Merge branch 'main' into folderize-loaders

consistency
[Hi-Dream LoRA] fix bug in validation (#11439 )
2025-04-29 00:28:06 +08:00 · 2025-04-29 00:26:28 +08:00 · 2025-04-28 06:22:32 -10:00 · 2025-04-29 00:11:52 +08:00 · 2025-04-29 00:04:19 +08:00 · 2025-04-29 00:02:38 +08:00
451 changed files with 38995 additions and 17201 deletions
@@ -180,6 +180,55 @@ jobs:
        pip install slack_sdk tabulate
        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY

+  run_torch_compile_tests:
+    name: PyTorch Compile CUDA tests
+
+    runs-on:
+      group: aws-g4dn-2xlarge
+
+    container:
+      image: diffusers/diffusers-pytorch-compile-cuda
+      options: --gpus 0 --shm-size "16gb" --ipc host
+
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: NVIDIA-SMI
+      run: |
+        nvidia-smi
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test,training]
+    - name: Environment
+      run: |
+        python utils/print_env.py
+    - name: Run torch compile tests on GPU
+      env:
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
+        RUN_COMPILE: yes
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -k "compile" --make-reports=tests_torch_compile_cuda tests/
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: cat reports/tests_torch_compile_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v4
+      with:
+        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:
@@ -417,7 +466,7 @@ jobs:
            additional_deps: ["peft"]
          - backend: "gguf"
            test_location: "gguf"
-            additional_deps: []
+            additional_deps: ["peft"]
          - backend: "torchao"
            test_location: "torchao"
            additional_deps: []
@@ -13,39 +13,5 @@ jobs:
    uses: huggingface/huggingface_hub/.github/workflows/style-bot-action.yml@main
    with:
      python_quality_dependencies: "[quality]"
-      pre_commit_script_name: "Download and Compare files from the main branch"
-      pre_commit_script: |
-        echo "Downloading the files from the main branch"
-
-        curl -o main_Makefile https://raw.githubusercontent.com/huggingface/diffusers/main/Makefile
-        curl -o main_setup.py https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/setup.py
-        curl -o main_check_doc_toc.py https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/utils/check_doc_toc.py
-
-        echo "Compare the files and raise error if needed"
-
-        diff_failed=0
-        if ! diff -q main_Makefile Makefile; then
-          echo "Error: The Makefile has changed. Please ensure it matches the main branch."
-          diff_failed=1
-        fi
-
-        if ! diff -q main_setup.py setup.py; then
-          echo "Error: The setup.py has changed. Please ensure it matches the main branch."
-          diff_failed=1
-        fi
-
-        if ! diff -q main_check_doc_toc.py utils/check_doc_toc.py; then
-          echo "Error: The utils/check_doc_toc.py has changed. Please ensure it matches the main branch."
-          diff_failed=1
-        fi
-
-        if [ $diff_failed -eq 1 ]; then
-          echo "❌ Error happened as we detected changes in the files that should not be changed ❌"
-          exit 1
-        fi
-
-        echo "No changes in the files. Proceeding..."
-        rm -rf main_Makefile main_setup.py main_check_doc_toc.py
-      style_command: "make style && make quality"
    secrets:
      bot_token: ${{ secrets.GITHUB_TOKEN }}
@@ -335,7 +335,7 @@ jobs:
    - name: Environment
      run: |
        python utils/print_env.py
-    - name: Run example tests on GPU
+    - name: Run torch compile tests on GPU
      env:
        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
        RUN_COMPILE: yes
@@ -28,9 +28,9 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
    python3 -m uv pip install --no-cache-dir \
-        torch==2.1.2 \
-        torchvision==0.16.2 \
-        torchaudio==2.1.2 \
+        torch \
+        torchvision \
+        torchaudio\
        onnxruntime \
        --extra-index-url https://download.pytorch.org/whl/cpu && \
    python3 -m uv pip install --no-cache-dir \
@@ -175,7 +175,7 @@
    title: gguf
  - local: quantization/torchao
    title: torchao
-  - local: quantization/quanto 
+  - local: quantization/quanto
    title: quanto
  title: Quantization Methods
 - sections:
@@ -265,19 +265,23 @@
    sections:
    - local: api/models/overview
      title: Overview
+    - local: api/models/auto_model
+      title: AutoModel
    - sections:
      - local: api/models/controlnet
        title: ControlNetModel
+      - local: api/models/controlnet_union
+        title: ControlNetUnionModel
      - local: api/models/controlnet_flux
        title: FluxControlNetModel
      - local: api/models/controlnet_hunyuandit
        title: HunyuanDiT2DControlNetModel
+      - local: api/models/controlnet_sana
+        title: SanaControlNetModel
      - local: api/models/controlnet_sd3
        title: SD3ControlNetModel
      - local: api/models/controlnet_sparsectrl
        title: SparseControlNetModel
-      - local: api/models/controlnet_union
-        title: ControlNetUnionModel
      title: ControlNets
    - sections:
      - local: api/models/allegro_transformer3d
@@ -286,30 +290,32 @@
        title: AuraFlowTransformer2DModel
      - local: api/models/cogvideox_transformer3d
        title: CogVideoXTransformer3DModel
-      - local: api/models/consisid_transformer3d
-        title: ConsisIDTransformer3DModel
      - local: api/models/cogview3plus_transformer2d
        title: CogView3PlusTransformer2DModel
      - local: api/models/cogview4_transformer2d
        title: CogView4Transformer2DModel
+      - local: api/models/consisid_transformer3d
+        title: ConsisIDTransformer3DModel
      - local: api/models/dit_transformer2d
        title: DiTTransformer2DModel
      - local: api/models/easyanimate_transformer3d
        title: EasyAnimateTransformer3DModel
      - local: api/models/flux_transformer
        title: FluxTransformer2DModel
+      - local: api/models/hidream_image_transformer
+        title: HiDreamImageTransformer2DModel
      - local: api/models/hunyuan_transformer2d
        title: HunyuanDiT2DModel
      - local: api/models/hunyuan_video_transformer_3d
        title: HunyuanVideoTransformer3DModel
      - local: api/models/latte_transformer3d
        title: LatteTransformer3DModel
-      - local: api/models/lumina_nextdit2d
-        title: LuminaNextDiT2DModel
-      - local: api/models/lumina2_transformer2d
-        title: Lumina2Transformer2DModel
      - local: api/models/ltx_video_transformer3d
        title: LTXVideoTransformer3DModel
+      - local: api/models/lumina2_transformer2d
+        title: Lumina2Transformer2DModel
+      - local: api/models/lumina_nextdit2d
+        title: LuminaNextDiT2DModel
      - local: api/models/mochi_transformer3d
        title: MochiTransformer3DModel
      - local: api/models/omnigen_transformer
@@ -318,10 +324,10 @@
        title: PixArtTransformer2DModel
      - local: api/models/prior_transformer
        title: PriorTransformer
-      - local: api/models/sd3_transformer2d
-        title: SD3Transformer2DModel
      - local: api/models/sana_transformer2d
        title: SanaTransformer2DModel
+      - local: api/models/sd3_transformer2d
+        title: SD3Transformer2DModel
      - local: api/models/stable_audio_transformer
        title: StableAudioDiTModel
      - local: api/models/transformer2d
@@ -336,10 +342,10 @@
        title: StableCascadeUNet
      - local: api/models/unet
        title: UNet1DModel
-      - local: api/models/unet2d
-        title: UNet2DModel
      - local: api/models/unet2d-cond
        title: UNet2DConditionModel
+      - local: api/models/unet2d
+        title: UNet2DModel
      - local: api/models/unet3d-cond
        title: UNet3DConditionModel
      - local: api/models/unet-motion
@@ -348,6 +354,10 @@
        title: UViT2DModel
      title: UNets
    - sections:
+      - local: api/models/asymmetricautoencoderkl
+        title: AsymmetricAutoencoderKL
+      - local: api/models/autoencoder_dc
+        title: AutoencoderDC
      - local: api/models/autoencoderkl
        title: AutoencoderKL
      - local: api/models/autoencoderkl_allegro
@@ -364,10 +374,6 @@
        title: AutoencoderKLMochi
      - local: api/models/autoencoder_kl_wan
        title: AutoencoderKLWan
-      - local: api/models/asymmetricautoencoderkl
-        title: AsymmetricAutoencoderKL
-      - local: api/models/autoencoder_dc
-        title: AutoencoderDC
      - local: api/models/consistency_decoder_vae
        title: ConsistencyDecoderVAE
      - local: api/models/autoencoder_oobleck
@@ -420,6 +426,8 @@
      title: ControlNet with Stable Diffusion 3
    - local: api/pipelines/controlnet_sdxl
      title: ControlNet with Stable Diffusion XL
+    - local: api/pipelines/controlnet_sana
+      title: ControlNet-Sana
    - local: api/pipelines/controlnetxs
      title: ControlNet-XS
    - local: api/pipelines/controlnetxs_sdxl
@@ -444,6 +452,8 @@
      title: Flux
    - local: api/pipelines/control_flux_inpaint
      title: FluxControlInpaint
+    - local: api/pipelines/hidream
+      title: HiDream-I1
    - local: api/pipelines/hunyuandit
      title: Hunyuan-DiT
    - local: api/pipelines/hunyuan_video
@@ -511,40 +521,40 @@
    - sections:
      - local: api/pipelines/stable_diffusion/overview
        title: Overview
-      - local: api/pipelines/stable_diffusion/text2img
-        title: Text-to-image
+      - local: api/pipelines/stable_diffusion/depth2img
+        title: Depth-to-image
+      - local: api/pipelines/stable_diffusion/gligen
+        title: GLIGEN (Grounded Language-to-Image Generation)
+      - local: api/pipelines/stable_diffusion/image_variation
+        title: Image variation
      - local: api/pipelines/stable_diffusion/img2img
        title: Image-to-image
      - local: api/pipelines/stable_diffusion/svd
        title: Image-to-video
      - local: api/pipelines/stable_diffusion/inpaint
        title: Inpainting
-      - local: api/pipelines/stable_diffusion/depth2img
-        title: Depth-to-image
-      - local: api/pipelines/stable_diffusion/image_variation
-        title: Image variation
+      - local: api/pipelines/stable_diffusion/k_diffusion
+        title: K-Diffusion
+      - local: api/pipelines/stable_diffusion/latent_upscale
+        title: Latent upscaler
+      - local: api/pipelines/stable_diffusion/ldm3d_diffusion
+        title: LDM3D Text-to-(RGB, Depth), Text-to-(RGB-pano, Depth-pano), LDM3D Upscaler
      - local: api/pipelines/stable_diffusion/stable_diffusion_safe
        title: Safe Stable Diffusion
+      - local: api/pipelines/stable_diffusion/sdxl_turbo
+        title: SDXL Turbo
      - local: api/pipelines/stable_diffusion/stable_diffusion_2
        title: Stable Diffusion 2
      - local: api/pipelines/stable_diffusion/stable_diffusion_3
        title: Stable Diffusion 3
      - local: api/pipelines/stable_diffusion/stable_diffusion_xl
        title: Stable Diffusion XL
-      - local: api/pipelines/stable_diffusion/sdxl_turbo
-        title: SDXL Turbo
-      - local: api/pipelines/stable_diffusion/latent_upscale
-        title: Latent upscaler
      - local: api/pipelines/stable_diffusion/upscale
        title: Super-resolution
-      - local: api/pipelines/stable_diffusion/k_diffusion
-        title: K-Diffusion
-      - local: api/pipelines/stable_diffusion/ldm3d_diffusion
-        title: LDM3D Text-to-(RGB, Depth), Text-to-(RGB-pano, Depth-pano), LDM3D Upscaler
      - local: api/pipelines/stable_diffusion/adapter
        title: T2I-Adapter
-      - local: api/pipelines/stable_diffusion/gligen
-        title: GLIGEN (Grounded Language-to-Image Generation)
+      - local: api/pipelines/stable_diffusion/text2img
+        title: Text-to-image
      title: Stable Diffusion
    - local: api/pipelines/stable_unclip
      title: Stable unCLIP
@@ -22,11 +22,11 @@ Learn how to load an IP-Adapter checkpoint and image in the IP-Adapter [loading]

