]] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`, *optional*):
+                torch device to place the resulting embeddings on
+            clean_caption (`bool`, defaults to `False`):
+                If `True`, the function will preprocess and clean the provided caption before encoding.
+            max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt.
+            complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`):
+                If `complex_human_instruction` is not empty, the function will use the complex Human instruction for
+                the prompt.
+        """
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if getattr(self, "tokenizer", None) is not None:
+            self.tokenizer.padding_side = "right"
+
+        prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+
+        # prepare complex human instruction
+        if not complex_human_instruction:
+            max_length_all = max_sequence_length
+        else:
+            chi_prompt = "\n".join(complex_human_instruction)
+            prompt = [chi_prompt + p for p in prompt]
+            num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt))
+            max_length_all = num_chi_prompt_tokens + max_sequence_length - 2
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_length_all,
+            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.to(device)
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)
+        prompt_embeds = prompt_embeds[0].to(dtype=dtype, device=device)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana_sprint.SanaSprintPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        clean_caption: bool = False,
+        max_sequence_length: int = 300,
+        complex_human_instruction: Optional[List[str]] = None,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                number of images that should be generated per prompt
+            device: (`torch.device`, *optional*):
+                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.
+            clean_caption (`bool`, defaults to `False`):
+                If `True`, the function will preprocess and clean the provided caption before encoding.
+            max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt.
+            complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`):
+                If `complex_human_instruction` is not empty, the function will use the complex Human instruction for
+                the prompt.
+        """
+
+        if device is None:
+            device = self._execution_device
+
+        if self.text_encoder is not None:
+            dtype = self.text_encoder.dtype
+        else:
+            dtype = None
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder, lora_scale)
+
+        if getattr(self, "tokenizer", None) is not None:
+            self.tokenizer.padding_side = "right"
+
+        # See Section 3.1. of the paper.
+        max_length = max_sequence_length
+        select_index = [0] + list(range(-max_length + 1, 0))
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                dtype=dtype,
+                clean_caption=clean_caption,
+                max_sequence_length=max_sequence_length,
+                complex_human_instruction=complex_human_instruction,
+            )
+
+            prompt_embeds = prompt_embeds[:, select_index]
+            prompt_attention_mask = prompt_attention_mask[:, select_index]
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1)
+
+        if self.text_encoder is not None:
+            if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # 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://huggingface.co/papers/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.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps
+    def get_timesteps(self, num_inference_steps, strength, device):
+        # get the original timestep using init_timestep
+        init_timestep = min(num_inference_steps * strength, num_inference_steps)
+
+        t_start = int(max(num_inference_steps - init_timestep, 0))
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+        if hasattr(self.scheduler, "set_begin_index"):
+            self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+        return timesteps, num_inference_steps - t_start
+
+    def check_inputs(
+        self,
+        prompt,
+        strength,
+        height,
+        width,
+        num_inference_steps,
+        timesteps,
+        max_timesteps,
+        intermediate_timesteps,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_embeds=None,
+        prompt_attention_mask=None,
+    ):
+        if strength < 0 or strength > 1:
+            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+        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 timesteps is not None and len(timesteps) != num_inference_steps + 1:
+            raise ValueError("If providing custom timesteps, `timesteps` must be of length `num_inference_steps + 1`.")
+
+        if timesteps is not None and max_timesteps is not None:
+            raise ValueError("If providing custom timesteps, `max_timesteps` should not be provided.")
+
+        if timesteps is None and max_timesteps is None:
+            raise ValueError("Should provide either `timesteps` or `max_timesteps`.")
+
+        if intermediate_timesteps is not None and num_inference_steps != 2:
+            raise ValueError("Intermediate timesteps for SCM is not supported when num_inference_steps != 2.")
+
+    # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+    def _text_preprocessing(self, text, clean_caption=False):
+        if clean_caption and not is_bs4_available():
+            logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if clean_caption and not is_ftfy_available():
+            logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if not isinstance(text, (tuple, list)):
+            text = [text]
+
+        def process(text: str):
+            if clean_caption:
+                text = self._clean_caption(text)
+                text = self._clean_caption(text)
+            else:
+                text = text.lower().strip()
+            return text
+
+        return [process(t) for t in text]
+
+    # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
+    def _clean_caption(self, caption):
+        caption = str(caption)
+        caption = ul.unquote_plus(caption)
+        caption = caption.strip().lower()
+        caption = re.sub("<person>", "person", caption)
+        # urls:
+        caption = re.sub(
+            r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        caption = re.sub(
+            r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        # html:
+        caption = BeautifulSoup(caption, features="html.parser").text
+
+        # @<nickname>
+        caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+        # 31C0—31EF CJK Strokes
+        # 31F0—31FF Katakana Phonetic Extensions
+        # 3200—32FF Enclosed CJK Letters and Months
+        # 3300—33FF CJK Compatibility
+        # 3400—4DBF CJK Unified Ideographs Extension A
+        # 4DC0—4DFF Yijing Hexagram Symbols
+        # 4E00—9FFF CJK Unified Ideographs
+        caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+        caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+        caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+        caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+        caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+        caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+        caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+        #######################################################
+
+        # все виды тире / all types of dash --> "-"
+        caption = re.sub(
+            r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+",  # noqa
+            "-",
+            caption,
+        )
+
+        # кавычки к одному стандарту
+        caption = re.sub(r"[`´«»“”¨]", '"', caption)
+        caption = re.sub(r"[‘’]", "'", caption)
+
+        # &quot;
+        caption = re.sub(r"&quot;?", "", caption)
+        # &amp
+        caption = re.sub(r"&amp", "", caption)
+
+        # ip addresses:
+        caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+        # article ids:
+        caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+        # \n
+        caption = re.sub(r"\\n", " ", caption)
+
+        # "#123"
+        caption = re.sub(r"#\d{1,3}\b", "", caption)
+        # "#12345.."
+        caption = re.sub(r"#\d{5,}\b", "", caption)
+        # "123456.."
+        caption = re.sub(r"\b\d{6,}\b", "", caption)
+        # filenames:
+        caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+        #
+        caption = re.sub(r"[\"\']{2,}", r'"', caption)  # """AUSVERKAUFT"""
+        caption = re.sub(r"[\.]{2,}", r" ", caption)  # """AUSVERKAUFT"""
+
+        caption = re.sub(self.bad_punct_regex, r" ", caption)  # ***AUSVERKAUFT***, #AUSVERKAUFT
+        caption = re.sub(r"\s+\.\s+", r" ", caption)  # " . "
+
+        # this-is-my-cute-cat / this_is_my_cute_cat
+        regex2 = re.compile(r"(?:\-|\_)")
+        if len(re.findall(regex2, caption)) > 3:
+            caption = re.sub(regex2, " ", caption)
+
+        caption = ftfy.fix_text(caption)
+        caption = html.unescape(html.unescape(caption))
+
+        caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption)  # jc6640
+        caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption)  # jc6640vc
+        caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption)  # 6640vc231
+
+        caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+        caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+        caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+        caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+        caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+        caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption)  # j2d1a2a...
+
+        caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+        caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+        caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+        caption = re.sub(r"\s+", " ", caption)
+
+        caption.strip()
+
+        caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+        caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+        caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+        caption = re.sub(r"^\.\S+$", "", caption)
+
+        return caption.strip()
+
+    def prepare_image(
+        self,
+        image: PipelineImageInput,
+        width: int,
+        height: int,
+        device: torch.device,
+        dtype: torch.dtype,
+    ):
+        if isinstance(image, torch.Tensor):
