update the description of StableDiffusionXLDenoiseLoopWrapper

* update

* update

.github/workflows/nightly_tests.yml

@@ -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:

.github/workflows/release_tests_fast.yml

@@ -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

docker/diffusers-onnxruntime-cpu/Dockerfile

@@ -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 \

docs/source/en/_toctree.yml

@@ -175,7 +175,7 @@
     title: gguf
   - local: quantization/torchao
     title: torchao
-  - local: quantization/quanto 
+  - local: quantization/quanto
     title: quanto
   title: Quantization Methods
 - sections:
@@ -270,16 +270,18 @@
     - 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
@@ -288,30 +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
@@ -320,10 +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
@@ -338,10 +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
@@ -350,6 +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
@@ -366,10 +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
@@ -422,6 +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
@@ -446,6 +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
@@ -513,40 +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

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

@@ -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>
@@ -56,6 +60,9 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse
 ## Mochi1LoraLoaderMixin
 
 [[autodoc]] loaders.lora_pipeline.Mochi1LoraLoaderMixin
+## AuraFlowLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.AuraFlowLoraLoaderMixin
 
 ## LTXVideoLoraLoaderMixin
 
@@ -73,10 +80,22 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse
 
 [[autodoc]] loaders.lora_pipeline.Lumina2LoraLoaderMixin
 
+## CogView4LoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.CogView4LoraLoaderMixin
+
+## WanLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.WanLoraLoaderMixin
+
 ## AmusedLoraLoaderMixin
 
 [[autodoc]] loaders.lora_pipeline.AmusedLoraLoaderMixin
 
+## HiDreamImageLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.HiDreamImageLoraLoaderMixin
+
 ## LoraBaseMixin
 
 [[autodoc]] loaders.lora_base.LoraBaseMixin

docs/source/en/api/models/autoencoderkl_allegro.md

@@ -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

docs/source/en/api/models/controlnet_sana.md

@@ -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
+

docs/source/en/api/models/hidream_image_transformer.md

@@ -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

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

@@ -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

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

@@ -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

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

@@ -347,7 +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]

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

@@ -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

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

@@ -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

docs/source/en/community_projects.md

@@ -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>

docs/source/en/quantization/bitsandbytes.md

@@ -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,

docs/source/en/quantization/torchao.md

@@ -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)

docs/source/en/quicktour.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

docs/source/en/training/adapt_a_model.md

@@ -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,

docs/source/en/training/cogvideox.md

@@ -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.

docs/source/en/training/distributed_inference.md

@@ -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",

docs/source/en/training/dreambooth.md

@@ -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"

docs/source/en/training/t2i_adapters.md

@@ -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(

docs/source/en/tutorials/inference_with_big_models.md

@@ -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(

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

@@ -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:
 

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

@@ -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,

examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py

@@ -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__)
 
@@ -1915,17 +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,
     )
@@ -1949,7 +1954,6 @@ def main(args):
                 lr_scheduler,
             )
         else:
-            print("I SHOULD BE HERE")
             transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
                 transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler
             )
@@ -1961,8 +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)
 
@@ -2171,7 +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,

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -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__)
 

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -80,7 +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__)
 

examples/cogvideo/train_cogvideox_image_to_video_lora.py

@@ -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__)
 

examples/cogvideo/train_cogvideox_lora.py

@@ -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__)
 

examples/cogview4-control/train_control_cogview4.py

@@ -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__)
 

examples/community/README.md

@@ -86,6 +86,7 @@ PIXART-α Controlnet pipeline | Implementation of the controlnet model for pixar
 | 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) |
+| Stable Diffusion 3 InstructPix2Pix Pipeline | Implementation of Stable Diffusion 3 InstructPix2Pix Pipeline | [Stable Diffusion 3 InstructPix2Pix Pipeline](#stable-diffusion-3-instructpix2pix-pipeline) | [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/BleachNick/SD3_UltraEdit_freeform) [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/CaptainZZZ/sd3-instructpix2pix) | [Jiayu Zhang](https://github.com/xduzhangjiayu) and [Haozhe Zhao](https://github.com/HaozheZhao)|
 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
@@ -5381,7 +5382,7 @@ pipe = DiffusionPipeline.from_pretrained(
 # Here we need use pipeline internal unet model
 pipe.unet = pipe.unet_model.from_pretrained(model_id, subfolder="unet", variant="fp16", use_safetensors=True)
 
-# Load aditional layers to the model
+# Load additional layers to the model
 pipe.unet.load_additional_layers(weight_path="proc_data/faithdiff/FaithDiff.bin", dtype=dtype)
 
 # Enable vae tiling
@@ -5432,4 +5433,50 @@ 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)
+[<img src="https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/faithdiff_restored.PNG" width="512px" height="512px"/>](https://imgsli.com/MzY1NzE2)
+
+
+# Stable Diffusion 3 InstructPix2Pix Pipeline
+This the implementation of the Stable Diffusion 3 InstructPix2Pix Pipeline, based on the HuggingFace Diffusers.
+
+## Example Usage
+This pipeline aims to edit image based on user's instruction by using SD3
+````py
+import torch
+from diffusers import SD3Transformer2DModel
+from diffusers import DiffusionPipeline
+from diffusers.utils import load_image
+
+
+resolution = 512
+image = load_image("https://hf.co/datasets/diffusers/diffusers-images-docs/resolve/main/mountain.png").resize(
+    (resolution, resolution)
+)
+edit_instruction = "Turn sky into a sunny one"
+
+
+pipe = DiffusionPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-3-medium-diffusers", custom_pipeline="pipeline_stable_diffusion_3_instruct_pix2pix", torch_dtype=torch.float16).to('cuda')
+
+pipe.transformer = SD3Transformer2DModel.from_pretrained("CaptainZZZ/sd3-instructpix2pix",torch_dtype=torch.float16).to('cuda')
+
+edited_image = pipe(
+    prompt=edit_instruction,
+    image=image,
+    height=resolution,
+    width=resolution,
+    guidance_scale=7.5,
+    image_guidance_scale=1.5,
+    num_inference_steps=30,
+).images[0]
+
+edited_image.save("edited_image.png")
+````
+|Original|Edited|
+|---|---|
+|![Original image](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/StableDiffusion3InstructPix2Pix/mountain.png)|![Edited image](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/StableDiffusion3InstructPix2Pix/edited.png)
+
+### Note
+This model is trained on 512x512, so input size is better on 512x512.
+For better editing performance, please refer to this powerful model https://huggingface.co/BleachNick/SD3_UltraEdit_freeform and Paper "UltraEdit: Instruction-based Fine-Grained Image
+Editing at Scale", many thanks to their contribution!

examples/community/dps_pipeline.py

@@ -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.

examples/community/fresco_v2v.py

@@ -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)
     """
 

examples/community/hd_painter.py

@@ -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
 

examples/community/marigold_depth_estimation.py

@@ -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):