 ## IPAdapterMixin

-[[autodoc]] loaders.ip_adapter.IPAdapterMixin
+[[autodoc]] loaders.ip_adapter.ip_adapter.IPAdapterMixin

 ## SD3IPAdapterMixin

-[[autodoc]] loaders.ip_adapter.SD3IPAdapterMixin
+[[autodoc]] loaders.ip_adapter.ip_adapter.SD3IPAdapterMixin
    - all
    - is_ip_adapter_active

@@ -20,11 +20,15 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 - [`FluxLoraLoaderMixin`] provides similar functions for [Flux](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux).
 - [`CogVideoXLoraLoaderMixin`] provides similar functions for [CogVideoX](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox).
 - [`Mochi1LoraLoaderMixin`] provides similar functions for [Mochi](https://huggingface.co/docs/diffusers/main/en/api/pipelines/mochi).
+- [`AuraFlowLoraLoaderMixin`] provides similar functions for [AuraFlow](https://huggingface.co/fal/AuraFlow).
 - [`LTXVideoLoraLoaderMixin`] provides similar functions for [LTX-Video](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ltx_video).
 - [`SanaLoraLoaderMixin`] provides similar functions for [Sana](https://huggingface.co/docs/diffusers/main/en/api/pipelines/sana).
 - [`HunyuanVideoLoraLoaderMixin`] provides similar functions for [HunyuanVideo](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hunyuan_video).
 - [`Lumina2LoraLoaderMixin`] provides similar functions for [Lumina2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/lumina2).
+- [`WanLoraLoaderMixin`] provides similar functions for [Wan](https://huggingface.co/docs/diffusers/main/en/api/pipelines/wan).
+- [`CogView4LoraLoaderMixin`] provides similar functions for [CogView4](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogview4).
 - [`AmusedLoraLoaderMixin`] is for the [`AmusedPipeline`].
+- [`HiDreamImageLoraLoaderMixin`] provides similar functions for [HiDream Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hidream)
 - [`LoraBaseMixin`] provides a base class with several utility methods to fuse, unfuse, unload, LoRAs and more.

 <Tip>
@@ -35,48 +39,71 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse

 ## StableDiffusionLoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.StableDiffusionLoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.StableDiffusionLoraLoaderMixin

 ## StableDiffusionXLLoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.StableDiffusionXLLoraLoaderMixin

 ## SD3LoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.SD3LoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.SD3LoraLoaderMixin

 ## FluxLoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.FluxLoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.FluxLoraLoaderMixin

 ## CogVideoXLoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.CogVideoXLoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.CogVideoXLoraLoaderMixin

 ## Mochi1LoraLoaderMixin

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

 ## LTXVideoLoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.LTXVideoLoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.LTXVideoLoraLoaderMixin

 ## SanaLoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.SanaLoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.SanaLoraLoaderMixin

 ## HunyuanVideoLoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.HunyuanVideoLoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.HunyuanVideoLoraLoaderMixin

 ## Lumina2LoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.Lumina2LoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.Lumina2LoraLoaderMixin
+
+## WanLoraLoaderMixin
+
+[[autodoc]] loaders.lora.lora_pipeline.WanLoraLoaderMixin
+
+## CogView4LoraLoaderMixin
+
+[[autodoc]] loaders.lora.lora_pipeline.CogView4LoraLoaderMixin
+
+## CogView4LoraLoaderMixin
+
+[[autodoc]] loaders.lora.lora_pipeline.CogView4LoraLoaderMixin
+
+## WanLoraLoaderMixin
+
+[[autodoc]] loaders.lora.lora_pipeline.WanLoraLoaderMixin

 ## AmusedLoraLoaderMixin

-[[autodoc]] loaders.lora_pipeline.AmusedLoraLoaderMixin
+[[autodoc]] loaders.lora.lora_pipeline.AmusedLoraLoaderMixin
+
+## HiDreamImageLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.HiDreamImageLoraLoaderMixin

 ## LoraBaseMixin

-[[autodoc]] loaders.lora_base.LoraBaseMixin
+[[autodoc]] loaders.lora.lora_base.LoraBaseMixin
@@ -12,7 +12,7 @@ specific language governing permissions and limitations under the License.

 # SD3Transformer2D

-This class is useful when *only* loading weights into a [`SD3Transformer2DModel`]. If you need to load weights into the text encoder or a text encoder and SD3Transformer2DModel, check [`SD3LoraLoaderMixin`](lora#diffusers.loaders.SD3LoraLoaderMixin) class instead.
+This class is useful when *only* loading weights into a [`SD3Transformer2DModel`]. If you need to load weights into the text encoder or a text encoder and [SD3Transformer2DModel], check [`SD3LoraLoaderMixin`](lora#diffusers.loaders.SD3LoraLoaderMixin) class instead.

 The [`SD3Transformer2DLoadersMixin`] class currently only loads IP-Adapter weights, but will be used in the future to save weights and load LoRAs.

@@ -24,6 +24,6 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse

 ## SD3Transformer2DLoadersMixin

-[[autodoc]] loaders.transformer_sd3.SD3Transformer2DLoadersMixin
+[[autodoc]] loaders.ip_adapter.transformer_sd3.SD3Transformer2DLoadersMixin
    - all
    - _load_ip_adapter_weights
@@ -0,0 +1,29 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# AutoModel
+
+The `AutoModel` is designed to make it easy to load a checkpoint without needing to know the specific model class. `AutoModel` automatically retrieves the correct model class from the checkpoint `config.json` file.
+
+```python
+from diffusers import AutoModel, AutoPipelineForText2Image
+
+unet = AutoModel.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", subfolder="unet")
+pipe = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", unet=unet)
+```
+
+
+## AutoModel
+
+[[autodoc]] AutoModel
+	- all
+	- from_pretrained
@@ -18,7 +18,7 @@ The model can be loaded with the following code snippet.
 ```python
 from diffusers import AutoencoderKLAllegro

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

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

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

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

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

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

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

 Flux is a series of text-to-image generation models based on diffusion transformers. To know more about Flux, check out the original [blog post](https://blackforestlabs.ai/announcing-black-forest-labs/) by the creators of Flux, Black Forest Labs.
@@ -346,7 +347,7 @@ image = pipe(
    height=1024,
    prompt="wearing sunglasses",
    negative_prompt="",
-    true_cfg=4.0,
+    true_cfg_scale=4.0,
    generator=torch.Generator().manual_seed(4444),
    ip_adapter_image=image,
 ).images[0]
@@ -0,0 +1,43 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+# HiDreamImage
+
+[HiDream-I1](https://huggingface.co/HiDream-ai) by HiDream.ai
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+## Available models
+
+The following models are available for the [`HiDreamImagePipeline`](text-to-image) pipeline:
+
+| Model name | Description |
+|:---|:---|
+| [`HiDream-ai/HiDream-I1-Full`](https://huggingface.co/HiDream-ai/HiDream-I1-Full) | - |
+| [`HiDream-ai/HiDream-I1-Dev`](https://huggingface.co/HiDream-ai/HiDream-I1-Dev) | - |
+| [`HiDream-ai/HiDream-I1-Fast`](https://huggingface.co/HiDream-ai/HiDream-I1-Fast) | - |
+
+## HiDreamImagePipeline
+
+[[autodoc]] HiDreamImagePipeline
+  - all
+  - __call__
+
+## HiDreamImagePipelineOutput
+
+[[autodoc]] pipelines.hidream_image.pipeline_output.HiDreamImagePipelineOutput
@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.

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

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

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

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

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

@@ -16,6 +16,7 @@

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

 [SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License. -->

-# SanaSprintPipeline
+# SANA-Sprint

 <div class="flex flex-wrap space-x-1">
  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.

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

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

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

 Stable Diffusion XL (SDXL) was proposed in [SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis](https://huggingface.co/papers/2307.01952) by Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach.
@@ -24,7 +24,7 @@

 ## Generating Videos with Wan 2.1

-We will first need to install some addtional dependencies.
+We will first need to install some additional dependencies.

 ```shell
 pip install -u ftfy imageio-ffmpeg imageio
@@ -133,6 +133,60 @@ output = pipe(
 export_to_video(output, "wan-i2v.mp4", fps=16)
 ```

+### First and Last Frame Interpolation
+
+```python
+import numpy as np
+import torch
+import torchvision.transforms.functional as TF
+from diffusers import AutoencoderKLWan, WanImageToVideoPipeline
+from diffusers.utils import export_to_video, load_image
+from transformers import CLIPVisionModel
+
+
+model_id = "Wan-AI/Wan2.1-FLF2V-14B-720P-diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(model_id, subfolder="image_encoder", torch_dtype=torch.float32)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id, vae=vae, image_encoder=image_encoder, torch_dtype=torch.bfloat16
+)
+pipe.to("cuda")
+
+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")
+
+def aspect_ratio_resize(image, pipe, max_area=720 * 1280):
+    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))
+    return image, height, width
+
+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)
+if last_frame.size != first_frame.size:
+    last_frame, _, _ = center_crop_resize(last_frame, height, width)
+
+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."
+
+output = pipe(
+    image=first_frame, last_image=last_frame, prompt=prompt, height=height, width=width, guidance_scale=5.5
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
+```
+
 ### Video to Video Generation

 ```python
@@ -83,4 +83,8 @@ Happy exploring, and thank you for being part of the Diffusers community!
    <td><a href="https://github.com/suzukimain/auto_diffusers"> Model Search </a></td>
    <td>Search models on Civitai and Hugging Face</td>
  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/beinsezii/skrample"> Skrample </a></td>
+    <td>Fully modular scheduler functions with 1st class diffusers integration.</td>
+  </tr>
 </table>
@@ -178,6 +178,9 @@ pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch
 # We can utilize the enable_group_offload method for Diffusers model implementations
 pipe.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True)

+# Uncomment the following to also allow recording the current streams.
+# pipe.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True, record_stream=True)
+
 # For any other model implementations, the apply_group_offloading function can be used
 apply_group_offloading(pipe.text_encoder, onload_device=onload_device, offload_type="block_level", num_blocks_per_group=2)
 apply_group_offloading(pipe.vae, onload_device=onload_device, offload_type="leaf_level")
@@ -205,6 +208,7 @@ Group offloading (for CUDA devices with support for asynchronous data transfer s
 - The `use_stream` parameter can be used with CUDA devices to enable prefetching layers for onload. It defaults to `False`. Layer prefetching allows overlapping computation and data transfer of model weights, which drastically reduces the overall execution time compared to other offloading methods. However, it can increase the CPU RAM usage significantly. Ensure that available CPU RAM that is at least twice the size of the model when setting `use_stream=True`. You can find more information about CUDA streams [here](https://pytorch.org/docs/stable/generated/torch.cuda.Stream.html)
 - If specifying `use_stream=True` on VAEs with tiling enabled, make sure to do a dummy forward pass (possibly with dummy inputs) before the actual inference to avoid device-mismatch errors. This may not work on all implementations. Please open an issue if you encounter any problems.
 - The parameter `low_cpu_mem_usage` can be set to `True` to reduce CPU memory usage when using streams for group offloading. This is useful when the CPU memory is the bottleneck, but it may counteract the benefits of using streams and increase the overall execution time. The CPU memory savings come from creating pinned-tensors on-the-fly instead of pre-pinning them. This parameter is better suited for using `leaf_level` offloading.
+- When using `use_stream=True`, users can additionally specify `record_stream=True` to get better speedups at the expense of slightly increased memory usage. Refer to the [official PyTorch docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html) to know more about this.