+            if image.ndim == 3:
+                image = image.unsqueeze(0)
+            # Resize if current dimensions do not match target dimensions.
+            if image.shape[2] != height or image.shape[3] != width:
+                image = F.interpolate(image, size=(height, width), mode="bilinear", align_corners=False)
+
+            image = self.image_processor.preprocess(image, height=height, width=width)
+
+        else:
+            image = self.image_processor.preprocess(image, height=height, width=width)
+
+        image = image.to(device=device, dtype=dtype)
+
+        return image
+
+    def prepare_latents(
+        self, image, timestep, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None
+    ):
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+
+        if image.shape[1] != num_channels_latents:
+            image = self.vae.encode(image).latent
+            image_latents = image * self.vae.config.scaling_factor * self.scheduler.config.sigma_data
+        else:
+            image_latents = image
+        if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
+            # expand init_latents for batch_size
+            additional_image_per_prompt = batch_size // image_latents.shape[0]
+            image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
+        elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
+            raise ValueError(
+                f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+            )
+        else:
+            image_latents = torch.cat([image_latents], dim=0)
+
+        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."
+            )
+
+        # adapt from https://github.com/huggingface/diffusers/blob/c36f8487df35895421c15f351c7d360bd680[…]/examples/research_projects/sana/train_sana_sprint_diffusers.py
+        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) * self.scheduler.config.sigma_data
+        latents = torch.cos(timestep) * image_latents + torch.sin(timestep) * noise
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @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,
+        num_inference_steps: int = 2,
+        timesteps: List[int] = None,
+        max_timesteps: float = 1.57080,
+        intermediate_timesteps: float = 1.3,
+        guidance_scale: float = 4.5,
+        image: PipelineImageInput = None,
+        strength: float = 0.6,
+        num_images_per_prompt: Optional[int] = 1,
+        height: int = 1024,
+        width: int = 1024,
+        eta: float = 0.0,
+        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,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        clean_caption: bool = False,
+        use_resolution_binning: 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 = 300,
+        complex_human_instruction: List[str] = [
+            "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:",
+            "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.",
+            "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.",
+            "Here are examples of how to transform or refine prompts:",
+            "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.",
+            "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.",
+            "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:",
+            "User Prompt: ",
+        ],
+    ) -> Union[SanaPipelineOutput, 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.
+            num_inference_steps (`int`, *optional*, defaults to 20):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            max_timesteps (`float`, *optional*, defaults to 1.57080):
+                The maximum timestep value used in the SCM scheduler.
+            intermediate_timesteps (`float`, *optional*, defaults to 1.3):
+                The intermediate timestep value used in SCM scheduler (only used when num_inference_steps=2).
+            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 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_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The width in pixels of the generated image.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            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.
+            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.IFPipelineOutput`] instead of a plain tuple.
+            attention_kwargs:
+                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).
+            clean_caption (`bool`, *optional*, defaults to `True`):
+                Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+                be installed. If the dependencies are not installed, the embeddings will be created from the raw
+                prompt.
+            use_resolution_binning (`bool` defaults to `True`):
+                If set to `True`, the requested height and width are first mapped to the closest resolutions using
+                `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to
+                the requested resolution. Useful for generating non-square images.
+            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 `300`):
+                Maximum sequence length to use with the `prompt`.
+            complex_human_instruction (`List[str]`, *optional*):
+                Instructions for complex human attention:
+                https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] 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
+        if use_resolution_binning:
+            if self.transformer.config.sample_size == 32:
+                aspect_ratio_bin = ASPECT_RATIO_1024_BIN
+            else:
+                raise ValueError("Invalid sample size")
+            orig_height, orig_width = height, width
+            height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin)
+
+        self.check_inputs(
+            prompt=prompt,
+            strength=strength,
+            height=height,
+            width=width,
+            num_inference_steps=num_inference_steps,
+            timesteps=timesteps,
+            max_timesteps=max_timesteps,
+            intermediate_timesteps=intermediate_timesteps,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            prompt_embeds=prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._interrupt = False
+
+        # 2. Default height and width to transformer
+        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
+        lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None
+
+        # 2. Preprocess image
+        init_image = self.prepare_image(image, width, height, device, self.vae.dtype)
+
+        # 3. Encode input prompt
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            clean_caption=clean_caption,
+            max_sequence_length=max_sequence_length,
+            complex_human_instruction=complex_human_instruction,
+            lora_scale=lora_scale,
+        )
+
+        # 5. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas=None,
+            max_timesteps=max_timesteps,
+            intermediate_timesteps=intermediate_timesteps,
+        )
+        if hasattr(self.scheduler, "set_begin_index"):
+            self.scheduler.set_begin_index(0)
+
+        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+        if num_inference_steps < 1:
+            raise ValueError(
+                f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
+                f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
+            )
+        latent_timestep = timesteps[:1]
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            init_image,
+            latent_timestep,
+            batch_size * num_images_per_prompt,
+            latent_channels,
+            height,
+            width,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # I think this is redundant given the scaling in prepare_latents
+        # latents = latents * self.scheduler.config.sigma_data
+
+        guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
+        guidance = guidance.expand(latents.shape[0]).to(prompt_embeds.dtype)
+        guidance = guidance * self.transformer.config.guidance_embeds_scale
+
+        # 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. Denoising loop
+        timesteps = timesteps[:-1]
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        transformer_dtype = self.transformer.dtype
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0])
+                latents_model_input = latents / self.scheduler.config.sigma_data
+
+                scm_timestep = torch.sin(timestep) / (torch.cos(timestep) + torch.sin(timestep))
+
+                scm_timestep_expanded = scm_timestep.view(-1, 1, 1, 1)
+                latent_model_input = latents_model_input * torch.sqrt(
+                    scm_timestep_expanded**2 + (1 - scm_timestep_expanded) ** 2
+                )
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    latent_model_input.to(dtype=transformer_dtype),
+                    encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype),
+                    encoder_attention_mask=prompt_attention_mask,
+                    guidance=guidance,
+                    timestep=scm_timestep,
+                    return_dict=False,
+                    attention_kwargs=self.attention_kwargs,
+                )[0]
+
+                noise_pred = (
+                    (1 - 2 * scm_timestep_expanded) * latent_model_input
+                    + (1 - 2 * scm_timestep_expanded + 2 * scm_timestep_expanded**2) * noise_pred
+                ) / torch.sqrt(scm_timestep_expanded**2 + (1 - scm_timestep_expanded) ** 2)
+                noise_pred = noise_pred.float() * self.scheduler.config.sigma_data
+
+                # compute previous image: x_t -> x_t-1
+                latents, denoised = self.scheduler.step(
+                    noise_pred, timestep, latents, **extra_step_kwargs, return_dict=False
+                )
+
+                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()
+
+        latents = denoised / self.scheduler.config.sigma_data
+        if output_type == "latent":
+            image = latents
+        else:
+            latents = latents.to(self.vae.dtype)
+            try:
+                image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            except torch.cuda.OutOfMemoryError as e:
+                warnings.warn(
+                    f"{e}. \n"
+                    f"Try to use VAE tiling for large images. For example: \n"
+                    f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)"
+                )
+            if use_resolution_binning:
+                image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height)
+
+        if not output_type == "latent":
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return SanaPipelineOutput(images=image)