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py

@@ -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

examples/consistency_distillation/train_lcm_distill_lora_sd_wds.py

@@ -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__)
 
@@ -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.

examples/consistency_distillation/train_lcm_distill_lora_sdxl.py

@@ -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.

examples/consistency_distillation/train_lcm_distill_lora_sdxl_wds.py

@@ -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__)
 
@@ -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.

examples/consistency_distillation/train_lcm_distill_sd_wds.py

@@ -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.

examples/consistency_distillation/train_lcm_distill_sdxl_wds.py

@@ -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.

examples/controlnet/README_flux.md

@@ -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 \

examples/controlnet/train_controlnet.py

@@ -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__)
 

examples/controlnet/train_controlnet_flax.py

@@ -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__)
 

examples/controlnet/train_controlnet_flux.py

@@ -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]
@@ -639,6 +639,15 @@ def parse_args(input_args=None):
         action="store_true",
         help="Enable model cpu offload and save memory.",
     )
+    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)
@@ -736,9 +745,13 @@ def get_train_dataset(args, accelerator):
 
 
 def prepare_train_dataset(dataset, accelerator):
+    interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
+    if interpolation is None:
+        raise ValueError(f"Unsupported interpolation mode {interpolation=}.")
+
     image_transforms = transforms.Compose(
         [
-            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.Resize(args.resolution, interpolation=interpolation),
             transforms.CenterCrop(args.resolution),
             transforms.ToTensor(),
             transforms.Normalize([0.5], [0.5]),
@@ -747,7 +760,7 @@ def prepare_train_dataset(dataset, accelerator):
 
     conditioning_image_transforms = transforms.Compose(
         [
-            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.Resize(args.resolution, interpolation=interpolation),
             transforms.CenterCrop(args.resolution),
             transforms.ToTensor(),
             transforms.Normalize([0.5], [0.5]),
@@ -1085,8 +1098,6 @@ 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 +1114,8 @@ 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.

examples/controlnet/train_controlnet_sd3.py

@@ -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__)
 

examples/controlnet/train_controlnet_sdxl.py

@@ -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():
@@ -134,7 +134,25 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
 
     for validation_prompt, validation_image in zip(validation_prompts, validation_images):
         validation_image = Image.open(validation_image).convert("RGB")
-        validation_image = validation_image.resize((args.resolution, args.resolution))
+
+        try:
+            interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper())
+        except (AttributeError, KeyError):
+            supported_interpolation_modes = [
+                f.lower() for f in dir(transforms.InterpolationMode) if not f.startswith("__") and not f.endswith("__")
+            ]
+            raise ValueError(
+                f"Interpolation mode {args.image_interpolation_mode} is not supported. "
+                f"Please select one of the following: {', '.join(supported_interpolation_modes)}"
+            )
+
+        transform = transforms.Compose(
+            [
+                transforms.Resize(args.resolution, interpolation=interpolation),
+                transforms.CenterCrop(args.resolution),
+            ]
+        )
+        validation_image = transform(validation_image)
 
         images = []
 
@@ -587,6 +605,15 @@ def parse_args(input_args=None):
             " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
         ),
     )
+    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)
@@ -732,9 +759,20 @@ def encode_prompt(prompt_batch, text_encoders, tokenizers, proportion_empty_prom
 
 
 def prepare_train_dataset(dataset, accelerator):
+    try:
+        interpolation_mode = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper())
+    except (AttributeError, KeyError):
+        supported_interpolation_modes = [
+            f.lower() for f in dir(transforms.InterpolationMode) if not f.startswith("__") and not f.endswith("__")
+        ]
+        raise ValueError(
+            f"Interpolation mode {args.image_interpolation_mode} is not supported. "
+            f"Please select one of the following: {', '.join(supported_interpolation_modes)}"
+        )
+
     image_transforms = transforms.Compose(
         [
-            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.Resize(args.resolution, interpolation=interpolation_mode),
             transforms.CenterCrop(args.resolution),
             transforms.ToTensor(),
             transforms.Normalize([0.5], [0.5]),
@@ -743,7 +781,7 @@ def prepare_train_dataset(dataset, accelerator):
 
     conditioning_image_transforms = transforms.Compose(
         [
-            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.Resize(args.resolution, interpolation=interpolation_mode),
             transforms.CenterCrop(args.resolution),
             transforms.ToTensor(),
         ]

examples/custom_diffusion/train_custom_diffusion.py

@@ -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__)
 

examples/dreambooth/README_hidream.md

@@ -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.

examples/dreambooth/requirements_hidream.txt

@@ -0,0 +1,8 @@
+accelerate>=1.4.0
+torchvision
+transformers>=4.50.0
+ftfy
+tensorboard
+Jinja2
+peft>=0.14.0
+sentencepiece

examples/dreambooth/test_dreambooth_lora_hidream.py

@@ -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"})

examples/dreambooth/train_dreambooth.py

@@ -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__)
 

examples/dreambooth/train_dreambooth_flax.py

@@ -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"))

examples/dreambooth/train_dreambooth_flux.py

@@ -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(
@@ -1407,17 +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,
     )
@@ -1444,8 +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)
 
@@ -1611,7 +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,

examples/dreambooth/train_dreambooth_lora.py

@@ -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]),

examples/dreambooth/train_dreambooth_lora_flux.py

@@ -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__)
 
@@ -1524,17 +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,
     )
@@ -1561,8 +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)
 
@@ -1738,7 +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,

examples/dreambooth/train_dreambooth_lora_lumina2.py

@@ -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,
             )
@@ -1050,7 +1050,7 @@ 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.")
@@ -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,

examples/dreambooth/train_dreambooth_lora_sana.py

@@ -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__)
 

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -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__)
 

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -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__)
 
@@ -1523,17 +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,
     )
@@ -1550,7 +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)

examples/dreambooth/train_dreambooth_sd3.py

@@ -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__)
 

examples/flux-control/train_control_flux.py

@@ -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__)
 

examples/flux-control/train_control_lora_flux.py

@@ -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__)
 

examples/instruct_pix2pix/train_instruct_pix2pix.py

@@ -58,7 +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")
 

examples/instruct_pix2pix/train_instruct_pix2pix_sdxl.py

@@ -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")
 

examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py

@@ -52,7 +52,7 @@ if is_wandb_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__, log_level="INFO")
 

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py

@@ -46,7 +46,7 @@ from diffusers.utils import check_min_version, is_wandb_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__, log_level="INFO")
 

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py

@@ -46,7 +46,7 @@ from diffusers.utils import check_min_version, is_wandb_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__, log_level="INFO")
 

examples/kandinsky2_2/text_to_image/train_text_to_image_prior.py

@@ -51,7 +51,7 @@ if is_wandb_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__, log_level="INFO")
 

examples/research_projects/flux_lora_quantization/train_dreambooth_lora_flux_miniature.py