 For more information about available parameters and an explanation of how group offloading works, refer to [`~hooks.group_offloading.apply_group_offloading`].

@@ -12,6 +12,9 @@ specific language governing permissions and limitations under the License.

 # Metal Performance Shaders (MPS)

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

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

 <Tip warning={true}>

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

 </Tip>

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

 ## Troubleshoot

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

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

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

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

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

 quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True,)

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

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

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

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

 quant_config = DiffusersBitsAndBytesConfig(load_in_4bit=True,)

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

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

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

 quantization_config = BitsAndBytesConfig(load_in_4bit=True)

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

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

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

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

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

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

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

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

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

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

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

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

-transformer_4bit = FluxTransformer2DModel.from_pretrained(
+transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
@@ -26,13 +26,13 @@ The example below only quantizes the weights to int8.

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

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

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

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

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

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

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

@@ -131,10 +131,10 @@ If you are using `torch<=2.6.0`, some quantization methods, such as `uint4wo`, c
 ```python
 import torch
 from accelerate import init_empty_weights
-from diffusers import FluxPipeline, FluxTransformer2DModel, TorchAoConfig
+from diffusers import FluxPipeline, AutoModel, TorchAoConfig

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

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

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

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

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

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

 model_id = "stable-diffusion-v1-5/stable-diffusion-v1-5"
-unet = UNet2DConditionModel.from_pretrained(
+unet = AutoModel.from_pretrained(
    model_id,
    subfolder="unet",
    in_channels=9,
@@ -216,7 +216,7 @@ Setting the `<ID_TOKEN>` is not necessary. From some limited experimentation, we
 > - The original repository uses a `lora_alpha` of `1`. We found this not suitable in many runs, possibly due to difference in modeling backends and training settings. Our recommendation is to set to the `lora_alpha` to either `rank` or `rank // 2`.
 > - If you're training on data whose captions generate bad results with the original model, a `rank` of 64 and above is good and also the recommendation by the team behind CogVideoX. If the generations are already moderately good on your training captions, a `rank` of 16/32 should work. We found that setting the rank too low, say `4`, is not ideal and doesn't produce promising results.
 > - The authors of CogVideoX recommend 4000 training steps and 100 training videos overall to achieve the best result. While that might yield the best results, we found from our limited experimentation that 2000 steps and 25 videos could also be sufficient.
-> - When using the Prodigy opitimizer for training, one can follow the recommendations from [this](https://huggingface.co/blog/sdxl_lora_advanced_script) blog. Prodigy tends to overfit quickly. From my very limited testing, I found a learning rate of `0.5` to be suitable in addition to `--prodigy_use_bias_correction`, `prodigy_safeguard_warmup` and `--prodigy_decouple`.
+> - When using the Prodigy optimizer for training, one can follow the recommendations from [this](https://huggingface.co/blog/sdxl_lora_advanced_script) blog. Prodigy tends to overfit quickly. From my very limited testing, I found a learning rate of `0.5` to be suitable in addition to `--prodigy_use_bias_correction`, `prodigy_safeguard_warmup` and `--prodigy_decouple`.
 > - The recommended learning rate by the CogVideoX authors and from our experimentation with Adam/AdamW is between `1e-3` and `1e-4` for a dataset of 25+ videos.
 >
 > Note that our testing is not exhaustive due to limited time for exploration. Our recommendation would be to play around with the different knobs and dials to find the best settings for your data.
@@ -165,10 +165,10 @@ flush()
 Load the diffusion transformer next which has 12.5B parameters. This time, set `device_map="auto"` to automatically distribute the model across two 16GB GPUs. The `auto` strategy is backed by [Accelerate](https://hf.co/docs/accelerate/index) and available as a part of the [Big Model Inference](https://hf.co/docs/accelerate/concept_guides/big_model_inference) feature. It starts by distributing a model across the fastest device first (GPU) before moving to slower devices like the CPU and hard drive if needed. The trade-off of storing model parameters on slower devices is slower inference latency.

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

-transformer = FluxTransformer2DModel.from_pretrained(
+transformer = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev", 
    subfolder="transformer",
    device_map="auto",
@@ -589,7 +589,7 @@ For stage 2 of DeepFloyd IF with DreamBooth, pay attention to these parameters:

 * `--learning_rate=5e-6`, use a lower learning rate with a smaller effective batch size
 * `--resolution=256`, the expected resolution for the upscaler
-* `--train_batch_size=2` and `--gradient_accumulation_steps=6`, to effectively train on images wiht faces requires larger batch sizes
+* `--train_batch_size=2` and `--gradient_accumulation_steps=6`, to effectively train on images with faces requires larger batch sizes

 ```bash
 export MODEL_NAME="DeepFloyd/IF-II-L-v1.0"
@@ -89,7 +89,7 @@ Many of the basic and important parameters are described in the [Text-to-image](

 As with the script parameters, a walkthrough of the training script is provided in the [Text-to-image](text2image#training-script) training guide. Instead, this guide takes a look at the T2I-Adapter relevant parts of the script.