src/diffusers/pipelines/semantic_stable_diffusion/pipeline_semantic_stable_diffusion.py

@@ -11,7 +11,7 @@ from ...pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyCh
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import deprecate, is_torch_xla_available, logging
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from .pipeline_output import SemanticStableDiffusionPipelineOutput
 
 
@@ -25,7 +25,8 @@ else:
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class SemanticStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
+class SemanticStableDiffusionPipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for text-to-image generation using Stable Diffusion with latent editing.
 

src/diffusers/pipelines/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py

@@ -37,7 +37,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
@@ -179,7 +179,9 @@ class AttendExciteAttnProcessor:
         return hidden_states
 
 
-class StableDiffusionAttendAndExcitePipeline(DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin):
+class StableDiffusionAttendAndExcitePipeline(
+    DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin
+):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion and Attend-and-Excite.
 
@@ -209,6 +211,8 @@ class StableDiffusionAttendAndExcitePipeline(DiffusionPipeline, StableDiffusionM
             A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
     """
 
+    _last_supported_version = "0.33.1"
+
     model_cpu_offload_seq = "text_encoder->unet->vae"
     _optional_components = ["safety_checker", "feature_extractor"]
     _exclude_from_cpu_offload = ["safety_checker"]

src/diffusers/pipelines/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py

@@ -40,7 +40,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
@@ -242,7 +242,11 @@ def preprocess_mask(mask, batch_size: int = 1):
 
 
 class StableDiffusionDiffEditPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin
+    DeprecatedPipelineMixin,
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    StableDiffusionLoraLoaderMixin,
 ):
     r"""
     <Tip warning={true}>
@@ -282,6 +286,8 @@ class StableDiffusionDiffEditPipeline(
             A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
     """
 
+    _last_supported_version = "0.33.1"
+
     model_cpu_offload_seq = "text_encoder->unet->vae"
     _optional_components = ["safety_checker", "feature_extractor", "inverse_scheduler"]
     _exclude_from_cpu_offload = ["safety_checker"]

src/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py

@@ -36,7 +36,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
@@ -108,7 +108,7 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class StableDiffusionGLIGENPipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionGLIGENPipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion with Grounded-Language-to-Image Generation (GLIGEN).
 