@@ -1088,7 +1088,7 @@ def main(args):
                 text_ids = batch["text_ids"].to(device=accelerator.device, dtype=weight_dtype)
                 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,

examples/t2i_adapter/train_t2i_adapter_sdxl.py

@@ -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__)
 

examples/text_to_image/train_text_to_image.py

@@ -57,7 +57,7 @@ if is_wandb_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__, log_level="INFO")
 
@@ -499,6 +499,15 @@ def parse_args():
             " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
         ),
     )
+    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.",
+    )
 
     args = parser.parse_args()
     env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
@@ -787,10 +796,17 @@ def main():
         )
         return inputs.input_ids
 
-    # Preprocessing the datasets.
+    # Get the specified interpolation method from the args
+    interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
+
+    # Raise an error if the interpolation method is invalid
+    if interpolation is None:
+        raise ValueError(f"Unsupported interpolation mode {args.image_interpolation_mode}.")
+
+    # Data preprocessing transformations
     train_transforms = transforms.Compose(
         [
-            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.Resize(args.resolution, interpolation=interpolation),  # Use dynamic interpolation method
             transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
             transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
             transforms.ToTensor(),

examples/text_to_image/train_text_to_image_flax.py

@@ -49,7 +49,7 @@ 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")
 
 logger = logging.getLogger(__name__)
 

examples/text_to_image/train_text_to_image_lora.py

@@ -56,7 +56,7 @@ 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")
 
@@ -418,6 +418,15 @@ def parse_args():
         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.",
+    )
 
     args = parser.parse_args()
     env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
@@ -649,10 +658,17 @@ def main():
         )
         return inputs.input_ids
 
-    # Preprocessing the datasets.
+    # Get the specified interpolation method from the args
+    interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
+
+    # Raise an error if the interpolation method is invalid
+    if interpolation is None:
+        raise ValueError(f"Unsupported interpolation mode {args.image_interpolation_mode}.")
+
+    # Data preprocessing transformations
     train_transforms = transforms.Compose(
         [
-            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.Resize(args.resolution, interpolation=interpolation),  # Use dynamic interpolation method
             transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
             transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
             transforms.ToTensor(),

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -68,7 +68,7 @@ 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():

examples/text_to_image/train_text_to_image_sdxl.py

@@ -55,7 +55,7 @@ from diffusers.utils.torch_utils import is_compiled_module
 
 
 # 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():

examples/textual_inversion/textual_inversion.py

@@ -81,7 +81,7 @@ else:
 
 
 # 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__)
 

examples/textual_inversion/textual_inversion_flax.py

@@ -56,7 +56,7 @@ else:
 # ------------------------------------------------------------------------------
 
 # 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__)
 

examples/textual_inversion/textual_inversion_sdxl.py

@@ -76,7 +76,7 @@ else:
 
 
 # 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__)
 

examples/unconditional_image_generation/train_unconditional.py

@@ -29,7 +29,7 @@ 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__, log_level="INFO")
 

examples/vqgan/train_vqgan.py

@@ -50,7 +50,7 @@ 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")
 

scripts/convert_sana_controlnet_to_diffusers.py

@@ -0,0 +1,216 @@
+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+from contextlib import nullcontext
+
+import torch
+from accelerate import init_empty_weights
+
+from diffusers import (
+    SanaControlNetModel,
+)
+from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.utils.import_utils import is_accelerate_available
+
+
+CTX = init_empty_weights if is_accelerate_available else nullcontext
+
+
+def main(args):
+    file_path = args.orig_ckpt_path
+
+    all_state_dict = torch.load(file_path, weights_only=True)
+    state_dict = all_state_dict.pop("state_dict")
+    converted_state_dict = {}
+
+    # Patch embeddings.
+    converted_state_dict["patch_embed.proj.weight"] = state_dict.pop("x_embedder.proj.weight")
+    converted_state_dict["patch_embed.proj.bias"] = state_dict.pop("x_embedder.proj.bias")
+
+    # Caption projection.
+    converted_state_dict["caption_projection.linear_1.weight"] = state_dict.pop("y_embedder.y_proj.fc1.weight")
+    converted_state_dict["caption_projection.linear_1.bias"] = state_dict.pop("y_embedder.y_proj.fc1.bias")
+    converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
+    converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
+
+    # AdaLN-single LN
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+        "t_embedder.mlp.0.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+        "t_embedder.mlp.2.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+    # Shared norm.
+    converted_state_dict["time_embed.linear.weight"] = state_dict.pop("t_block.1.weight")
+    converted_state_dict["time_embed.linear.bias"] = state_dict.pop("t_block.1.bias")
+
+    # y norm
+    converted_state_dict["caption_norm.weight"] = state_dict.pop("attention_y_norm.weight")
+
+    # Positional embedding interpolation scale.
+    interpolation_scale = {512: None, 1024: None, 2048: 1.0, 4096: 2.0}
+
+    # ControlNet Input Projection.
+    converted_state_dict["input_block.weight"] = state_dict.pop("controlnet.0.before_proj.weight")
+    converted_state_dict["input_block.bias"] = state_dict.pop("controlnet.0.before_proj.bias")
+
+    for depth in range(7):
+        # Transformer blocks.
+        converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.scale_shift_table"
+        )
+
+        # Linear Attention is all you need 🤘
+        # Self attention.
+        q, k, v = torch.chunk(state_dict.pop(f"controlnet.{depth}.copied_block.attn.qkv.weight"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.attn.proj.bias"
+        )
+
+        # Feed-forward.
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.inverted_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.inverted_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.depth_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.depth_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_point.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.point_conv.conv.weight"
+        )
+
+        # Cross-attention.
+        q = state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.q_linear.weight")
+        q_bias = state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.q_linear.bias")
+        k, v = torch.chunk(state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.kv_linear.weight"), 2, dim=0)
+        k_bias, v_bias = torch.chunk(
+            state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.kv_linear.bias"), 2, dim=0
+        )
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.cross_attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.cross_attn.proj.bias"
+        )
+
+        # ControlNet After Projection
+        converted_state_dict[f"controlnet_blocks.{depth}.weight"] = state_dict.pop(
+            f"controlnet.{depth}.after_proj.weight"
+        )
+        converted_state_dict[f"controlnet_blocks.{depth}.bias"] = state_dict.pop(f"controlnet.{depth}.after_proj.bias")
+
+    # ControlNet
+    with CTX():
+        controlnet = SanaControlNetModel(
+            num_attention_heads=model_kwargs[args.model_type]["num_attention_heads"],
+            attention_head_dim=model_kwargs[args.model_type]["attention_head_dim"],
+            num_layers=model_kwargs[args.model_type]["num_layers"],
+            num_cross_attention_heads=model_kwargs[args.model_type]["num_cross_attention_heads"],
+            cross_attention_head_dim=model_kwargs[args.model_type]["cross_attention_head_dim"],
+            cross_attention_dim=model_kwargs[args.model_type]["cross_attention_dim"],
+            caption_channels=2304,
+            sample_size=args.image_size // 32,
+            interpolation_scale=interpolation_scale[args.image_size],
+        )
+
+    if is_accelerate_available():
+        load_model_dict_into_meta(controlnet, converted_state_dict)
+    else:
+        controlnet.load_state_dict(converted_state_dict, strict=True, assign=True)
+
+    num_model_params = sum(p.numel() for p in controlnet.parameters())
+    print(f"Total number of controlnet parameters: {num_model_params}")
+
+    controlnet = controlnet.to(weight_dtype)
+    controlnet.load_state_dict(converted_state_dict, strict=True)
+
+    print(f"Saving Sana ControlNet in Diffusers format in {args.dump_path}.")
+    controlnet.save_pretrained(args.dump_path)
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+VARIANT_MAPPING = {
+    "fp32": None,
+    "fp16": "fp16",
+    "bf16": "bf16",
+}
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--orig_ckpt_path", default=None, type=str, required=True, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--image_size",
+        default=1024,
+        type=int,
+        choices=[512, 1024, 2048, 4096],
+        required=False,
+        help="Image size of pretrained model, 512, 1024, 2048 or 4096.",
+    )
+    parser.add_argument(
+        "--model_type",
+        default="SanaMS_1600M_P1_ControlNet_D7",
+        type=str,
+        choices=["SanaMS_1600M_P1_ControlNet_D7", "SanaMS_600M_P1_ControlNet_D7"],
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+    parser.add_argument("--dtype", default="fp16", type=str, choices=["fp32", "fp16", "bf16"], help="Weight dtype.")
+
+    args = parser.parse_args()
+
+    model_kwargs = {
+        "SanaMS_1600M_P1_ControlNet_D7": {
+            "num_attention_heads": 70,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 20,
+            "cross_attention_head_dim": 112,
+            "cross_attention_dim": 2240,
+            "num_layers": 7,
+        },
+        "SanaMS_600M_P1_ControlNet_D7": {
+            "num_attention_heads": 36,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 16,
+            "cross_attention_head_dim": 72,
+            "cross_attention_dim": 1152,
+            "num_layers": 7,
+        },
+    }
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    weight_dtype = DTYPE_MAPPING[args.dtype]
+    variant = VARIANT_MAPPING[args.dtype]
+
+    main(args)