-The training script begins by preparing the dataset. This incudes [tokenizing](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L674) the prompt and [applying transforms](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L714) to the images and conditioning images.
+The training script begins by preparing the dataset. This includes [tokenizing](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L674) the prompt and [applying transforms](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L714) to the images and conditioning images.

 ```py
 conditioning_image_transforms = transforms.Compose(
@@ -32,9 +32,9 @@ The denoiser checkpoint can also have multiple shards and supports inference tha
 For example, let's save a sharded checkpoint for the [SDXL UNet](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main/unet):

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

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

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

-unet = UNet2DConditionModel.from_pretrained(
+unet = AutoModel.from_pretrained(
    "sayakpaul/sdxl-unet-sharded", torch_dtype=torch.float16
 )
 pipeline = StableDiffusionXLPipeline.from_pretrained(
@@ -95,6 +95,23 @@ Use the Space below to gauge a pipeline's memory requirements before you downloa
    ></iframe>
 </div>

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

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

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

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

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

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

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

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

 </Tip>

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

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

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

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

-base_unet = UNet2DConditionModel.from_pretrained(
+base_unet = AutoModel.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    use_safetensors=True,
@@ -1,7 +1,8 @@
-accelerate>=0.16.0
+accelerate>=0.31.0
 torchvision
-transformers>=4.25.1
+transformers>=4.41.2
 ftfy
 tensorboard
 Jinja2
-peft==0.7.0
+peft>=0.11.1
+sentencepiece
@@ -24,7 +24,7 @@ import re
 import shutil
 from contextlib import nullcontext
 from pathlib import Path
-from typing import List, Optional, Union
+from typing import List, Optional

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

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

 logger = get_logger(__name__)

@@ -228,10 +228,20 @@ def log_validation(

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

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

    for tracker in accelerator.trackers:
        phase_name = "test" if is_final_validation else "validation"
@@ -657,6 +667,7 @@ def parse_args(input_args=None):
    parser.add_argument(
        "--adam_weight_decay_text_encoder", type=float, default=1e-03, help="Weight decay to use for text_encoder"
    )
+
    parser.add_argument(
        "--lora_layers",
        type=str,
@@ -666,6 +677,7 @@ def parse_args(input_args=None):
            'E.g. - "to_k,to_q,to_v,to_out.0" will result in lora training of attention layers only. For more examples refer to https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/README_flux.md'
        ),
    )
+
    parser.add_argument(
        "--adam_epsilon",
        type=float,
@@ -738,6 +750,15 @@ def parse_args(input_args=None):
            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
        ),
    )
+    parser.add_argument(
+        "--upcast_before_saving",
+        action="store_true",
+        default=False,
+        help=(
+            "Whether to upcast the trained transformer layers to float32 before saving (at the end of training). "
+            "Defaults to precision dtype used for training to save memory"
+        ),
+    )
    parser.add_argument(
        "--prior_generation_precision",
        type=str,
@@ -818,9 +839,9 @@ class TokenEmbeddingsHandler:
        idx = 0
        for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
            assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )

            self.inserting_toks = inserting_toks
            special_tokens_dict = {"additional_special_tokens": self.inserting_toks}
@@ -1147,7 +1168,7 @@ def tokenize_prompt(tokenizer, prompt, max_sequence_length, add_special_tokens=F
    return text_input_ids


-def _get_t5_prompt_embeds(
+def _encode_prompt_with_t5(
    text_encoder,
    tokenizer,
    max_sequence_length=512,
@@ -1176,7 +1197,10 @@ def _get_t5_prompt_embeds(

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

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

    _, seq_len, _ = prompt_embeds.shape
@@ -1188,7 +1212,7 @@ def _get_t5_prompt_embeds(
    return prompt_embeds


-def _get_clip_prompt_embeds(
+def _encode_prompt_with_clip(
    text_encoder,
    tokenizer,
    prompt: str,
@@ -1217,9 +1241,13 @@ def _get_clip_prompt_embeds(

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

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

    # duplicate text embeddings for each generation per prompt, using mps friendly method
    prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
@@ -1238,136 +1266,35 @@ def encode_prompt(
    text_input_ids_list=None,
 ):
    prompt = [prompt] if isinstance(prompt, str) else prompt
-    batch_size = len(prompt)
-    dtype = text_encoders[0].dtype
+    if hasattr(text_encoders[0], "module"):
+        dtype = text_encoders[0].module.dtype
+    else:
+        dtype = text_encoders[0].dtype

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

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

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

    return prompt_embeds, pooled_prompt_embeds, text_ids


-# CustomFlowMatchEulerDiscreteScheduler was taken from ostris ai-toolkit trainer:
-# https://github.com/ostris/ai-toolkit/blob/9ee1ef2a0a2a9a02b92d114a95f21312e5906e54/toolkit/samplers/custom_flowmatch_sampler.py#L95
-class CustomFlowMatchEulerDiscreteScheduler(FlowMatchEulerDiscreteScheduler):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        with torch.no_grad():
-            # create weights for timesteps
-            num_timesteps = 1000
-
-            # generate the multiplier based on cosmap loss weighing
-            # this is only used on linear timesteps for now
-
-            # cosine map weighing is higher in the middle and lower at the ends
-            # bot = 1 - 2 * self.sigmas + 2 * self.sigmas ** 2
-            # cosmap_weighing = 2 / (math.pi * bot)
-
-            # sigma sqrt weighing is significantly higher at the end and lower at the beginning
-            sigma_sqrt_weighing = (self.sigmas**-2.0).float()
-            # clip at 1e4 (1e6 is too high)
-            sigma_sqrt_weighing = torch.clamp(sigma_sqrt_weighing, max=1e4)
-            # bring to a mean of 1
-            sigma_sqrt_weighing = sigma_sqrt_weighing / sigma_sqrt_weighing.mean()
-
-            # Create linear timesteps from 1000 to 0
-            timesteps = torch.linspace(1000, 0, num_timesteps, device="cpu")
-
-            self.linear_timesteps = timesteps
-            # self.linear_timesteps_weights = cosmap_weighing
-            self.linear_timesteps_weights = sigma_sqrt_weighing
-
-            # self.sigmas = self.get_sigmas(timesteps, n_dim=1, dtype=torch.float32, device='cpu')
-            pass
-
-    def get_weights_for_timesteps(self, timesteps: torch.Tensor) -> torch.Tensor:
-        # Get the indices of the timesteps
-        step_indices = [(self.timesteps == t).nonzero().item() for t in timesteps]
-
-        # Get the weights for the timesteps
-        weights = self.linear_timesteps_weights[step_indices].flatten()
-
-        return weights
-
-    def get_sigmas(self, timesteps: torch.Tensor, n_dim, dtype, device) -> torch.Tensor:
-        sigmas = self.sigmas.to(device=device, dtype=dtype)
-        schedule_timesteps = self.timesteps.to(device)
-        timesteps = timesteps.to(device)
-        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
-
-        sigma = sigmas[step_indices].flatten()
-        while len(sigma.shape) < n_dim:
-            sigma = sigma.unsqueeze(-1)
-
-        return sigma
-
-    def add_noise(
-        self,
-        original_samples: torch.Tensor,
-        noise: torch.Tensor,
-        timesteps: torch.Tensor,
-    ) -> torch.Tensor:
-        ## ref https://github.com/huggingface/diffusers/blob/fbe29c62984c33c6cf9cf7ad120a992fe6d20854/examples/dreambooth/train_dreambooth_sd3.py#L1578
-        ## Add noise according to flow matching.
-        ## zt = (1 - texp) * x + texp * z1
-
-        # sigmas = get_sigmas(timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
-        # noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise
-
-        # timestep needs to be in [0, 1], we store them in [0, 1000]
-        # noisy_sample = (1 - timestep) * latent + timestep * noise
-        t_01 = (timesteps / 1000).to(original_samples.device)
-        noisy_model_input = (1 - t_01) * original_samples + t_01 * noise
-
-        # n_dim = original_samples.ndim
-        # sigmas = self.get_sigmas(timesteps, n_dim, original_samples.dtype, original_samples.device)
-        # noisy_model_input = (1.0 - sigmas) * original_samples + sigmas * noise
-        return noisy_model_input
-
-    def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
-        return sample
-
-    def set_train_timesteps(self, num_timesteps, device, linear=False):
-        if linear:
-            timesteps = torch.linspace(1000, 0, num_timesteps, device=device)
-            self.timesteps = timesteps
-            return timesteps
-        else:
-            # distribute them closer to center. Inference distributes them as a bias toward first
-            # Generate values from 0 to 1
-            t = torch.sigmoid(torch.randn((num_timesteps,), device=device))
-
-            # Scale and reverse the values to go from 1000 to 0
-            timesteps = (1 - t) * 1000
-
-            # Sort the timesteps in descending order
-            timesteps, _ = torch.sort(timesteps, descending=True)
-
-            self.timesteps = timesteps.to(device=device)
-
-            return timesteps
-
-
 def main(args):
    if args.report_to == "wandb" and args.hub_token is not None:
        raise ValueError(
@@ -1499,7 +1426,7 @@ def main(args):
    )

    # Load scheduler and models
-    noise_scheduler = CustomFlowMatchEulerDiscreteScheduler.from_pretrained(
+    noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="scheduler"
    )
    noise_scheduler_copy = copy.deepcopy(noise_scheduler)
@@ -1619,7 +1546,6 @@ def main(args):
        target_modules=target_modules,
    )
    transformer.add_adapter(transformer_lora_config)
-
    if args.train_text_encoder:
        text_lora_config = LoraConfig(
            r=args.rank,
@@ -1679,7 +1605,7 @@ def main(args):
        lora_state_dict = FluxPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1727,7 +1653,6 @@ def main(args):
        cast_training_params(models, dtype=torch.float32)

    transformer_lora_parameters = list(filter(lambda p: p.requires_grad, transformer.parameters()))
-
    if args.train_text_encoder:
        text_lora_parameters_one = list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
    # if we use textual inversion, we freeze all parameters except for the token embeddings
@@ -1737,7 +1662,8 @@ def main(args):
        for name, param in text_encoder_one.named_parameters():
            if "token_embedding" in name:
                # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-                param.data = param.to(dtype=torch.float32)
+                if args.mixed_precision == "fp16":
+                    param.data = param.to(dtype=torch.float32)
                param.requires_grad = True
                text_lora_parameters_one.append(param)
            else:
@@ -1747,7 +1673,8 @@ def main(args):
            for name, param in text_encoder_two.named_parameters():
                if "shared" in name:
                    # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-                    param.data = param.to(dtype=torch.float32)
+                    if args.mixed_precision == "fp16":
+                        param.data = param.to(dtype=torch.float32)
                    param.requires_grad = True
                    text_lora_parameters_two.append(param)
                else:
@@ -1828,6 +1755,7 @@ def main(args):
            optimizer_class = bnb.optim.AdamW8bit
        else:
            optimizer_class = torch.optim.AdamW
+
        optimizer = optimizer_class(
            params_to_optimize,
            betas=(args.adam_beta1, args.adam_beta2),
@@ -1987,17 +1915,22 @@ def main(args):
            free_memory()

    # 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 * accelerator.num_processes * num_update_steps_per_epoch
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )
@@ -2032,8 +1965,14 @@ def main(args):

    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps:
+            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)

@@ -2125,7 +2064,7 @@ def main(args):
                if args.train_text_encoder:
                    text_encoder_one.train()
                    # set top parameter requires_grad = True for gradient checkpointing works
-                    accelerator.unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
+                    unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
                elif args.train_text_encoder_ti:  # textual inversion / pivotal tuning
                    text_encoder_one.train()
                if args.enable_t5_ti:
@@ -2137,6 +2076,11 @@ def main(args):
                pivoted_tr = True

        for step, batch in enumerate(train_dataloader):
+            models_to_accumulate = [transformer]
+            if not freeze_text_encoder:
+                models_to_accumulate.extend([text_encoder_one])
+                if args.enable_t5_ti:
+                    models_to_accumulate.extend([text_encoder_two])
            if pivoted_te:
                # stopping optimization of text_encoder params
                optimizer.param_groups[te_idx]["lr"] = 0.0
@@ -2145,7 +2089,7 @@ def main(args):
                logger.info(f"PIVOT TRANSFORMER {epoch}")
                optimizer.param_groups[0]["lr"] = 0.0

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

                # encode batch prompts when custom prompts are provided for each image -
@@ -2189,7 +2133,7 @@ def main(args):
                model_input = (model_input - vae_config_shift_factor) * vae_config_scaling_factor
                model_input = model_input.to(dtype=weight_dtype)

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

                latent_image_ids = FluxPipeline._prepare_latent_image_ids(
                    model_input.shape[0],
@@ -2228,7 +2172,7 @@ def main(args):
                )

                # handle guidance
-                if transformer.config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                    guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                    guidance = guidance.expand(model_input.shape[0])
                else:
@@ -2237,7 +2181,7 @@ def main(args):
                # Predict the noise residual
                model_pred = transformer(
                    hidden_states=packed_noisy_model_input,
-                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
+                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transformer model (we should not keep it but I want to keep the inputs same for the model for testing)
                    timestep=timesteps / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
@@ -2288,16 +2232,26 @@ def main(args):
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    if not freeze_text_encoder:
-                        if args.train_text_encoder:
+                        if args.train_text_encoder:  # text encoder tuning
                            params_to_clip = itertools.chain(transformer.parameters(), text_encoder_one.parameters())
                        elif pure_textual_inversion:
-                            params_to_clip = itertools.chain(
-                                text_encoder_one.parameters(), text_encoder_two.parameters()
-                            )
+                            if args.enable_t5_ti:
+                                params_to_clip = itertools.chain(
+                                    text_encoder_one.parameters(), text_encoder_two.parameters()
+                                )
+                            else:
+                                params_to_clip = itertools.chain(text_encoder_one.parameters())
                        else:
-                            params_to_clip = itertools.chain(
-                                transformer.parameters(), text_encoder_one.parameters(), text_encoder_two.parameters()
-                            )
+                            if args.enable_t5_ti:
+                                params_to_clip = itertools.chain(
+                                    transformer.parameters(),
+                                    text_encoder_one.parameters(),
+                                    text_encoder_two.parameters(),
+                                )
+                            else:
+                                params_to_clip = itertools.chain(
+                                    transformer.parameters(), text_encoder_one.parameters()
+                                )
                    else:
                        params_to_clip = itertools.chain(transformer.parameters())
                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
@@ -2339,6 +2293,10 @@ def main(args):

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

            logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
@@ -2351,14 +2309,16 @@ def main(args):
        if accelerator.is_main_process:
            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                # create pipeline
-                if freeze_text_encoder:
+                if freeze_text_encoder:  # no text encoder one, two optimizations
                    text_encoder_one, text_encoder_two = load_text_encoders(text_encoder_cls_one, text_encoder_cls_two)
+                    text_encoder_one.to(weight_dtype)
+                    text_encoder_two.to(weight_dtype)
                pipeline = FluxPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
-                    transformer=accelerator.unwrap_model(transformer),
+                    text_encoder=unwrap_model(text_encoder_one),
+                    text_encoder_2=unwrap_model(text_encoder_two),
+                    transformer=unwrap_model(transformer),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
@@ -2372,21 +2332,21 @@ def main(args):
                    epoch=epoch,
                    torch_dtype=weight_dtype,
                )
-                images = None
-                del pipeline
-
                if freeze_text_encoder:
                    del text_encoder_one, text_encoder_two
                    free_memory()
-                elif args.train_text_encoder:
-                    del text_encoder_two
-                    free_memory()
+
+                images = None
+                del pipeline

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

        if args.train_text_encoder:
@@ -2428,8 +2388,8 @@ def main(args):
                accelerator=accelerator,
                pipeline_args=pipeline_args,
                epoch=epoch,
-                torch_dtype=weight_dtype,
                is_final_validation=True,
+                torch_dtype=weight_dtype,
            )

        save_model_card(
@@ -2452,6 +2412,7 @@ def main(args):
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )
+
        images = None
        del pipeline

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


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

 logger = get_logger(__name__)

@@ -200,7 +200,8 @@ Special VAE used for training: {vae_path}.
        "diffusers",
        "diffusers-training",
        lora,
-        "template:sd-lora" "stable-diffusion",
+        "template:sd-lora",
+        "stable-diffusion",
        "stable-diffusion-diffusers",
    ]
    model_card = populate_model_card(model_card, tags=tags)
@@ -724,9 +725,9 @@ class TokenEmbeddingsHandler:
        idx = 0
        for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
            assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )

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

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

        lora_state_dict, network_alphas = StableDiffusionPipeline.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
        if incompatible_keys is not None:
@@ -71,6 +71,7 @@ from diffusers.utils import (
    convert_unet_state_dict_to_peft,
    is_wandb_available,
 )
+from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module

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

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

 logger = get_logger(__name__)

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

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

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

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


 def log_validation(
@@ -891,9 +890,9 @@ class TokenEmbeddingsHandler:
        idx = 0
        for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
            assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )

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

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

        lora_state_dict, network_alphas = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
        if incompatible_keys is not None:
@@ -720,7 +720,7 @@ def main(args):
    # Train!
    logger.info("***** Running training *****")
    logger.info(f"  Num training steps = {args.max_train_steps}")
-    logger.info(f"  Instantaneous batch size per device = { args.train_batch_size}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")

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

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

 logger = get_logger(__name__)

@@ -1138,7 +1138,7 @@ def main(args):
        lora_state_dict = CogVideoXImageToVideoPipeline.lora_state_dict(input_dir)

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

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

 logger = get_logger(__name__)

@@ -1159,7 +1159,7 @@ def main(args):
        lora_state_dict = CogVideoXPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -59,7 +59,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -10,7 +10,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif

 | Example                                                                                                                               | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                                              | Colab                                                                                                                                                                                                              |                                                        Author |
 |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
-|Spatiotemporal Skip Guidance (STG)|[Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664) (CVPR 2025) enhances video diffusion models by generating a weaker model through layer skipping and using it as guidance, improving fidelity in models like HunyuanVideo, LTXVideo, and Mochi.|[Spatiotemporal Skip Guidance](#spatiotemporal-skip-guidance)|-|[Junha Hyung](https://junhahyung.github.io/), [Kinam Kim](https://kinam0252.github.io/)|
+|Spatiotemporal Skip Guidance (STG)|[Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664) (CVPR 2025) enhances video diffusion models by generating a weaker model through layer skipping and using it as guidance, improving fidelity in models like HunyuanVideo, LTXVideo, and Mochi.|[Spatiotemporal Skip Guidance](#spatiotemporal-skip-guidance)|-|[Junha Hyung](https://junhahyung.github.io/), [Kinam Kim](https://kinam0252.github.io/), and [Ednaordinary](https://github.com/Ednaordinary)|
 |Adaptive Mask Inpainting|Adaptive Mask Inpainting algorithm from [Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models](https://github.com/snuvclab/coma) (ECCV '24, Oral) provides a way to insert human inside the scene image without altering the background, by inpainting with adapting mask.|[Adaptive Mask Inpainting](#adaptive-mask-inpainting)|-|[Hyeonwoo Kim](https://sshowbiz.xyz),[Sookwan Han](https://jellyheadandrew.github.io)|
 |Flux with CFG|[Flux with CFG](https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md) provides an implementation of using CFG in [Flux](https://blackforestlabs.ai/announcing-black-forest-labs/).|[Flux with CFG](#flux-with-cfg)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/flux_with_cfg.ipynb)|[Linoy Tsaban](https://github.com/linoytsaban), [Apolinário](https://github.com/apolinario), and [Sayak Paul](https://github.com/sayakpaul)|
 |Differential Diffusion|[Differential Diffusion](https://github.com/exx8/differential-diffusion) modifies an image according to a text prompt, and according to a map that specifies the amount of change in each region.|[Differential Diffusion](#differential-diffusion)|[![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/exx8/differential-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/exx8/differential-diffusion/blob/main/examples/SD2.ipynb)|[Eran Levin](https://github.com/exx8) and [Ohad Fried](https://www.ohadf.com/)|
@@ -85,7 +85,7 @@ PIXART-α Controlnet pipeline | Implementation of the controlnet model for pixar
 | Stable Diffusion XL Attentive Eraser Pipeline |[[AAAI2025 Oral] Attentive Eraser](https://github.com/Anonym0u3/AttentiveEraser) is a novel tuning-free method that enhances object removal capabilities in pre-trained diffusion models.|[Stable Diffusion XL Attentive Eraser Pipeline](#stable-diffusion-xl-attentive-eraser-pipeline)|-|[Wenhao Sun](https://github.com/Anonym0u3) and [Benlei Cui](https://github.com/Benny079)|
 | Perturbed-Attention Guidance |StableDiffusionPAGPipeline is a modification of StableDiffusionPipeline to support Perturbed-Attention Guidance (PAG).|[Perturbed-Attention Guidance](#perturbed-attention-guidance)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/perturbed_attention_guidance.ipynb)|[Hyoungwon Cho](https://github.com/HyoungwonCho)|
 | CogVideoX DDIM Inversion Pipeline | Implementation of DDIM inversion and guided attention-based editing denoising process on CogVideoX. | [CogVideoX DDIM Inversion Pipeline](#cogvideox-ddim-inversion-pipeline) | - | [LittleNyima](https://github.com/LittleNyima) |
-
+| FaithDiff Stable Diffusion XL Pipeline | Implementation of [(CVPR 2025) FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolutionUnleashing Diffusion Priors for Faithful Image Super-resolution](https://arxiv.org/abs/2411.18824) - FaithDiff is a faithful image super-resolution method that leverages latent diffusion models by actively adapting the diffusion prior and jointly fine-tuning its components (encoder and diffusion model) with an alignment module to ensure high fidelity and structural consistency. | [FaithDiff Stable Diffusion XL Pipeline](#faithdiff-stable-diffusion-xl-pipeline) | [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/jychen9811/FaithDiff) | [Junyang Chen, Jinshan Pan, Jiangxin Dong, IMAG Lab, (Adapted by Eliseu Silva)](https://github.com/JyChen9811/FaithDiff) |
 To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.