@@ -135,6 +135,8 @@ class StableDiffusionGLIGENPipeline(DiffusionPipeline, StableDiffusionMixin):
             A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
     """
 
+    _last_supported_version = "0.33.1"
+
     _optional_components = ["safety_checker", "feature_extractor"]
     model_cpu_offload_seq = "text_encoder->unet->vae"
     _exclude_from_cpu_offload = ["safety_checker"]

src/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py

@@ -41,7 +41,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.clip_image_project_model import CLIPImageProjection
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
@@ -160,7 +160,7 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class StableDiffusionGLIGENTextImagePipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionGLIGENTextImagePipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion with Grounded-Language-to-Image Generation (GLIGEN).
 
@@ -193,6 +193,8 @@ class StableDiffusionGLIGENTextImagePipeline(DiffusionPipeline, StableDiffusionM
             A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
     """
 
+    _last_supported_version = "0.33.1"
+
     model_cpu_offload_seq = "text_encoder->unet->vae"
     _optional_components = ["safety_checker", "feature_extractor"]
     _exclude_from_cpu_offload = ["safety_checker"]

src/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py

@@ -42,7 +42,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
 
 
@@ -64,7 +64,11 @@ class ModelWrapper:
 
 
 class StableDiffusionKDiffusionPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin
+    DeprecatedPipelineMixin,
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    StableDiffusionLoraLoaderMixin,
 ):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
@@ -105,6 +109,8 @@ class StableDiffusionKDiffusionPipeline(
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 
+    _last_supported_version = "0.33.1"
+
     model_cpu_offload_seq = "text_encoder->unet->vae"
     _optional_components = ["safety_checker", "feature_extractor"]
     _exclude_from_cpu_offload = ["safety_checker"]

src/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py

@@ -48,7 +48,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 
 
@@ -88,6 +88,7 @@ class ModelWrapper:
 
 
 class StableDiffusionXLKDiffusionPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
     FromSingleFileMixin,
@@ -95,6 +96,8 @@ class StableDiffusionXLKDiffusionPipeline(
     TextualInversionLoaderMixin,
     IPAdapterMixin,
 ):
+    _last_supported_version = "0.33.1"
+
     r"""
     Pipeline for text-to-image generation using Stable Diffusion XL and k-diffusion.
 

src/diffusers/pipelines/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py

@@ -37,7 +37,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
@@ -178,6 +178,7 @@ class LDM3DPipelineOutput(BaseOutput):
 
 
 class StableDiffusionLDM3DPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
     TextualInversionLoaderMixin,
@@ -185,6 +186,8 @@ class StableDiffusionLDM3DPipeline(
     StableDiffusionLoraLoaderMixin,
     FromSingleFileMixin,
 ):
+    _last_supported_version = "0.33.1"
+
     r"""
     Pipeline for text-to-image and 3D generation using LDM3D.
 

src/diffusers/pipelines/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py

@@ -33,7 +33,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
@@ -156,12 +156,15 @@ def retrieve_timesteps(
 
 
 class StableDiffusionPanoramaPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
     TextualInversionLoaderMixin,
     StableDiffusionLoraLoaderMixin,
     IPAdapterMixin,
 ):
+    _last_supported_version = "0.33.1"
+
     r"""
     Pipeline for text-to-image generation using MultiDiffusion.
 

src/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py

@@ -14,7 +14,7 @@ from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import deprecate, is_torch_xla_available, logging
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from . import StableDiffusionSafePipelineOutput
 from .safety_checker import SafeStableDiffusionSafetyChecker
 
@@ -29,7 +29,9 @@ else:
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class StableDiffusionPipelineSafe(DiffusionPipeline, StableDiffusionMixin, IPAdapterMixin):
+class StableDiffusionPipelineSafe(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin, IPAdapterMixin):
+    _last_supported_version = "0.33.1"
+
     r"""
     Pipeline based on the [`StableDiffusionPipeline`] for text-to-image generation using Safe Latent Diffusion.
 

src/diffusers/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py

@@ -34,7 +34,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
@@ -107,7 +107,11 @@ class CrossAttnStoreProcessor:
 
 
 # Modified to get self-attention guidance scale in this paper (https://huggingface.co/papers/2210.00939) as an input
-class StableDiffusionSAGPipeline(DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, IPAdapterMixin):
+class StableDiffusionSAGPipeline(
+    DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, IPAdapterMixin
+):
+    _last_supported_version = "0.33.1"
+
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py

@@ -33,7 +33,7 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from . import TextToVideoSDPipelineOutput
 
 
@@ -68,8 +68,13 @@ EXAMPLE_DOC_STRING = """
 
 
 class TextToVideoSDPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin
+    DeprecatedPipelineMixin,
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    StableDiffusionLoraLoaderMixin,
 ):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for text-to-video generation.
 

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py

@@ -34,7 +34,7 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from . import TextToVideoSDPipelineOutput
 
 
@@ -103,8 +103,13 @@ def retrieve_latents(
 
 
 class VideoToVideoSDPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin
+    DeprecatedPipelineMixin,
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    StableDiffusionLoraLoaderMixin,
 ):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for text-guided video-to-video generation.
 