scripts/convert_vae_pt_to_diffusers.py

@@ -53,7 +53,12 @@ def custom_convert_ldm_vae_checkpoint(checkpoint, config):
     }
 
     for i in range(num_down_blocks):
-        resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]
+        resnets = [
+            key
+            for key in down_blocks[i]
+            if f"down.{i}" in key and f"down.{i}.downsample" not in key and "attn" not in key
+        ]
+        attentions = [key for key in down_blocks[i] if f"down.{i}.attn" in key]
 
         if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
             new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop(
@@ -67,6 +72,10 @@ def custom_convert_ldm_vae_checkpoint(checkpoint, config):
         meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"}
         assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
 
+        paths = renew_vae_attention_paths(attentions)
+        meta_path = {"old": f"down.{i}.attn", "new": f"down_blocks.{i}.attentions"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
     mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
     num_mid_res_blocks = 2
     for i in range(1, num_mid_res_blocks + 1):
@@ -85,8 +94,11 @@ def custom_convert_ldm_vae_checkpoint(checkpoint, config):
     for i in range(num_up_blocks):
         block_id = num_up_blocks - 1 - i
         resnets = [
-            key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key
+            key
+            for key in up_blocks[block_id]
+            if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key and "attn" not in key
         ]
+        attentions = [key for key in up_blocks[block_id] if f"up.{block_id}.attn" in key]
 
         if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
             new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[
@@ -100,6 +112,10 @@ def custom_convert_ldm_vae_checkpoint(checkpoint, config):
         meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"}
         assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
 
+        paths = renew_vae_attention_paths(attentions)
+        meta_path = {"old": f"up.{block_id}.attn", "new": f"up_blocks.{i}.attentions"}
+        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
     mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
     num_mid_res_blocks = 2
     for i in range(1, num_mid_res_blocks + 1):

scripts/convert_wan_to_diffusers.py

@@ -39,6 +39,24 @@ TRANSFORMER_KEYS_RENAME_DICT = {
     "img_emb.proj.1": "condition_embedder.image_embedder.ff.net.0.proj",
     "img_emb.proj.3": "condition_embedder.image_embedder.ff.net.2",
     "img_emb.proj.4": "condition_embedder.image_embedder.norm2",
+    # for the FLF2V model
+    "img_emb.emb_pos": "condition_embedder.image_embedder.pos_embed",
+    # Add attention component mappings
+    "self_attn.q": "attn1.to_q",
+    "self_attn.k": "attn1.to_k",
+    "self_attn.v": "attn1.to_v",
+    "self_attn.o": "attn1.to_out.0",
+    "self_attn.norm_q": "attn1.norm_q",
+    "self_attn.norm_k": "attn1.norm_k",
+    "cross_attn.q": "attn2.to_q",
+    "cross_attn.k": "attn2.to_k",
+    "cross_attn.v": "attn2.to_v",
+    "cross_attn.o": "attn2.to_out.0",
+    "cross_attn.norm_q": "attn2.norm_q",
+    "cross_attn.norm_k": "attn2.norm_k",
+    "attn2.to_k_img": "attn2.add_k_proj",
+    "attn2.to_v_img": "attn2.add_v_proj",
+    "attn2.norm_k_img": "attn2.norm_added_k",
 }
 
 TRANSFORMER_SPECIAL_KEYS_REMAP = {}
@@ -135,6 +153,28 @@ def get_transformer_config(model_type: str) -> Dict[str, Any]:
                 "text_dim": 4096,
             },
         }
+    elif model_type == "Wan-FLF2V-14B-720P":
+        config = {
+            "model_id": "ypyp/Wan2.1-FLF2V-14B-720P",  # This is just a placeholder
+            "diffusers_config": {
+                "image_dim": 1280,
+                "added_kv_proj_dim": 5120,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 40,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "rope_max_seq_len": 1024,
+                "pos_embed_seq_len": 257 * 2,
+            },
+        }
     return config
 
 
@@ -393,11 +433,12 @@ if __name__ == "__main__":
     vae = convert_vae()
     text_encoder = UMT5EncoderModel.from_pretrained("google/umt5-xxl")
     tokenizer = AutoTokenizer.from_pretrained("google/umt5-xxl")
+    flow_shift = 16.0 if "FLF2V" in args.model_type else 3.0
     scheduler = UniPCMultistepScheduler(
-        prediction_type="flow_prediction", use_flow_sigmas=True, num_train_timesteps=1000, flow_shift=3.0
+        prediction_type="flow_prediction", use_flow_sigmas=True, num_train_timesteps=1000, flow_shift=flow_shift
     )
 