 ```py
@@ -124,7 +124,6 @@ pipe = pipe.to("cuda")
 #--------Option--------#
 prompt = "A close-up of a beautiful woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues, realistic style."
 stg_applied_layers_idx = [34]
-stg_mode = "STG"
 stg_scale = 1.0 # 0.0 for CFG
 #----------------------#

@@ -5334,3 +5333,103 @@ output = pipeline_for_inversion(
 pipeline.export_latents_to_video(output.inverse_latents[-1], "path/to/inverse_video.mp4", fps=8)
 pipeline.export_latents_to_video(output.recon_latents[-1], "path/to/recon_video.mp4", fps=8)
 ```
+# FaithDiff Stable Diffusion XL Pipeline
+
+[Project](https://jychen9811.github.io/FaithDiff_page/) / [GitHub](https://github.com/JyChen9811/FaithDiff/)
+
+This the implementation of the FaithDiff pipeline for SDXL, adapted to use the HuggingFace Diffusers.
+
+For more details see the project links above.
+
+## Example Usage
+
+This example upscale and restores a low-quality image. The input image has a resolution of 512x512 and will be upscaled at a scale of 2x, to a final resolution of 1024x1024. It is possible to upscale to a larger scale, but it is recommended that the input image be at least 1024x1024 in these cases. To upscale this image by 4x, for example, it would be recommended to re-input the result into a new 2x processing, thus performing progressive scaling.
+
+````py
+import random
+import numpy as np
+import torch
+from diffusers import DiffusionPipeline, AutoencoderKL, UniPCMultistepScheduler
+from huggingface_hub import hf_hub_download
+from diffusers.utils import load_image
+from PIL import Image
+
+device = "cuda"
+dtype = torch.float16
+MAX_SEED = np.iinfo(np.int32).max
+
+# Download weights for additional unet layers
+model_file = hf_hub_download(
+    "jychen9811/FaithDiff",
+    filename="FaithDiff.bin", local_dir="./proc_data/faithdiff", local_dir_use_symlinks=False
+)
+
+# Initialize the models and pipeline
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=dtype)
+
+model_id = "SG161222/RealVisXL_V4.0"
+pipe = DiffusionPipeline.from_pretrained(
+    model_id,
+    torch_dtype=dtype,
+    vae=vae,
+    unet=None, #<- Do not load with original model.
+    custom_pipeline="pipeline_faithdiff_stable_diffusion_xl",    
+    use_safetensors=True,
+    variant="fp16",
+).to(device)
+
+# Here we need use pipeline internal unet model
+pipe.unet = pipe.unet_model.from_pretrained(model_id, subfolder="unet", variant="fp16", use_safetensors=True)
+
+# Load additional layers to the model
+pipe.unet.load_additional_layers(weight_path="proc_data/faithdiff/FaithDiff.bin", dtype=dtype)
+
+# Enable vae tiling
+pipe.set_encoder_tile_settings()
+pipe.enable_vae_tiling()
+
+# Optimization
+pipe.enable_model_cpu_offload()
+
+# Set selected scheduler
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+
+#input params
+prompt = "The image features a woman in her 55s with blonde hair and a white shirt, smiling at the camera. She appears to be in a good mood and is wearing a white scarf around her neck. "
+upscale = 2 # scale here
+start_point = "lr" # or "noise"
+latent_tiled_overlap = 0.5
+latent_tiled_size = 1024
+
+# Load image
+lq_image = load_image("https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/woman.png")
+original_height = lq_image.height
+original_width = lq_image.width
+print(f"Current resolution: H:{original_height} x W:{original_width}")
+
+width = original_width * int(upscale)
+height = original_height * int(upscale)
+print(f"Final resolution: H:{height} x W:{width}")
+
+# Restoration
+image = lq_image.resize((width, height), Image.LANCZOS)
+input_image, width_init, height_init, width_now, height_now = pipe.check_image_size(image)
+
+generator = torch.Generator(device=device).manual_seed(random.randint(0, MAX_SEED))
+gen_image = pipe(lr_img=input_image, 
+                 prompt = prompt,                  
+                 num_inference_steps=20, 
+                 guidance_scale=5, 
+                 generator=generator, 
+                 start_point=start_point, 
+                 height = height_now, 
+                 width=width_now, 
+                 overlap=latent_tiled_overlap, 
+                 target_size=(latent_tiled_size, latent_tiled_size)
+                ).images[0]
+
+cropped_image = gen_image.crop((0, 0, width_init, height_init))
+cropped_image.save("data/result.png")
+````
+### Result
+[<img src="https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/faithdiff_restored.PNG" width="512px" height="512px"/>](https://imgsli.com/MzY1NzE2)
@@ -1103,7 +1103,7 @@ class AdaptiveMaskInpaintPipeline(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `default_mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -312,9 +312,9 @@ if __name__ == "__main__":
                    # These are the coordinates of the output image
                    out_coordinates = np.arange(1, out_length + 1)

-                    # since both scale-factor and output size can be provided simulatneously, perserving the center of the image requires shifting
-                    # the output coordinates. the deviation is because out_length doesn't necesary equal in_length*scale.
-                    # to keep the center we need to subtract half of this deivation so that we get equal margins for boths sides and center is preserved.
+                    # since both scale-factor and output size can be provided simultaneously, preserving the center of the image requires shifting
+                    # the output coordinates. the deviation is because out_length doesn't necessary equal in_length*scale.
+                    # to keep the center we need to subtract half of this deviation so that we get equal margins for both sides and center is preserved.
                    shifted_out_coordinates = out_coordinates - (out_length - in_length * scale) / 2

                    # These are the matching positions of the output-coordinates on the input image coordinates.
@@ -351,7 +351,7 @@ def my_forward(
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
            added_cond_kwargs: (`dict`, *optional*):
-                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
                are passed along to the UNet blocks.

        Returns:
@@ -864,9 +864,9 @@ def get_flow_and_interframe_paras(flow_model, imgs):
 class AttentionControl:
    """
    Control FRESCO-based attention
-    * enable/diable spatial-guided attention
-    * enable/diable temporal-guided attention
-    * enable/diable cross-frame attention
+    * enable/disable spatial-guided attention
+    * enable/disable temporal-guided attention
+    * enable/disable cross-frame attention
    * collect intermediate attention feature (for spatial-guided attention)
    """

@@ -34,7 +34,7 @@ class RASGAttnProcessor:
        temb: Optional[torch.Tensor] = None,
        scale: float = 1.0,
    ) -> torch.Tensor:
-        # Same as the default AttnProcessor up untill the part where similarity matrix gets saved
+        # Same as the default AttnProcessor up until the part where similarity matrix gets saved
        downscale_factor = self.mask_resoltuion // hidden_states.shape[1]
        residual = hidden_states

@@ -686,7 +686,7 @@ class StableDiffusionHDPainterPipeline(StableDiffusionInpaintPipeline):
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -362,7 +362,7 @@ class ImageToImageInpaintingPipeline(DiffusionPipeline):
                f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                " `pipeline.unet` or your `mask_image` or `image` input."
            )

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

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

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

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

            finally:
@@ -43,7 +43,7 @@ from diffusers.utils import BaseOutput, check_min_version


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


 class MarigoldDepthOutput(BaseOutput):
@@ -701,7 +701,7 @@ class StableDiffusionXLControlNetTileSRPipeline(
            raise ValueError("`max_tile_size` cannot be None.")
        elif not isinstance(max_tile_size, int) or max_tile_size not in (1024, 1280):
            raise ValueError(
-                f"`max_tile_size` has to be in 1024 or 1280 but is {max_tile_size} of type" f" {type(max_tile_size)}."
+                f"`max_tile_size` has to be in 1024 or 1280 but is {max_tile_size} of type {type(max_tile_size)}."
            )
        if tile_gaussian_sigma is None:
            raise ValueError("`tile_gaussian_sigma` cannot be None.")
@@ -488,7 +488,7 @@ class FluxDifferentialImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
        if padding_mask_crop is not None:
            if not isinstance(image, PIL.Image.Image):
                raise ValueError(
-                    f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
+                    f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}."
                )
            if not isinstance(mask_image, PIL.Image.Image):
                raise ValueError(
@@ -496,7 +496,7 @@ class FluxDifferentialImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                    f" {type(mask_image)}."
                )
            if output_type != "pil":
-                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
+                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.")

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

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

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

        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
            raise ValueError(
@@ -2050,7 +2050,7 @@ class StableDiffusionXL_AE_Pipeline(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -33,7 +33,6 @@ from diffusers import DiffusionPipeline
 from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
 from diffusers.loaders import (
    FromSingleFileMixin,
-    StableDiffusionLoraLoaderMixin,
    StableDiffusionXLLoraLoaderMixin,
    TextualInversionLoaderMixin,
 )
@@ -300,7 +299,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):


 class StableDiffusionXLControlNetAdapterInpaintPipeline(
-    DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin, StableDiffusionLoraLoaderMixin
+    DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin
 ):
    r"""
    Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter
@@ -1578,7 +1577,7 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -0,0 +1,661 @@
+# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+import types
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import ftfy
+import regex as re
+import torch
+from transformers import AutoTokenizer, UMT5EncoderModel
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import WanLoraLoaderMixin
+from diffusers.models import AutoencoderKLWan, WanTransformer3DModel
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.wan.pipeline_output import WanPipelineOutput
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from diffusers import AutoencoderKLWan
+        >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
+        >>> from examples.community.pipeline_stg_wan import WanSTGPipeline
+
+        >>> # Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+        >>> model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
+        >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+        >>> pipe = WanSTGPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+        >>> flow_shift = 5.0  # 5.0 for 720P, 3.0 for 480P
+        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+        >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [8] # Layer indices from 0 to 39 for 14b or 0 to 29 for 1.3b
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+
+        >>> output = pipe(
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     height=720,
+        ...     width=1280,
+        ...     num_frames=81,
+        ...     guidance_scale=5.0,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ... ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=16)
+        ```
+"""
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def prompt_clean(text):
+    text = whitespace_clean(basic_clean(text))
+    return text
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    rotary_emb: torch.Tensor,
+) -> torch.Tensor:
+    return hidden_states
+
+
+def forward_without_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    rotary_emb: torch.Tensor,
+) -> torch.Tensor:
+    shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+        self.scale_shift_table + temb.float()
+    ).chunk(6, dim=1)
+
+    # 1. Self-attention
+    norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
+    attn_output = self.attn1(hidden_states=norm_hidden_states, rotary_emb=rotary_emb)
+    hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
+
+    # 2. Cross-attention
+    norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
+    attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+    hidden_states = hidden_states + attn_output
+
+    # 3. Feed-forward
+    norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(hidden_states)
+    ff_output = self.ffn(norm_hidden_states)
+    hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
+
+    return hidden_states
+
+
+class WanSTGPipeline(DiffusionPipeline, WanLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation using Wan.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        tokenizer ([`T5Tokenizer`]):
+            Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer),
+            specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        transformer ([`WanTransformer3DModel`]):
+            Conditional Transformer to denoise the input latents.
+        scheduler ([`UniPCMultistepScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: UMT5EncoderModel,
+        transformer: WanTransformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt = [prompt_clean(u) for u in prompt]
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
+        seq_lens = mask.gt(0).sum(dim=1).long()
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
+        )
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        negative_prompt,
+        height,
+        width,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif negative_prompt is not None and (
+            not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list)
+        ):
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: int = 16,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_latent_frames,
+            int(height) // self.vae_scale_factor_spatial,
+            int(width) // self.vae_scale_factor_spatial,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "np",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+        stg_applied_layers_idx: Optional[List[int]] = [3, 8, 16],
+        stg_scale: Optional[float] = 0.0,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `480`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `832`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `81`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `5.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
+                The dtype to use for the torch.amp.autocast.
+
+        Examples:
+
+        Returns:
+            [`~WanPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where
+                the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._stg_scale = stg_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                latent_model_input = latents.to(transformer_dtype)
+                timestep = t.expand(latents.shape[0])
+
+                if self.do_spatio_temporal_guidance:
+                    for idx, block in enumerate(self.transformer.blocks):
+                        block.forward = types.MethodType(forward_without_stg, block)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    if self.do_spatio_temporal_guidance:
+                        for idx, block in enumerate(self.transformer.blocks):
+                            if idx in stg_applied_layers_idx:
+                                block.forward = types.MethodType(forward_with_stg, block)
+                        noise_perturb = self.transformer(
+                            hidden_states=latent_model_input,
+                            timestep=timestep,
+                            encoder_hidden_states=prompt_embeds,
+                            attention_kwargs=attention_kwargs,
+                            return_dict=False,
+                        )[0]
+                        noise_pred = (
+                            noise_uncond
+                            + guidance_scale * (noise_pred - noise_uncond)
+                            + self._stg_scale * (noise_pred - noise_perturb)
+                        )
+                    else:
+                        noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return WanPipelineOutput(frames=video)
@@ -288,8 +288,7 @@ class UFOGenScheduler(SchedulerMixin, ConfigMixin):