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py

@@ -24,7 +24,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionSafetyChecker
 
 
@@ -296,12 +296,14 @@ def create_motion_field_and_warp_latents(motion_field_strength_x, motion_field_s
 
 
 class TextToVideoZeroPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
     TextualInversionLoaderMixin,
     StableDiffusionLoraLoaderMixin,
     FromSingleFileMixin,
 ):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for zero-shot text-to-video generation using Stable Diffusion.
 

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py

@@ -35,7 +35,7 @@ from ...utils import (
     unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 
 
 if is_invisible_watermark_available():
@@ -346,11 +346,13 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
 
 
 class TextToVideoZeroSDXLPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
     StableDiffusionXLLoraLoaderMixin,
     TextualInversionLoaderMixin,
 ):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for zero-shot text-to-video generation using Stable Diffusion XL.
 

src/diffusers/pipelines/unclip/pipeline_unclip.py

@@ -24,7 +24,7 @@ from ...models import PriorTransformer, UNet2DConditionModel, UNet2DModel
 from ...schedulers import UnCLIPScheduler
 from ...utils import is_torch_xla_available, logging
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput
 from .text_proj import UnCLIPTextProjModel
 
 
@@ -38,7 +38,7 @@ else:
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class UnCLIPPipeline(DiffusionPipeline):
+class UnCLIPPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
     """
     Pipeline for text-to-image generation using unCLIP.
 
@@ -69,6 +69,7 @@ class UnCLIPPipeline(DiffusionPipeline):
 
     """
 
+    _last_supported_version = "0.33.1"
     _exclude_from_cpu_offload = ["prior"]
 
     prior: PriorTransformer

src/diffusers/pipelines/unclip/pipeline_unclip_image_variation.py

@@ -29,7 +29,7 @@ from ...models import UNet2DConditionModel, UNet2DModel
 from ...schedulers import UnCLIPScheduler
 from ...utils import is_torch_xla_available, logging
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput
 from .text_proj import UnCLIPTextProjModel
 
 
@@ -43,7 +43,7 @@ else:
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class UnCLIPImageVariationPipeline(DiffusionPipeline):
+class UnCLIPImageVariationPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
     """
     Pipeline to generate image variations from an input image using UnCLIP.
 
@@ -73,6 +73,7 @@ class UnCLIPImageVariationPipeline(DiffusionPipeline):
             Scheduler used in the super resolution denoising process (a modified [`DDPMScheduler`]).
     """
 
+    _last_supported_version = "0.33.1"
     decoder: UNet2DConditionModel
     text_proj: UnCLIPTextProjModel
     text_encoder: CLIPTextModelWithProjection

src/diffusers/pipelines/unidiffuser/pipeline_unidiffuser.py

@@ -28,7 +28,7 @@ from ...utils import (
 )
 from ...utils.outputs import BaseOutput
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline
 from .modeling_text_decoder import UniDiffuserTextDecoder
 from .modeling_uvit import UniDiffuserModel
 
@@ -62,7 +62,7 @@ class ImageTextPipelineOutput(BaseOutput):
     text: Optional[Union[List[str], List[List[str]]]]
 
 
-class UniDiffuserPipeline(DiffusionPipeline):
+class UniDiffuserPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
     r"""
     Pipeline for a bimodal image-text model which supports unconditional text and image generation, text-conditioned
     image generation, image-conditioned text generation, and joint image-text generation.
@@ -96,6 +96,7 @@ class UniDiffuserPipeline(DiffusionPipeline):
             original UniDiffuser paper uses the [`DPMSolverMultistepScheduler`] scheduler.
     """
 
+    _last_supported_version = "0.33.1"
     # TODO: support for moving submodules for components with enable_model_cpu_offload
     model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae->text_decoder"
 

src/diffusers/pipelines/wan/__init__.py

@@ -24,6 +24,7 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["pipeline_wan"] = ["WanPipeline"]
     _import_structure["pipeline_wan_i2v"] = ["WanImageToVideoPipeline"]
+    _import_structure["pipeline_wan_vace"] = ["WanVACEPipeline"]
     _import_structure["pipeline_wan_video2video"] = ["WanVideoToVideoPipeline"]
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -35,6 +36,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     else:
         from .pipeline_wan import WanPipeline
         from .pipeline_wan_i2v import WanImageToVideoPipeline
+        from .pipeline_wan_vace import WanVACEPipeline
         from .pipeline_wan_video2video import WanVideoToVideoPipeline
 
 else:

src/diffusers/pipelines/wan/pipeline_wan_vace.py

@@ -0,0 +1,950 @@
+# 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
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import PIL.Image
+import regex as re
+import torch
+from transformers import AutoTokenizer, UMT5EncoderModel
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PipelineImageInput
+from ...loaders import WanLoraLoaderMixin
+from ...models import AutoencoderKLWan, WanVACETransformer3DModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import WanPipelineOutput
+
+
+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
+
+if is_ftfy_available():
+    import ftfy
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> import PIL.Image
+        >>> from diffusers import AutoencoderKLWan, WanVACEPipeline
+        >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
+        >>> from diffusers.utils import export_to_video, load_image
+        def prepare_video_and_mask(first_img: PIL.Image.Image, last_img: PIL.Image.Image, height: int, width: int, num_frames: int):
+            first_img = first_img.resize((width, height))
+            last_img = last_img.resize((width, height))
+            frames = []
+            frames.append(first_img)
+            # Ideally, this should be 127.5 to match original code, but they perform computation on numpy arrays
+            # whereas we are passing PIL images. If you choose to pass numpy arrays, you can set it to 127.5 to
+            # match the original code.
+            frames.extend([PIL.Image.new("RGB", (width, height), (128, 128, 128))] * (num_frames - 2))
+            frames.append(last_img)
+            mask_black = PIL.Image.new("L", (width, height), 0)
+            mask_white = PIL.Image.new("L", (width, height), 255)
+            mask = [mask_black, *[mask_white] * (num_frames - 2), mask_black]
+            return frames, mask
+
+        >>> # Available checkpoints: Wan-AI/Wan2.1-VACE-1.3B-diffusers, Wan-AI/Wan2.1-VACE-14B-diffusers
+        >>> model_id = "Wan-AI/Wan2.1-VACE-1.3B-diffusers"
+        >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+        >>> pipe = WanVACEPipeline.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)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective."
+        >>> 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"
+        >>> first_frame = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png"
+        ... )
+        >>> last_frame = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png>>> "
+        ... )
+
+        >>> height = 512
+        >>> width = 512
+        >>> num_frames = 81
+        >>> video, mask = prepare_video_and_mask(first_frame, last_frame, height, width, num_frames)
+
+        >>> output = pipe(
+        ...     video=video,
+        ...     mask=mask,
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     height=height,
+        ...     width=width,
+        ...     num_frames=num_frames,
+        ...     num_inference_steps=30,
+        ...     guidance_scale=5.0,
+        ...     generator=torch.Generator().manual_seed(42),
+        ... ).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
+
+
+# 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 WanVACEPipeline(DiffusionPipeline, WanLoraLoaderMixin):
+    r"""
+    Pipeline for controllable 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: WanVACETransformer3DModel,
+        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)
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds
+    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
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.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,
+        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,
+        video=None,
+        mask=None,
+        reference_images=None,
+    ):
+        base = self.vae_scale_factor_spatial * self.transformer.config.patch_size[1]
+        if height % base != 0 or width % base != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by {base} 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)}")
+
+        if video is not None:
+            if mask is not None:
+                if len(video) != len(mask):
+                    raise ValueError(
+                        f"Length of `video` {len(video)} and `mask` {len(mask)} do not match. Please make sure that"
+                        " they have the same length."
+                    )
+            if reference_images is not None:
+                is_pil_image = isinstance(reference_images, PIL.Image.Image)
+                is_list_of_pil_images = isinstance(reference_images, list) and all(
+                    isinstance(ref_img, PIL.Image.Image) for ref_img in reference_images
+                )
+                is_list_of_list_of_pil_images = isinstance(reference_images, list) and all(
+                    isinstance(ref_img, list) and all(isinstance(ref_img_, PIL.Image.Image) for ref_img_ in ref_img)
+                    for ref_img in reference_images
+                )
+                if not (is_pil_image or is_list_of_pil_images or is_list_of_list_of_pil_images):
+                    raise ValueError(
+                        "`reference_images` has to be of type `PIL.Image.Image` or `list` of `PIL.Image.Image`, or "
+                        "`list` of `list` of `PIL.Image.Image`, but is {type(reference_images)}"
+                    )
+                if is_list_of_list_of_pil_images and len(reference_images) != 1:
+                    raise ValueError(
+                        "The pipeline only supports generating one video at a time at the moment. When passing a list "
+                        "of list of reference images, where the outer list corresponds to the batch size and the inner "
+                        "list corresponds to list of conditioning images per video, please make sure to only pass "
+                        "one inner list of reference images (i.e., `[[<image1>, <image2>, ...]]`"
+                    )
+        elif mask is not None:
+            raise ValueError("`mask` can only be passed if `video` is passed as well.")
+
+    def preprocess_conditions(
+        self,
+        video: Optional[List[PipelineImageInput]] = None,
+        mask: Optional[List[PipelineImageInput]] = None,
+        reference_images: Optional[Union[PIL.Image.Image, List[PIL.Image.Image], List[List[PIL.Image.Image]]]] = None,
+        batch_size: int = 1,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+    ):
+        if video is not None:
+            base = self.vae_scale_factor_spatial * self.transformer.config.patch_size[1]
+            video_height, video_width = self.video_processor.get_default_height_width(video[0])
+
+            if video_height * video_width > height * width:
+                scale = min(width / video_width, height / video_height)
+                video_height, video_width = int(video_height * scale), int(video_width * scale)
+
+            if video_height % base != 0 or video_width % base != 0:
+                logger.warning(
+                    f"Video height and width should be divisible by {base}, but got {video_height} and {video_width}. "
+                )
+                video_height = (video_height // base) * base
+                video_width = (video_width // base) * base
+
+            assert video_height * video_width <= height * width
+