-    if "I2V" in args.model_type:
+    if "I2V" in args.model_type or "FLF2V" in args.model_type:
         image_encoder = CLIPVisionModelWithProjection.from_pretrained(
             "laion/CLIP-ViT-H-14-laion2B-s32B-b79K", torch_dtype=torch.bfloat16
         )

setup.py

@@ -269,7 +269,7 @@ version_range_max = max(sys.version_info[1], 10) + 1
 
 setup(
     name="diffusers",
-    version="0.33.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="0.34.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     description="State-of-the-art diffusion in PyTorch and JAX.",
     long_description=open("README.md", "r", encoding="utf-8").read(),
     long_description_content_type="text/markdown",

src/diffusers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.33.0.dev0"
+__version__ = "0.34.0.dev0"
 
 from typing import TYPE_CHECKING
 
@@ -34,10 +34,12 @@ from .utils import (
 
 _import_structure = {
     "configuration_utils": ["ConfigMixin"],
+    "guiders": [],
     "hooks": [],
     "loaders": ["FromOriginalModelMixin"],
     "models": [],
     "pipelines": [],
+    "modular_pipelines": [],
     "quantizers.quantization_config": [],
     "schedulers": [],
     "utils": [
@@ -130,12 +132,26 @@ except OptionalDependencyNotAvailable:
     _import_structure["utils.dummy_pt_objects"] = [name for name in dir(dummy_pt_objects) if not name.startswith("_")]
 
 else:
+    _import_structure["guiders"].extend(
+        [
+            "AdaptiveProjectedGuidance",
+            "AutoGuidance",
+            "ClassifierFreeGuidance",
+            "ClassifierFreeZeroStarGuidance",
+            "SkipLayerGuidance",
+            "SmoothedEnergyGuidance",
+            "TangentialClassifierFreeGuidance",
+        ]
+    )
     _import_structure["hooks"].extend(
         [
             "FasterCacheConfig",
             "HookRegistry",
             "PyramidAttentionBroadcastConfig",
+            "LayerSkipConfig",
+            "SmoothedEnergyGuidanceConfig",
             "apply_faster_cache",
+            "apply_layer_skip",
             "apply_pyramid_attention_broadcast",
         ]
     )
@@ -171,6 +187,7 @@ else:
             "FluxControlNetModel",
             "FluxMultiControlNetModel",
             "FluxTransformer2DModel",
+            "HiDreamImageTransformer2DModel",
             "HunyuanDiT2DControlNetModel",
             "HunyuanDiT2DModel",
             "HunyuanDiT2DMultiControlNetModel",
@@ -189,6 +206,7 @@ else:
             "OmniGenTransformer2DModel",
             "PixArtTransformer2DModel",
             "PriorTransformer",
+            "SanaControlNetModel",
             "SanaTransformer2DModel",
             "SD3ControlNetModel",
             "SD3MultiControlNetModel",
@@ -243,6 +261,15 @@ else:
             "StableDiffusionMixin",
         ]
     )
+    _import_structure["modular_pipelines"].extend(
+        [
+            "ModularLoader",
+            "ModularPipeline",
+            "ModularPipelineBlocks",
+            "ComponentSpec",
+            "ComponentsManager",
+        ]
+    )
     _import_structure["quantizers"] = ["DiffusersQuantizer"]
     _import_structure["schedulers"].extend(
         [
@@ -266,6 +293,7 @@ else:
             "EulerDiscreteScheduler",
             "FlowMatchEulerDiscreteScheduler",
             "FlowMatchHeunDiscreteScheduler",
+            "FlowMatchLCMScheduler",
             "HeunDiscreteScheduler",
             "IPNDMScheduler",
             "KarrasVeScheduler",
@@ -368,6 +396,7 @@ else:
             "FluxInpaintPipeline",
             "FluxPipeline",
             "FluxPriorReduxPipeline",
+            "HiDreamImagePipeline",
             "HunyuanDiTControlNetPipeline",
             "HunyuanDiTPAGPipeline",
             "HunyuanDiTPipeline",
@@ -426,6 +455,7 @@ else:
             "PixArtSigmaPAGPipeline",
             "PixArtSigmaPipeline",
             "ReduxImageEncoder",
+            "SanaControlNetPipeline",
             "SanaPAGPipeline",
             "SanaPipeline",
             "SanaSprintPipeline",
@@ -518,6 +548,24 @@ else:
         ]
     )
 
+
+try:
+    if not (is_torch_available() and is_transformers_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_transformers_objects"] = [
+        name for name in dir(dummy_torch_and_transformers_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["modular_pipelines"].extend(
+        [
+            "StableDiffusionXLAutoPipeline",
+            "StableDiffusionXLModularLoader",
+        ]
+    )
 try:
     if not (is_torch_available() and is_transformers_available() and is_opencv_available()):
         raise OptionalDependencyNotAvailable()
@@ -723,10 +771,22 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from .utils.dummy_pt_objects import *  # noqa F403
     else:
+        from .guiders import (
+            AdaptiveProjectedGuidance,
+            AutoGuidance,
+            ClassifierFreeGuidance,
+            ClassifierFreeZeroStarGuidance,
+            SkipLayerGuidance,
+            SmoothedEnergyGuidance,
+            TangentialClassifierFreeGuidance,
+        )
         from .hooks import (
             FasterCacheConfig,
             HookRegistry,
+            LayerSkipConfig,
             PyramidAttentionBroadcastConfig,
+            SmoothedEnergyGuidanceConfig,
+            apply_layer_skip,
             apply_faster_cache,
             apply_pyramid_attention_broadcast,
         )
@@ -761,6 +821,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             FluxControlNetModel,
             FluxMultiControlNetModel,
             FluxTransformer2DModel,
+            HiDreamImageTransformer2DModel,
             HunyuanDiT2DControlNetModel,
             HunyuanDiT2DModel,
             HunyuanDiT2DMultiControlNetModel,
@@ -779,6 +840,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             OmniGenTransformer2DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
+            SanaControlNetModel,
             SanaTransformer2DModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,
@@ -832,6 +894,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ScoreSdeVePipeline,
             StableDiffusionMixin,
         )
+        from .modular_pipelines import (
+            ModularLoader,
+            ModularPipeline,
+            ModularPipelineBlocks,
+            ComponentSpec,
+            ComponentsManager,
+        )
         from .quantizers import DiffusersQuantizer
         from .schedulers import (
             AmusedScheduler,
@@ -854,6 +923,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             EulerDiscreteScheduler,
             FlowMatchEulerDiscreteScheduler,
             FlowMatchHeunDiscreteScheduler,
+            FlowMatchLCMScheduler,
             HeunDiscreteScheduler,
             IPNDMScheduler,
             KarrasVeScheduler,
@@ -937,6 +1007,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             FluxInpaintPipeline,
             FluxPipeline,
             FluxPriorReduxPipeline,
+            HiDreamImagePipeline,
             HunyuanDiTControlNetPipeline,
             HunyuanDiTPAGPipeline,
             HunyuanDiTPipeline,
@@ -995,6 +1066,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PixArtSigmaPAGPipeline,
             PixArtSigmaPipeline,
             ReduxImageEncoder,
+            SanaControlNetPipeline,
             SanaPAGPipeline,
             SanaPipeline,
             SanaSprintPipeline,
@@ -1084,7 +1156,16 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             WuerstchenDecoderPipeline,
             WuerstchenPriorPipeline,
         )
-
+    try:
+        if not (is_torch_available() and is_transformers_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .modular_pipelines import (
+            StableDiffusionXLAutoPipeline,
+            StableDiffusionXLModularLoader,
+        )
     try:
         if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
             raise OptionalDependencyNotAvailable()