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

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

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

 logger = get_logger(__name__)

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

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

    return kohya_ss_state_dict
@@ -889,7 +889,7 @@ def main(args):
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
-        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be devide by the number of processes assuming batches are multiplied by the number of processes
+        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be divided by the number of processes assuming batches are multiplied by the number of processes
    )

    # Make one log on every process with the configuration for debugging.
@@ -66,7 +66,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -721,7 +721,7 @@ def main(args):
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
-        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be devide by the number of processes assuming batches are multiplied by the number of processes
+        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be divided by the number of processes assuming batches are multiplied by the number of processes
    )

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

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

 logger = get_logger(__name__)

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

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

    return kohya_ss_state_dict
@@ -884,7 +884,7 @@ def main(args):
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
-        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be devide by the number of processes assuming batches are multiplied by the number of processes
+        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be divided by the number of processes assuming batches are multiplied by the number of processes
    )

    # Make one log on every process with the configuration for debugging.
@@ -72,7 +72,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -854,7 +854,7 @@ def main(args):
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
-        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be devide by the number of processes assuming batches are multiplied by the number of processes
+        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be divided by the number of processes assuming batches are multiplied by the number of processes
    )

    # Make one log on every process with the configuration for debugging.
@@ -78,7 +78,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -894,7 +894,7 @@ def main(args):
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
-        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be devide by the number of processes assuming batches are multiplied by the number of processes
+        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be divided by the number of processes assuming batches are multiplied by the number of processes
    )

    # Make one log on every process with the configuration for debugging.
@@ -6,7 +6,19 @@ Training script provided by LibAI, which is an institution dedicated to the prog
 > [!NOTE]
 > **Memory consumption**
 >
-> Flux can be quite expensive to run on consumer hardware devices and as a result, ControlNet training of it comes with higher memory requirements than usual. 
+> Flux can be quite expensive to run on consumer hardware devices and as a result, ControlNet training of it comes with higher memory requirements than usual.
+
+Here is a gpu memory consumption for reference, tested on a single A100 with 80G.
+
+| period | GPU |
+| - | - | 
+| load as float32 | ~70G |
+| mv transformer and vae to bf16 | ~48G |
+| pre compute txt embeddings | ~62G |
+| **offload te to cpu** | ~30G |
+| training | ~58G |
+| validation | ~71G |
+

 > **Gated access**
 >
@@ -98,8 +110,9 @@ accelerate launch train_controlnet_flux.py \
    --validation_image "./conditioning_image_1.png" "./conditioning_image_2.png" \
    --validation_prompt "red circle with blue background" "cyan circle with brown floral background" \
    --train_batch_size=1 \
-    --gradient_accumulation_steps=4 \
+    --gradient_accumulation_steps=16 \
    --report_to="wandb" \
+    --lr_scheduler="cosine" \
    --num_double_layers=4 \
    --num_single_layers=0 \
    --seed=42 \
@@ -60,7 +60,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -927,17 +927,22 @@ def main(args):
    )

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

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )
@@ -962,8 +967,14 @@ def main(args):

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

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

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

 logger = logging.getLogger(__name__)

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

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

 logger = get_logger(__name__)
 if is_torch_npu_available():
@@ -148,7 +148,7 @@ def log_validation(
                    pooled_prompt_embeds=pooled_prompt_embeds,
                    control_image=validation_image,
                    num_inference_steps=28,
-                    controlnet_conditioning_scale=0.7,
+                    controlnet_conditioning_scale=1,
                    guidance_scale=3.5,
                    generator=generator,
                ).images[0]
@@ -1085,8 +1085,6 @@ def main(args):
        return {"prompt_embeds": prompt_embeds, "pooled_prompt_embeds": pooled_prompt_embeds, "text_ids": text_ids}

    train_dataset = get_train_dataset(args, accelerator)
-    text_encoders = [text_encoder_one, text_encoder_two]
-    tokenizers = [tokenizer_one, tokenizer_two]
    compute_embeddings_fn = functools.partial(
        compute_embeddings,
        flux_controlnet_pipeline=flux_controlnet_pipeline,
@@ -1103,7 +1101,8 @@ def main(args):
            compute_embeddings_fn, batched=True, new_fingerprint=new_fingerprint, batch_size=50
        )

-    del text_encoders, tokenizers, text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two
+    text_encoder_one.to("cpu")
+    text_encoder_two.to("cpu")
    free_memory()

    # Then get the training dataset ready to be passed to the dataloader.
@@ -17,6 +17,7 @@ import argparse
 import contextlib
 import copy
 import functools
+import gc
 import logging
 import math
 import os
@@ -52,6 +53,7 @@ from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_density_for_timestep_sampling, compute_loss_weighting_for_sd3, free_memory
 from diffusers.utils import check_min_version, is_wandb_available, make_image_grid
 from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
+from diffusers.utils.testing_utils import backend_empty_cache
 from diffusers.utils.torch_utils import is_compiled_module


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

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

 logger = get_logger(__name__)

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

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

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

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

        images = []

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

            images.append(image)
@@ -655,6 +695,7 @@ def make_train_dataset(args, tokenizer_one, tokenizer_two, tokenizer_three, acce
            dataset = load_dataset(
                args.train_data_dir,
                cache_dir=args.cache_dir,
+                trust_remote_code=True,
            )
        # See more about loading custom images at
        # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
@@ -61,7 +61,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)
 if is_torch_npu_available():
@@ -50,9 +50,11 @@ def retrieve(class_prompt, class_data_dir, num_class_images):
    total = 0
    pbar = tqdm(desc="downloading real regularization images", total=num_class_images)

-    with open(f"{class_data_dir}/caption.txt", "w") as f1, open(f"{class_data_dir}/urls.txt", "w") as f2, open(
-        f"{class_data_dir}/images.txt", "w"
-    ) as f3:
+    with (
+        open(f"{class_data_dir}/caption.txt", "w") as f1,
+        open(f"{class_data_dir}/urls.txt", "w") as f2,
+        open(f"{class_data_dir}/images.txt", "w") as f3,
+    ):
        while total < num_class_images:
            images = class_images[count]
            count += 1
@@ -63,7 +63,7 @@ from diffusers.utils.import_utils import is_xformers_available


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

 logger = get_logger(__name__)