+            video = self.video_processor.preprocess_video(video, video_height, video_width)
+            image_size = (video_height, video_width)  # Use the height/width of video (with possible rescaling)
+        else:
+            video = torch.zeros(batch_size, 3, num_frames, height, width, dtype=dtype, device=device)
+            image_size = (height, width)  # Use the height/width provider by user
+
+        if mask is not None:
+            mask = self.video_processor.preprocess_video(mask, image_size[0], image_size[1])
+            mask = torch.clamp((mask + 1) / 2, min=0, max=1)
+        else:
+            mask = torch.ones_like(video)
+
+        video = video.to(dtype=dtype, device=device)
+        mask = mask.to(dtype=dtype, device=device)
+
+        # Make a list of list of images where the outer list corresponds to video batch size and the inner list
+        # corresponds to list of conditioning images per video
+        if reference_images is None or isinstance(reference_images, PIL.Image.Image):
+            reference_images = [[reference_images] for _ in range(video.shape[0])]
+        elif isinstance(reference_images, (list, tuple)) and isinstance(next(iter(reference_images)), PIL.Image.Image):
+            reference_images = [reference_images]
+        elif (
+            isinstance(reference_images, (list, tuple))
+            and isinstance(next(iter(reference_images)), list)
+            and isinstance(next(iter(reference_images[0])), PIL.Image.Image)
+        ):
+            reference_images = reference_images
+        else:
+            raise ValueError(
+                "`reference_images` has to be of type `PIL.Image.Image` or `list` of `PIL.Image.Image`, or "
+                "`list` of `list` of `PIL.Image.Image`, but is {type(reference_images)}"
+            )
+
+        if video.shape[0] != len(reference_images):
+            raise ValueError(
+                f"Batch size of `video` {video.shape[0]} and length of `reference_images` {len(reference_images)} does not match."
+            )
+
+        ref_images_lengths = [len(reference_images_batch) for reference_images_batch in reference_images]
+        if any(l != ref_images_lengths[0] for l in ref_images_lengths):
+            raise ValueError(
+                f"All batches of `reference_images` should have the same length, but got {ref_images_lengths}. Support for this "
+                "may be added in the future."
+            )
+
+        reference_images_preprocessed = []
+        for i, reference_images_batch in enumerate(reference_images):
+            preprocessed_images = []
+            for j, image in enumerate(reference_images_batch):
+                if image is None:
+                    continue
+                image = self.video_processor.preprocess(image, None, None)
+                img_height, img_width = image.shape[-2:]
+                scale = min(image_size[0] / img_height, image_size[1] / img_width)
+                new_height, new_width = int(img_height * scale), int(img_width * scale)
+                resized_image = torch.nn.functional.interpolate(
+                    image, size=(new_height, new_width), mode="bilinear", align_corners=False
+                ).squeeze(0)  # [C, H, W]
+                top = (image_size[0] - new_height) // 2
+                left = (image_size[1] - new_width) // 2
+                canvas = torch.ones(3, *image_size, device=device, dtype=dtype)
+                canvas[:, top : top + new_height, left : left + new_width] = resized_image
+                preprocessed_images.append(canvas)
+            reference_images_preprocessed.append(preprocessed_images)
+
+        return video, mask, reference_images_preprocessed
+
+    def prepare_video_latents(
+        self,
+        video: torch.Tensor,
+        mask: torch.Tensor,
+        reference_images: Optional[List[List[torch.Tensor]]] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        device: Optional[torch.device] = None,
+    ) -> torch.Tensor:
+        device = device or self._execution_device
+
+        if isinstance(generator, list):
+            # TODO: support this
+            raise ValueError("Passing a list of generators is not yet supported. This may be supported in the future.")
+
+        if reference_images is None:
+            # For each batch of video, we set no re
+            # ference image (as one or more can be passed by user)
+            reference_images = [[None] for _ in range(video.shape[0])]
+        else:
+            if video.shape[0] != len(reference_images):
+                raise ValueError(
+                    f"Batch size of `video` {video.shape[0]} and length of `reference_images` {len(reference_images)} does not match."
+                )
+
+        if video.shape[0] != 1:
+            # TODO: support this
+            raise ValueError(
+                "Generating with more than one video is not yet supported. This may be supported in the future."
+            )
+
+        vae_dtype = self.vae.dtype
+        video = video.to(dtype=vae_dtype)
+
+        latents_mean = torch.tensor(self.vae.config.latents_mean, device=device, dtype=torch.float32).view(
+            1, self.vae.config.z_dim, 1, 1, 1
+        )
+        latents_std = 1.0 / torch.tensor(self.vae.config.latents_std, device=device, dtype=torch.float32).view(
+            1, self.vae.config.z_dim, 1, 1, 1
+        )
+
+        if mask is None:
+            latents = retrieve_latents(self.vae.encode(video), generator, sample_mode="argmax").unbind(0)
+            latents = ((latents.float() - latents_mean) * latents_std).to(vae_dtype)
+        else:
+            mask = mask.to(dtype=vae_dtype)
+            mask = torch.where(mask > 0.5, 1.0, 0.0)
+            inactive = video * (1 - mask)
+            reactive = video * mask
+            inactive = retrieve_latents(self.vae.encode(inactive), generator, sample_mode="argmax")
+            reactive = retrieve_latents(self.vae.encode(reactive), generator, sample_mode="argmax")
+            inactive = ((inactive.float() - latents_mean) * latents_std).to(vae_dtype)
+            reactive = ((reactive.float() - latents_mean) * latents_std).to(vae_dtype)
+            latents = torch.cat([inactive, reactive], dim=1)
+
+        latent_list = []
+        for latent, reference_images_batch in zip(latents, reference_images):
+            for reference_image in reference_images_batch:
+                assert reference_image.ndim == 3
+                reference_image = reference_image.to(dtype=vae_dtype)
+                reference_image = reference_image[None, :, None, :, :]  # [1, C, 1, H, W]
+                reference_latent = retrieve_latents(self.vae.encode(reference_image), generator, sample_mode="argmax")
+                reference_latent = ((reference_latent.float() - latents_mean) * latents_std).to(vae_dtype)
+                reference_latent = reference_latent.squeeze(0)  # [C, 1, H, W]
+                reference_latent = torch.cat([reference_latent, torch.zeros_like(reference_latent)], dim=0)
+                latent = torch.cat([reference_latent.squeeze(0), latent], dim=1)
+            latent_list.append(latent)
+        return torch.stack(latent_list)
+
+    def prepare_masks(
+        self,
+        mask: torch.Tensor,
+        reference_images: Optional[List[torch.Tensor]] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+    ) -> torch.Tensor:
+        if isinstance(generator, list):