src/diffusers/guiders/__init__.py

@@ -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.
+
+from typing import Union
+
+from ..utils import is_torch_available
+
+
+if is_torch_available():
+    from .adaptive_projected_guidance import AdaptiveProjectedGuidance
+    from .auto_guidance import AutoGuidance
+    from .classifier_free_guidance import ClassifierFreeGuidance
+    from .classifier_free_zero_star_guidance import ClassifierFreeZeroStarGuidance
+    from .skip_layer_guidance import SkipLayerGuidance
+    from .smoothed_energy_guidance import SmoothedEnergyGuidance
+    from .tangential_classifier_free_guidance import TangentialClassifierFreeGuidance
+
+    GuiderType = Union[AdaptiveProjectedGuidance, AutoGuidance, ClassifierFreeGuidance, ClassifierFreeZeroStarGuidance, SkipLayerGuidance, SmoothedEnergyGuidance, TangentialClassifierFreeGuidance]

src/diffusers/guiders/adaptive_projected_guidance.py

@@ -0,0 +1,184 @@
+# 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.
+
+import math
+from typing import Optional, List, TYPE_CHECKING, Dict, Union, Tuple
+
+import torch
+
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class AdaptiveProjectedGuidance(BaseGuidance):
+    """
+    Adaptive Projected Guidance (APG): https://huggingface.co/papers/2410.02416
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        adaptive_projected_guidance_momentum (`float`, defaults to `None`):
+            The momentum parameter for the adaptive projected guidance. Disabled if set to `None`.
+        adaptive_projected_guidance_rescale (`float`, defaults to `15.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.0`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `1.0`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond"]
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        adaptive_projected_guidance_momentum: Optional[float] = None,
+        adaptive_projected_guidance_rescale: float = 15.0,
+        eta: float = 1.0,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+        start: float = 0.0,
+        stop: float = 1.0,
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.adaptive_projected_guidance_momentum = adaptive_projected_guidance_momentum
+        self.adaptive_projected_guidance_rescale = adaptive_projected_guidance_rescale
+        self.eta = eta
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+        self.momentum_buffer = None
+
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        if self._step == 0:
+            if self.adaptive_projected_guidance_momentum is not None:
+                self.momentum_buffer = MomentumBuffer(self.adaptive_projected_guidance_momentum)
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor:
+        pred = None
+
+        if not self._is_apg_enabled():
+            pred = pred_cond
+        else:
+            pred = normalized_guidance(
+                pred_cond,
+                pred_uncond,
+                self.guidance_scale,
+                self.momentum_buffer,
+                self.eta,
+                self.adaptive_projected_guidance_rescale,
+                self.use_original_formulation,
+            )
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred, {}
+
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_apg_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_apg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close
+
+
+class MomentumBuffer:
+    def __init__(self, momentum: float):
+        self.momentum = momentum
+        self.running_average = 0
+
+    def update(self, update_value: torch.Tensor):
+        new_average = self.momentum * self.running_average
+        self.running_average = update_value + new_average
+
+
+def normalized_guidance(
+    pred_cond: torch.Tensor,
+    pred_uncond: torch.Tensor,
+    guidance_scale: float,
+    momentum_buffer: Optional[MomentumBuffer] = None,
+    eta: float = 1.0,
+    norm_threshold: float = 0.0,
+    use_original_formulation: bool = False,
+):
+    diff = pred_cond - pred_uncond
+    dim = [-i for i in range(1, len(diff.shape))]
+    
+    if momentum_buffer is not None:
+        momentum_buffer.update(diff)
+        diff = momentum_buffer.running_average
+    
+    if norm_threshold > 0:
+        ones = torch.ones_like(diff)
+        diff_norm = diff.norm(p=2, dim=dim, keepdim=True)
+        scale_factor = torch.minimum(ones, norm_threshold / diff_norm)
+        diff = diff * scale_factor
+    
+    v0, v1 = diff.double(), pred_cond.double()
+    v1 = torch.nn.functional.normalize(v1, dim=dim)
+    v0_parallel = (v0 * v1).sum(dim=dim, keepdim=True) * v1
+    v0_orthogonal = v0 - v0_parallel
+    diff_parallel, diff_orthogonal = v0_parallel.type_as(diff), v0_orthogonal.type_as(diff)
+    normalized_update = diff_orthogonal + eta * diff_parallel
+    
+    pred = pred_cond if use_original_formulation else pred_uncond
+    pred = pred + guidance_scale * normalized_update
+    
+    return pred