@@ -731,18 +731,18 @@ def main(args):
            if not class_images_dir.exists():
                class_images_dir.mkdir(parents=True, exist_ok=True)
            if args.real_prior:
-                assert (
-                    class_images_dir / "images"
-                ).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
-                assert (
-                    len(list((class_images_dir / "images").iterdir())) == args.num_class_images
-                ), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
-                assert (
-                    class_images_dir / "caption.txt"
-                ).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
-                assert (
-                    class_images_dir / "images.txt"
-                ).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
+                assert (class_images_dir / "images").exists(), (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
+                assert len(list((class_images_dir / "images").iterdir())) == args.num_class_images, (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
+                assert (class_images_dir / "caption.txt").exists(), (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
+                assert (class_images_dir / "images.txt").exists(), (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
                concept["class_prompt"] = os.path.join(class_images_dir, "caption.txt")
                concept["class_data_dir"] = os.path.join(class_images_dir, "images.txt")
                args.concepts_list[i] = concept
@@ -0,0 +1,119 @@
+# DreamBooth training example for HiDream Image
+
+[DreamBooth](https://arxiv.org/abs/2208.12242) is a method to personalize text2image models like stable diffusion given just a few (3~5) images of a subject.
+
+The `train_dreambooth_lora_hidream.py` script shows how to implement the training procedure with [LoRA](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) and adapt it for [HiDream Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/). 
+
+
+This will also allow us to push the trained model parameters to the Hugging Face Hub platform.
+
+## Running locally with PyTorch
+
+### Installing the dependencies
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+**Important**
+
+To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
+
+```bash
+git clone https://github.com/huggingface/diffusers
+cd diffusers
+pip install -e .
+```
+
+Then cd in the `examples/dreambooth` folder and run
+```bash
+pip install -r requirements_hidream.txt
+```
+
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+Or for a default accelerate configuration without answering questions about your environment
+
+```bash
+accelerate config default
+```
+
+Or if your environment doesn't support an interactive shell (e.g., a notebook)
+
+```python
+from accelerate.utils import write_basic_config
+write_basic_config()
+```
+
+When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.14.0` installed in your environment.
+
+
+### 3d icon example
+
+For this example we will use some 3d icon images: https://huggingface.co/datasets/linoyts/3d_icon.
+
+This will also allow us to push the trained LoRA parameters to the Hugging Face Hub platform.
+
+Now, we can launch training using:
+> [!NOTE]
+> The following training configuration prioritizes lower memory consumption by using gradient checkpointing, 
+> 8-bit Adam optimizer, latent caching, offloading, no validation.
+> all text embeddings are pre-computed to save memory.
+```bash
+export MODEL_NAME="HiDream-ai/HiDream-I1-Dev"
+export INSTANCE_DIR="linoyts/3d_icon"
+export OUTPUT_DIR="trained-hidream-lora"
+
+accelerate launch train_dreambooth_lora_hidream.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --dataset_name=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --mixed_precision="bf16" \
+  --instance_prompt="3d icon" \
+  --caption_column="prompt"\
+  --validation_prompt="a 3dicon, a llama eating ramen" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --use_8bit_adam \
+  --rank=8 \
+  --learning_rate=2e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant_with_warmup" \
+  --lr_warmup_steps=100 \
+  --max_train_steps=1000 \
+  --cache_latents\
+  --gradient_checkpointing \
+  --validation_epochs=25 \
+  --seed="0" \
+  --push_to_hub
+```
+
+For using `push_to_hub`, make you're logged into your Hugging Face account:
+
+```bash
+huggingface-cli login
+```
+
+To better track our training experiments, we're using the following flags in the command above:
+
+* `report_to="wandb` will ensure the training runs are tracked on [Weights and Biases](https://wandb.ai/site). To use it, be sure to install `wandb` with `pip install wandb`. Don't forget to call `wandb login <your_api_key>` before training if you haven't done it before.
+* `validation_prompt` and `validation_epochs` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
+
+## Notes
+
+Additionally, we welcome you to explore the following CLI arguments:
+
+* `--lora_layers`: The transformer modules to apply LoRA training on. Please specify the layers in a comma seperated. E.g. - "to_k,to_q,to_v" will result in lora training of attention layers only.
+* `--rank`: The rank of the LoRA layers. The higher the rank, the more parameters are trained. The default is 16.
+
+We provide several options for optimizing memory optimization:
+
+* `--offload`: When enabled, we will offload the text encoder and VAE to CPU, when they are not used.
+* `cache_latents`: When enabled, we will pre-compute the latents from the input images with the VAE and remove the VAE from memory once done.
+* `--use_8bit_adam`: When enabled, we will use the 8bit version of AdamW provided by the `bitsandbytes` library.
+
+Refer to the [official documentation](https://huggingface.co/docs/diffusers/main/en/api/pipelines/) of the `HiDreamImagePipeline` to know more about the model.
@@ -0,0 +1,8 @@
+accelerate>=1.4.0
+torchvision
+transformers>=4.50.0
+ftfy
+tensorboard
+Jinja2
+peft>=0.14.0
+sentencepiece
@@ -0,0 +1,220 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# 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 logging
+import os
+import sys
+import tempfile
+
+import safetensors
+
+
+sys.path.append("..")
+from test_examples_utils import ExamplesTestsAccelerate, run_command  # noqa: E402
+
+
+logging.basicConfig(level=logging.DEBUG)
+
+logger = logging.getLogger()
+stream_handler = logging.StreamHandler(sys.stdout)
+logger.addHandler(stream_handler)
+
+
+class DreamBoothLoRAHiDreamImage(ExamplesTestsAccelerate):
+    instance_data_dir = "docs/source/en/imgs"
+    pretrained_model_name_or_path = "hf-internal-testing/tiny-hidream-i1-pipe"
+    text_encoder_4_path = "hf-internal-testing/tiny-random-LlamaForCausalLM"
+    tokenizer_4_path = "hf-internal-testing/tiny-random-LlamaForCausalLM"
+    script_path = "examples/dreambooth/train_dreambooth_lora_hidream.py"
+    transformer_layer_type = "double_stream_blocks.0.block.attn1.to_k"
+
+    def test_dreambooth_lora_hidream(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --pretrained_text_encoder_4_name_or_path {self.text_encoder_4_path}
+                --pretrained_tokenizer_4_name_or_path {self.tokenizer_4_path}
+                --instance_data_dir {self.instance_data_dir}
+                --resolution 32
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                --max_sequence_length 16
+                """.split()
+
+            test_args.extend(["--instance_prompt", ""])
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            # when not training the text encoder, all the parameters in the state dict should start
+            # with `"transformer"` in their names.
+            starts_with_transformer = all(key.startswith("transformer") for key in lora_state_dict.keys())
+            self.assertTrue(starts_with_transformer)
+
+    def test_dreambooth_lora_latent_caching(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --pretrained_text_encoder_4_name_or_path {self.text_encoder_4_path}
+                --pretrained_tokenizer_4_name_or_path {self.tokenizer_4_path}
+                --instance_data_dir {self.instance_data_dir}
+                --resolution 32
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --cache_latents
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                --max_sequence_length 16
+                """.split()
+
+            test_args.extend(["--instance_prompt", ""])
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            # when not training the text encoder, all the parameters in the state dict should start
+            # with `"transformer"` in their names.
+            starts_with_transformer = all(key.startswith("transformer") for key in lora_state_dict.keys())
+            self.assertTrue(starts_with_transformer)
+
+    def test_dreambooth_lora_layers(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --pretrained_text_encoder_4_name_or_path {self.text_encoder_4_path}
+                --pretrained_tokenizer_4_name_or_path {self.tokenizer_4_path}
+                --instance_data_dir {self.instance_data_dir}
+                --resolution 32
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --cache_latents
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lora_layers {self.transformer_layer_type}
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                --max_sequence_length 16
+                """.split()
+
+            test_args.extend(["--instance_prompt", ""])
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            # when not training the text encoder, all the parameters in the state dict should start
+            # with `"transformer"` in their names. In this test, we only params of
+            # `self.transformer_layer_type` should be in the state dict.
+            starts_with_transformer = all(self.transformer_layer_type in key for key in lora_state_dict)
+            self.assertTrue(starts_with_transformer)
+
+    def test_dreambooth_lora_hidream_checkpointing_checkpoints_total_limit(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --pretrained_text_encoder_4_name_or_path {self.text_encoder_4_path}
+            --pretrained_tokenizer_4_name_or_path {self.tokenizer_4_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --resolution=32
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=6
+            --checkpoints_total_limit=2
+            --checkpointing_steps=2
+            --max_sequence_length 16
+            """.split()
+
+            test_args.extend(["--instance_prompt", ""])
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual(
+                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
+                {"checkpoint-4", "checkpoint-6"},
+            )
+
+    def test_dreambooth_lora_hidream_checkpointing_checkpoints_total_limit_removes_multiple_checkpoints(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --pretrained_text_encoder_4_name_or_path {self.text_encoder_4_path}
+            --pretrained_tokenizer_4_name_or_path {self.tokenizer_4_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --resolution=32
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=4
+            --checkpointing_steps=2
+            --max_sequence_length 16
+            """.split()
+
+            test_args.extend(["--instance_prompt", ""])
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-2", "checkpoint-4"})
+
+            resume_run_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --pretrained_text_encoder_4_name_or_path {self.text_encoder_4_path}
+            --pretrained_tokenizer_4_name_or_path {self.tokenizer_4_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --resolution=32
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=8
+            --checkpointing_steps=2
+            --resume_from_checkpoint=checkpoint-4
+            --checkpoints_total_limit=2
+            --max_sequence_length 16
+            """.split()
+
+            resume_run_args.extend(["--instance_prompt", ""])
+            run_command(self._launch_args + resume_run_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-6", "checkpoint-8"})
@@ -63,7 +63,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -1014,7 +1014,7 @@ def main(args):

    if args.train_text_encoder and unwrap_model(text_encoder).dtype != torch.float32:
        raise ValueError(
-            f"Text encoder loaded as datatype {unwrap_model(text_encoder).dtype}." f" {low_precision_error_string}"
+            f"Text encoder loaded as datatype {unwrap_model(text_encoder).dtype}. {low_precision_error_string}"
        )

    # Enable TF32 for faster training on Ampere GPUs,
@@ -35,7 +35,7 @@ from diffusers.utils import check_min_version


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

 # Cache compiled models across invocations of this script.
 cc.initialize_cache(os.path.expanduser("~/.cache/jax/compilation_cache"))
@@ -65,7 +65,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -618,6 +618,15 @@ def parse_args(input_args=None):
        ),
    )
    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+    parser.add_argument(
+        "--image_interpolation_mode",
+        type=str,
+        default="lanczos",
+        choices=[
+            f.lower() for f in dir(transforms.InterpolationMode) if not f.startswith("__") and not f.endswith("__")
+        ],
+        help="The image interpolation method to use for resizing images.",
+    )

    if input_args is not None:
        args = parser.parse_args(input_args)
@@ -737,7 +746,10 @@ class DreamBoothDataset(Dataset):
            self.instance_images.extend(itertools.repeat(img, repeats))

        self.pixel_values = []
-        train_resize = transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR)
+        interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
+        if interpolation is None:
+            raise ValueError(f"Unsupported interpolation mode {interpolation=}.")
+        train_resize = transforms.Resize(size, interpolation=interpolation)
        train_crop = transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size)
        train_flip = transforms.RandomHorizontalFlip(p=1.0)
        train_transforms = transforms.Compose(
@@ -895,7 +907,10 @@ def _encode_prompt_with_t5(

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

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

    _, seq_len, _ = prompt_embeds.shape
@@ -936,9 +951,13 @@ def _encode_prompt_with_clip(

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

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

    # duplicate text embeddings for each generation per prompt, using mps friendly method
    prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
@@ -958,7 +977,12 @@ def encode_prompt(
 ):
    prompt = [prompt] if isinstance(prompt, str) else prompt
    batch_size = len(prompt)
-    dtype = text_encoders[0].dtype
+
+    if hasattr(text_encoders[0], "module"):
+        dtype = text_encoders[0].module.dtype
+    else:
+        dtype = text_encoders[0].dtype
+
    device = device if device is not None else text_encoders[1].device
    pooled_prompt_embeds = _encode_prompt_with_clip(
        text_encoder=text_encoders[0],
@@ -1395,17 +1419,22 @@ def main(args):
                tokens_two = torch.cat([tokens_two, class_tokens_two], dim=0)