+            # TODO: support this
+            raise ValueError("Passing a list of generators is not yet supported. This may be supported in the future.")
+
+        if reference_images is None:
+            # For each batch of video, we set no reference image (as one or more can be passed by user)
+            reference_images = [[None] for _ in range(mask.shape[0])]
+        else:
+            if mask.shape[0] != len(reference_images):
+                raise ValueError(
+                    f"Batch size of `mask` {mask.shape[0]} and length of `reference_images` {len(reference_images)} does not match."
+                )
+
+        if mask.shape[0] != 1:
+            # TODO: support this
+            raise ValueError(
+                "Generating with more than one video is not yet supported. This may be supported in the future."
+            )
+
+        transformer_patch_size = self.transformer.config.patch_size[1]
+
+        mask_list = []
+        for mask_, reference_images_batch in zip(mask, reference_images):
+            num_channels, num_frames, height, width = mask_.shape
+            new_num_frames = (num_frames + self.vae_scale_factor_temporal - 1) // self.vae_scale_factor_temporal
+            new_height = height // (self.vae_scale_factor_spatial * transformer_patch_size) * transformer_patch_size
+            new_width = width // (self.vae_scale_factor_spatial * transformer_patch_size) * transformer_patch_size
+            mask_ = mask_[0, :, :, :]
+            mask_ = mask_.view(
+                num_frames, new_height, self.vae_scale_factor_spatial, new_width, self.vae_scale_factor_spatial
+            )
+            mask_ = mask_.permute(2, 4, 0, 1, 3).flatten(0, 1)  # [8x8, num_frames, new_height, new_width]
+            mask_ = torch.nn.functional.interpolate(
+                mask_.unsqueeze(0), size=(new_num_frames, new_height, new_width), mode="nearest-exact"
+            ).squeeze(0)
+            num_ref_images = len(reference_images_batch)
+            if num_ref_images > 0:
+                mask_padding = torch.zeros_like(mask_[:, :num_ref_images, :, :])
+                mask_ = torch.cat([mask_, mask_padding], dim=1)
+            mask_list.append(mask_)
+        return torch.stack(mask_list)
+
+    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 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,
+        video: Optional[List[PipelineImageInput]] = None,
+        mask: Optional[List[PipelineImageInput]] = None,
+        reference_images: Optional[List[PipelineImageInput]] = None,
+        conditioning_scale: Union[float, List[float], torch.Tensor] = 1.0,
+        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,
+    ):
+        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 `"np"`):
+                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
+
+        # Simplification of implementation for now
+        if not isinstance(prompt, str):
+            raise ValueError("Passing a list of prompts is not yet supported. This may be supported in the future.")
+        if num_videos_per_prompt != 1:
+            raise ValueError(
+                "Generating multiple videos per prompt is not yet supported. This may be supported in the future."
+            )
+
+        # 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,
+            video,
+            mask,
+            reference_images,
+        )
+
+        if num_frames % self.vae_scale_factor_temporal != 1:
+            logger.warning(
+                f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number."
+            )
+            num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1
+        num_frames = max(num_frames, 1)
+
+        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]
+
+        vae_dtype = self.vae.dtype
+        transformer_dtype = self.transformer.dtype
+
+        if isinstance(conditioning_scale, (int, float)):
+            conditioning_scale = [conditioning_scale] * len(self.transformer.config.vace_layers)
+        if isinstance(conditioning_scale, list):
+            if len(conditioning_scale) != len(self.transformer.config.vace_layers):
+                raise ValueError(
+                    f"Length of `conditioning_scale` {len(conditioning_scale)} does not match number of layers {len(self.transformer.config.vace_layers)}."
+                )
+            conditioning_scale = torch.tensor(conditioning_scale)
+        if isinstance(conditioning_scale, torch.Tensor):
+            if conditioning_scale.size(0) != len(self.transformer.config.vace_layers):
+                raise ValueError(
+                    f"Length of `conditioning_scale` {conditioning_scale.size(0)} does not match number of layers {len(self.transformer.config.vace_layers)}."
+                )
+            conditioning_scale = conditioning_scale.to(device=device, dtype=transformer_dtype)
+
+        # 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,
+        )
+
+        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
+        video, mask, reference_images = self.preprocess_conditions(
+            video,
+            mask,
+            reference_images,
+            batch_size,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+        )
+        num_reference_images = len(reference_images[0])
+
+        conditioning_latents = self.prepare_video_latents(video, mask, reference_images, generator, device)
+        mask = self.prepare_masks(mask, reference_images, generator)
+        conditioning_latents = torch.cat([conditioning_latents, mask], dim=1)
+        conditioning_latents = conditioning_latents.to(transformer_dtype)
+
+        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 + num_reference_images * self.vae_scale_factor_temporal,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        if conditioning_latents.shape[2] != latents.shape[2]:
+            logger.warning(
+                "The number of frames in the conditioning latents does not match the number of frames to be generated. Generation quality may be affected."
+            )
+
+        # 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])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    control_hidden_states=conditioning_latents,
+                    control_hidden_states_scale=conditioning_scale,
+                    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,
+                        control_hidden_states=conditioning_latents,
+                        control_hidden_states_scale=conditioning_scale,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    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[:, :, num_reference_images:]
+            latents = latents.to(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)

src/diffusers/pipelines/wuerstchen/pipeline_wuerstchen.py

@@ -21,7 +21,7 @@ from transformers import CLIPTextModel, CLIPTokenizer
 from ...schedulers import DDPMWuerstchenScheduler
 from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput
 from .modeling_paella_vq_model import PaellaVQModel
 from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt
 
@@ -56,7 +56,7 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class WuerstchenDecoderPipeline(DiffusionPipeline):
+class WuerstchenDecoderPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
     """
     Pipeline for generating images from the Wuerstchen model.