src/diffusers/guiders/auto_guidance.py

@@ -0,0 +1,177 @@
+# 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.
+
+import math
+from typing import List, Optional, Union, TYPE_CHECKING, Dict, Tuple
+
+import torch
+
+from ..hooks import HookRegistry, LayerSkipConfig
+from ..hooks.layer_skip import _apply_layer_skip_hook
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class AutoGuidance(BaseGuidance):
+    """
+    AutoGuidance: https://huggingface.co/papers/2406.02507
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        auto_guidance_layers (`int` or `List[int]`, *optional*):
+            The layer indices to apply skip layer guidance to. Can be a single integer or a list of integers. If not
+            provided, `skip_layer_config` must be provided.
+        auto_guidance_config (`LayerSkipConfig` or `List[LayerSkipConfig]`, *optional*):
+            The configuration for the skip layer guidance. Can be a single `LayerSkipConfig` or a list of
+            `LayerSkipConfig`. If not provided, `skip_layer_guidance_layers` must be provided.
+        dropout (`float`, *optional*):
+            The dropout probability for autoguidance on the enabled skip layers (either with `auto_guidance_layers` or
+            `auto_guidance_config`). If not provided, the dropout probability will be set to 1.0.
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.0`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `1.0`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond"]
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        auto_guidance_layers: Optional[Union[int, List[int]]] = None,
+        auto_guidance_config: Union[LayerSkipConfig, List[LayerSkipConfig]] = None,
+        dropout: Optional[float] = None,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+        start: float = 0.0,
+        stop: float = 1.0,
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.auto_guidance_layers = auto_guidance_layers
+        self.auto_guidance_config = auto_guidance_config
+        self.dropout = dropout
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+
+        if auto_guidance_layers is None and auto_guidance_config is None:
+            raise ValueError(
+                "Either `auto_guidance_layers` or `auto_guidance_config` must be provided to enable Skip Layer Guidance."
+            )
+        if auto_guidance_layers is not None and auto_guidance_config is not None:
+            raise ValueError("Only one of `auto_guidance_layers` or `auto_guidance_config` can be provided.")
+        if (dropout is None and auto_guidance_layers is not None) or (dropout is not None and auto_guidance_layers is None):
+            raise ValueError("`dropout` must be provided if `auto_guidance_layers` is provided.")
+
+        if auto_guidance_layers is not None:
+            if isinstance(auto_guidance_layers, int):
+                auto_guidance_layers = [auto_guidance_layers]
+            if not isinstance(auto_guidance_layers, list):
+                raise ValueError(
+                    f"Expected `auto_guidance_layers` to be an int or a list of ints, but got {type(auto_guidance_layers)}."
+                )
+            auto_guidance_config = [LayerSkipConfig(layer, fqn="auto", dropout=dropout) for layer in auto_guidance_layers]
+
+        if isinstance(auto_guidance_config, LayerSkipConfig):
+            auto_guidance_config = [auto_guidance_config]
+
+        if not isinstance(auto_guidance_config, list):
+            raise ValueError(
+                f"Expected `auto_guidance_config` to be a LayerSkipConfig or a list of LayerSkipConfig, but got {type(auto_guidance_config)}."
+            )
+
+        self.auto_guidance_config = auto_guidance_config
+        self._auto_guidance_hook_names = [f"AutoGuidance_{i}" for i in range(len(self.auto_guidance_config))]
+
+    def prepare_models(self, denoiser: torch.nn.Module) -> None:
+        self._count_prepared += 1
+        if self._is_ag_enabled() and self.is_unconditional:
+            for name, config in zip(self._auto_guidance_hook_names, self.auto_guidance_config):
+                _apply_layer_skip_hook(denoiser, config, name=name)
+    
+    def cleanup_models(self, denoiser: torch.nn.Module) -> None:
+        if self._is_ag_enabled() and self.is_unconditional:
+            for name in self._auto_guidance_hook_names:
+                registry = HookRegistry.check_if_exists_or_initialize(denoiser)
+                registry.remove_hook(name, recurse=True)
+    
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor:
+        pred = None
+
+        if not self._is_ag_enabled():
+            pred = pred_cond
+        else:
+            shift = pred_cond - pred_uncond
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+        
+        return pred, {}
+    
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_ag_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_ag_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close

src/diffusers/guiders/classifier_free_guidance.py

@@ -0,0 +1,132 @@
+# 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.
+
+import math
+from typing import Optional, List, TYPE_CHECKING, Dict, Union, Tuple
+
+import torch
+
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class ClassifierFreeGuidance(BaseGuidance):
+    """
+    Classifier-free guidance (CFG): https://huggingface.co/papers/2207.12598
+    
+    CFG is a technique used to improve generation quality and condition-following in diffusion models. It works by
+    jointly training a model on both conditional and unconditional data, and using a weighted sum of the two during
+    inference. This allows the model to tradeoff between generation quality and sample diversity.
+    The original paper proposes scaling and shifting the conditional distribution based on the difference between
+    conditional and unconditional predictions. [x_pred = x_cond + scale * (x_cond - x_uncond)]
+    
+    Diffusers implemented the scaling and shifting on the unconditional prediction instead based on the [Imagen
+    paper](https://huggingface.co/papers/2205.11487), which is equivalent to what the original paper proposed in
+    theory. [x_pred = x_uncond + scale * (x_cond - x_uncond)]
+    
+    The intution behind the original formulation can be thought of as moving the conditional distribution estimates
+    further away from the unconditional distribution estimates, while the diffusers-native implementation can be
+    thought of as moving the unconditional distribution towards the conditional distribution estimates to get rid of
+    the unconditional predictions (usually negative features like "bad quality, bad anotomy, watermarks", etc.)
+    
+    The `use_original_formulation` argument can be set to `True` to use the original CFG formulation mentioned in the
+    paper. By default, we use the diffusers-native implementation that has been in the codebase for a long time.
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.0`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `1.0`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond"]
+
+    def __init__(
+        self, guidance_scale: float = 7.5, guidance_rescale: float = 0.0, use_original_formulation: bool = False, start: float = 0.0, stop: float = 1.0
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+    
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor:
+        pred = None
+
+        if not self._is_cfg_enabled():
+            pred = pred_cond
+        else:
+            shift = pred_cond - pred_uncond
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred, {}
+
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_cfg_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_cfg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close