    # 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 * accelerator.num_processes * num_update_steps_per_epoch
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )
@@ -1432,8 +1461,14 @@ def main(args):

    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps:
+            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)

@@ -1590,7 +1625,7 @@ def main(args):
                )

                # handle guidance
-                if accelerator.unwrap_model(transformer).config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                    guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                    guidance = guidance.expand(model_input.shape[0])
                else:
@@ -1599,7 +1634,7 @@ def main(args):
                # Predict the noise residual
                model_pred = transformer(
                    hidden_states=packed_noisy_model_input,
-                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
+                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transformer model (we should not keep it but I want to keep the inputs same for the model for testing)
                    timestep=timesteps / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
@@ -1716,9 +1751,9 @@ def main(args):
                pipeline = FluxPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one, keep_fp32_wrapper=False),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two, keep_fp32_wrapper=False),
-                    transformer=accelerator.unwrap_model(transformer, keep_fp32_wrapper=False),
+                    text_encoder=unwrap_model(text_encoder_one, keep_fp32_wrapper=False),
+                    text_encoder_2=unwrap_model(text_encoder_two, keep_fp32_wrapper=False),
+                    transformer=unwrap_model(transformer, keep_fp32_wrapper=False),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
@@ -74,7 +74,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -524,6 +524,15 @@ def parse_args(input_args=None):
        default=4,
        help=("The dimension of the LoRA update matrices."),
    )
+    parser.add_argument(
+        "--image_interpolation_mode",
+        type=str,
+        default="lanczos",
+        choices=[
+            f.lower() for f in dir(transforms.InterpolationMode) if not f.startswith("__") and not f.endswith("__")
+        ],
+        help="The image interpolation method to use for resizing images.",
+    )

    if input_args is not None:
        args = parser.parse_args(input_args)
@@ -601,9 +610,13 @@ class DreamBoothDataset(Dataset):
        else:
            self.class_data_root = None

+        interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
+        if interpolation is None:
+            raise ValueError(f"Unsupported interpolation mode {interpolation=}.")
+
        self.image_transforms = transforms.Compose(
            [
-                transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
+                transforms.Resize(size, interpolation=interpolation),
                transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
                transforms.ToTensor(),
                transforms.Normalize([0.5], [0.5]),
@@ -982,7 +995,7 @@ def main(args):

        lora_state_dict, network_alphas = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")

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

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

 logger = get_logger(__name__)

@@ -177,16 +177,25 @@ def log_validation(
        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
        f" {args.validation_prompt}."
    )
-    pipeline = pipeline.to(accelerator.device)
+    pipeline = pipeline.to(accelerator.device, dtype=torch_dtype)
    pipeline.set_progress_bar_config(disable=True)

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

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

    for tracker in accelerator.trackers:
        phase_name = "test" if is_final_validation else "validation"
@@ -203,8 +212,7 @@ def log_validation(
            )

    del pipeline
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
+    free_memory()

    return images

@@ -932,7 +940,10 @@ def _encode_prompt_with_t5(

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

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

    _, seq_len, _ = prompt_embeds.shape
@@ -973,9 +984,13 @@ def _encode_prompt_with_clip(

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

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

    # duplicate text embeddings for each generation per prompt, using mps friendly method
    prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
@@ -994,7 +1009,11 @@ def encode_prompt(
    text_input_ids_list=None,
 ):
    prompt = [prompt] if isinstance(prompt, str) else prompt
-    dtype = text_encoders[0].dtype
+
+    if hasattr(text_encoders[0], "module"):
+        dtype = text_encoders[0].module.dtype
+    else:
+        dtype = text_encoders[0].dtype

    pooled_prompt_embeds = _encode_prompt_with_clip(
        text_encoder=text_encoders[0],
@@ -1275,7 +1294,7 @@ def main(args):
        lora_state_dict = FluxPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1505,17 +1524,22 @@ def main(args):
            free_memory()

    # 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 * accelerator.num_processes * num_update_steps_per_epoch
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )
@@ -1542,8 +1566,14 @@ def main(args):

    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps:
+            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)

@@ -1619,7 +1649,7 @@ def main(args):
        if args.train_text_encoder:
            text_encoder_one.train()
            # set top parameter requires_grad = True for gradient checkpointing works
-            accelerator.unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
+            unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)

        for step, batch in enumerate(train_dataloader):
            models_to_accumulate = [transformer]
@@ -1710,7 +1740,7 @@ def main(args):
                )

                # handle guidance
-                if accelerator.unwrap_model(transformer).config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                    guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                    guidance = guidance.expand(model_input.shape[0])
                else:
@@ -1719,7 +1749,7 @@ def main(args):
                # Predict the noise residual
                model_pred = transformer(
                    hidden_states=packed_noisy_model_input,
-                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
+                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transformer model (we should not keep it but I want to keep the inputs same for the model for testing)
                    timestep=timesteps / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
@@ -1828,9 +1858,9 @@ def main(args):
                pipeline = FluxPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
-                    transformer=accelerator.unwrap_model(transformer),
+                    text_encoder=unwrap_model(text_encoder_one),
+                    text_encoder_2=unwrap_model(text_encoder_two),
+                    transformer=unwrap_model(transformer),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
@@ -48,7 +48,7 @@ import diffusers
 from diffusers import (
    AutoencoderKL,
    FlowMatchEulerDiscreteScheduler,
-    Lumina2Text2ImgPipeline,
+    Lumina2Pipeline,
    Lumina2Transformer2DModel,
 )
 from diffusers.optimization import get_scheduler
@@ -72,7 +72,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -898,7 +898,7 @@ def main(args):
        cur_class_images = len(list(class_images_dir.iterdir()))

        if cur_class_images < args.num_class_images:
-            pipeline = Lumina2Text2ImgPipeline.from_pretrained(
+            pipeline = Lumina2Pipeline.from_pretrained(
                args.pretrained_model_name_or_path,
                torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16,
                revision=args.revision,
@@ -990,7 +990,7 @@ def main(args):
    text_encoder.to(dtype=torch.bfloat16)

    # Initialize a text encoding pipeline and keep it to CPU for now.
-    text_encoding_pipeline = Lumina2Text2ImgPipeline.from_pretrained(
+    text_encoding_pipeline = Lumina2Pipeline.from_pretrained(
        args.pretrained_model_name_or_path,
        vae=None,
        transformer=None,
@@ -1034,7 +1034,7 @@ def main(args):
                # make sure to pop weight so that corresponding model is not saved again
                weights.pop()

-            Lumina2Text2ImgPipeline.save_lora_weights(
+            Lumina2Pipeline.save_lora_weights(
                output_dir,
                transformer_lora_layers=transformer_lora_layers_to_save,
            )
@@ -1050,10 +1050,10 @@ def main(args):
            else:
                raise ValueError(f"unexpected save model: {model.__class__}")

-        lora_state_dict = Lumina2Text2ImgPipeline.lora_state_dict(input_dir)
+        lora_state_dict = Lumina2Pipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1473,7 +1473,7 @@ def main(args):
        if accelerator.is_main_process:
            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                # create pipeline
-                pipeline = Lumina2Text2ImgPipeline.from_pretrained(
+                pipeline = Lumina2Pipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    transformer=accelerator.unwrap_model(transformer),
                    revision=args.revision,
@@ -1503,14 +1503,14 @@ def main(args):
            transformer = transformer.to(weight_dtype)
        transformer_lora_layers = get_peft_model_state_dict(transformer)

-        Lumina2Text2ImgPipeline.save_lora_weights(
+        Lumina2Pipeline.save_lora_weights(
            save_directory=args.output_dir,
            transformer_lora_layers=transformer_lora_layers,
        )

        # Final inference
        # Load previous pipeline
-        pipeline = Lumina2Text2ImgPipeline.from_pretrained(
+        pipeline = Lumina2Pipeline.from_pretrained(
            args.pretrained_model_name_or_path,
            revision=args.revision,
            variant=args.variant,
@@ -71,7 +71,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -1064,7 +1064,7 @@ def main(args):
        lora_state_dict = SanaPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -72,7 +72,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -1355,7 +1355,7 @@ def main(args):
        lora_state_dict = StableDiffusion3Pipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -79,7 +79,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -118,7 +118,7 @@ def save_model_card(
            )

    model_description = f"""
-# {'SDXL' if 'playground' not in base_model else 'Playground'} LoRA DreamBooth - {repo_id}
+# {"SDXL" if "playground" not in base_model else "Playground"} LoRA DreamBooth - {repo_id}

 <Gallery />

@@ -669,6 +669,16 @@ def parse_args(input_args=None):
        ),
    )

+    parser.add_argument(
+        "--image_interpolation_mode",
+        type=str,
+        default="lanczos",
+        choices=[
+            f.lower() for f in dir(transforms.InterpolationMode) if not f.startswith("__") and not f.endswith("__")
+        ],
+        help="The image interpolation method to use for resizing images.",
+    )
+
    if input_args is not None:
        args = parser.parse_args(input_args)
    else:
@@ -790,7 +800,12 @@ class DreamBoothDataset(Dataset):
        self.original_sizes = []
        self.crop_top_lefts = []
        self.pixel_values = []
-        train_resize = transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR)
+
+        interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
+        if interpolation is None:
+            raise ValueError(f"Unsupported interpolation mode {interpolation=}.")
+        train_resize = transforms.Resize(size, interpolation=interpolation)
+
        train_crop = transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size)
        train_flip = transforms.RandomHorizontalFlip(p=1.0)
        train_transforms = transforms.Compose(
@@ -1271,7 +1286,7 @@ def main(args):

        lora_state_dict, network_alphas = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
        if incompatible_keys is not None:
@@ -1508,17 +1523,22 @@ def main(args):
                tokens_two = torch.cat([tokens_two, class_tokens_two], dim=0)

    # 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 * accelerator.num_processes * num_update_steps_per_epoch
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )
@@ -1535,7 +1555,14 @@ def main(args):

    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps:
+            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."
+            )
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    # Afterwards we recalculate our number of training epochs
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
@@ -63,7 +63,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -54,7 +54,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -57,7 +57,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__)

@@ -91,9 +91,9 @@ def log_validation(flux_transformer, args, accelerator, weight_dtype, step, is_f
            torch_dtype=weight_dtype,
        )
        pipeline.load_lora_weights(args.output_dir)
-        assert (
-            pipeline.transformer.config.in_channels == initial_channels * 2
-        ), f"{pipeline.transformer.config.in_channels=}"
+        assert pipeline.transformer.config.in_channels == initial_channels * 2, (
+            f"{pipeline.transformer.config.in_channels=}"
+        )

    pipeline.to(accelerator.device)
    pipeline.set_progress_bar_config(disable=True)
@@ -954,7 +954,7 @@ def main(args):

            lora_state_dict = FluxControlPipeline.lora_state_dict(input_dir)
            transformer_lora_state_dict = {
-                f'{k.replace("transformer.", "")}': v
+                f"{k.replace('transformer.', '')}": v
                for k, v in lora_state_dict.items()
                if k.startswith("transformer.") and "lora" in k
            }
@@ -49,6 +49,7 @@ from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionInstructPix2P
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import EMAModel
 from diffusers.utils import check_min_version, deprecate, is_wandb_available
+from diffusers.utils.constants import DIFFUSERS_REQUEST_TIMEOUT
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module

@@ -57,7 +58,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__, log_level="INFO")

@@ -418,7 +419,7 @@ def convert_to_np(image, resolution):


 def download_image(url):
-    image = PIL.Image.open(requests.get(url, stream=True).raw)
+    image = PIL.Image.open(requests.get(url, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)
    image = PIL.ImageOps.exif_transpose(image)
    image = image.convert("RGB")
    return image
@@ -60,7 +60,7 @@ if is_wandb_available():
    import wandb

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

 logger = get_logger(__name__, log_level="INFO")

--- a/Show More
+++ b/Show More