src/diffusers/guiders/classifier_free_zero_star_guidance.py

@@ -0,0 +1,148 @@
+# 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.
+
+import math
+from typing import Optional, List, TYPE_CHECKING, Dict, Union, Tuple
+
+import torch
+
+from .guider_utils import BaseGuidance, rescale_noise_cfg
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+class ClassifierFreeZeroStarGuidance(BaseGuidance):
+    """
+    Classifier-free Zero* (CFG-Zero*): https://huggingface.co/papers/2503.18886
+    
+    This is an implementation of the Classifier-Free Zero* guidance technique, which is a variant of classifier-free
+    guidance. It proposes zero initialization of the noise predictions for the first few steps of the diffusion
+    process, and also introduces an optimal rescaling factor for the noise predictions, which can help in improving the
+    quality of generated images.
+    
+    The authors of the paper suggest setting zero initialization in the first 4% of the inference steps.
+    
+    Args:
+        guidance_scale (`float`, defaults to `7.5`):
+            The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text
+            prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and
+            deterioration of image quality.
+        zero_init_steps (`int`, defaults to `1`):
+            The number of inference steps for which the noise predictions are zeroed out (see Section 4.2).
+        guidance_rescale (`float`, defaults to `0.0`):
+            The rescale factor applied to the noise predictions. This is used to improve image quality and fix
+            overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+            Flawed](https://huggingface.co/papers/2305.08891).
+        use_original_formulation (`bool`, defaults to `False`):
+            Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default,
+            we use the diffusers-native implementation that has been in the codebase for a long time. See
+            [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details.
+        start (`float`, defaults to `0.01`):
+            The fraction of the total number of denoising steps after which guidance starts.
+        stop (`float`, defaults to `0.2`):
+            The fraction of the total number of denoising steps after which guidance stops.
+    """
+
+    _input_predictions = ["pred_cond", "pred_uncond"]
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        zero_init_steps: int = 1,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+        start: float = 0.0,
+        stop: float = 1.0,
+    ):
+        super().__init__(start, stop)
+
+        self.guidance_scale = guidance_scale
+        self.zero_init_steps = zero_init_steps
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+    
+    def prepare_inputs(self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None) -> List["BlockState"]:
+        
+        if input_fields is None:
+            input_fields = self._input_fields
+        
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for i in range(self.num_conditions):
+            data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i])
+            data_batches.append(data_batch)
+        return data_batches
+
+    def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor:
+        pred = None
+
+        if self._step < self.zero_init_steps:
+            pred = torch.zeros_like(pred_cond)
+        elif not self._is_cfg_enabled():
+            pred = pred_cond
+        else:
+            pred_cond_flat = pred_cond.flatten(1)
+            pred_uncond_flat = pred_uncond.flatten(1)
+            alpha = cfg_zero_star_scale(pred_cond_flat, pred_uncond_flat)
+            alpha = alpha.view(-1, *(1,) * (len(pred_cond.shape) - 1))
+            pred_uncond = pred_uncond * alpha
+            shift = pred_cond - pred_uncond
+            pred = pred_cond if self.use_original_formulation else pred_uncond
+            pred = pred + self.guidance_scale * shift
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred, {}
+    
+    @property
+    def is_conditional(self) -> bool:
+        return self._count_prepared == 1
+
+    @property
+    def num_conditions(self) -> int:
+        num_conditions = 1
+        if self._is_cfg_enabled():
+            num_conditions += 1
+        return num_conditions
+
+    def _is_cfg_enabled(self) -> bool:
+        if not self._enabled:
+            return False
+        
+        is_within_range = True
+        if self._num_inference_steps is not None:
+            skip_start_step = int(self._start * self._num_inference_steps)
+            skip_stop_step = int(self._stop * self._num_inference_steps)
+            is_within_range = skip_start_step <= self._step < skip_stop_step
+        
+        is_close = False
+        if self.use_original_formulation:
+            is_close = math.isclose(self.guidance_scale, 0.0)
+        else:
+            is_close = math.isclose(self.guidance_scale, 1.0)
+        
+        return is_within_range and not is_close
+
+
+def cfg_zero_star_scale(cond: torch.Tensor, uncond: torch.Tensor, eps: float = 1e-8) -> torch.Tensor:
+    cond_dtype = cond.dtype
+    cond = cond.float()
+    uncond = uncond.float()
+    dot_product = torch.sum(cond * uncond, dim=1, keepdim=True)
+    squared_norm = torch.sum(uncond**2, dim=1, keepdim=True) + eps
+    # st_star = v_cond^T * v_uncond / ||v_uncond||^2
+    scale = dot_product / squared_norm
+    return scale.to(dtype=cond_dtype)

src/diffusers/guiders/guider_utils.py

@@ -0,0 +1,215 @@
+# 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.
+
+from typing import TYPE_CHECKING, Any, Dict, List, Tuple, Union
+
+import torch
+
+from ..utils import get_logger
+
+
+if TYPE_CHECKING:
+    from ..modular_pipelines.modular_pipeline import BlockState
+
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class BaseGuidance:
+    r"""Base class providing the skeleton for implementing guidance techniques."""
+
+    _input_predictions = None
+    _identifier_key = "__guidance_identifier__"
+
+    def __init__(self, start: float = 0.0, stop: float = 1.0):
+        self._start = start
+        self._stop = stop
+        self._step: int = None
+        self._num_inference_steps: int = None
+        self._timestep: torch.LongTensor = None
+        self._count_prepared = 0
+        self._input_fields: Dict[str, Union[str, Tuple[str, str]]] = None
+        self._enabled = True
+
+        if not (0.0 <= start < 1.0):
+            raise ValueError(
+                f"Expected `start` to be between 0.0 and 1.0, but got {start}."
+            )
+        if not (start <= stop <= 1.0):
+            raise ValueError(
+                f"Expected `stop` to be between {start} and 1.0, but got {stop}."
+            )
+
+        if self._input_predictions is None or not isinstance(self._input_predictions, list):
+            raise ValueError(
+                "`_input_predictions` must be a list of required prediction names for the guidance technique."
+            )
+
+    def disable(self):
+        self._enabled = False
+    
+    def enable(self):
+        self._enabled = True
+    
+    def set_state(self, step: int, num_inference_steps: int, timestep: torch.LongTensor) -> None:
+        self._step = step
+        self._num_inference_steps = num_inference_steps
+        self._timestep = timestep
+        self._count_prepared = 0
+
+    def set_input_fields(self, **kwargs: Dict[str, Union[str, Tuple[str, str]]]) -> None:
+        """
+        Set the input fields for the guidance technique. The input fields are used to specify the names of the
+        returned attributes containing the prepared data after `prepare_inputs` is called. The prepared data is
+        obtained from the values of the provided keyword arguments to this method.
+
+        Args:
+            **kwargs (`Dict[str, Union[str, Tuple[str, str]]]`):
+                A dictionary where the keys are the names of the fields that will be used to store the data once
+                it is prepared with `prepare_inputs`. The values can be either a string or a tuple of length 2,
+                which is used to look up the required data provided for preparation.
+
+                If a string is provided, it will be used as the conditional data (or unconditional if used with
+                a guidance method that requires it). If a tuple of length 2 is provided, the first element must
+                be the conditional data identifier and the second element must be the unconditional data identifier
+                or None.
+
+                Example:
+                
+                ```
+                data = {"prompt_embeds": <some tensor>, "negative_prompt_embeds": <some tensor>, "latents": <some tensor>}
+
+                BaseGuidance.set_input_fields(
+                    latents="latents",
+                    prompt_embeds=("prompt_embeds", "negative_prompt_embeds"),
+                )
+                ```
+        """
+        for key, value in kwargs.items():
+            is_string = isinstance(value, str)
+            is_tuple_of_str_with_len_2 = isinstance(value, tuple) and len(value) == 2 and all(isinstance(v, str) for v in value)
+            if not (is_string or is_tuple_of_str_with_len_2):
+                raise ValueError(
+                    f"Expected `set_input_fields` to be called with a string or a tuple of string with length 2, but got {type(value)} for key {key}."
+                )
+        self._input_fields = kwargs
+    
+    def prepare_models(self, denoiser: torch.nn.Module) -> None:
+        """
+        Prepares the models for the guidance technique on a given batch of data. This method should be overridden in
+        subclasses to implement specific model preparation logic.
+        """
+        self._count_prepared += 1
+    
+    def cleanup_models(self, denoiser: torch.nn.Module) -> None:
+        """
+        Cleans up the models for the guidance technique after a given batch of data. This method should be overridden in
+        subclasses to implement specific model cleanup logic. It is useful for removing any hooks or other stateful
+        modifications made during `prepare_models`.
+        """
+        pass
+    
+    def prepare_inputs(self, data: "BlockState") -> List["BlockState"]:
+        raise NotImplementedError("BaseGuidance::prepare_inputs must be implemented in subclasses.")
+
+    def __call__(self, data: List["BlockState"]) -> Any:
+        if not all(hasattr(d, "noise_pred") for d in data):
+            raise ValueError("Expected all data to have `noise_pred` attribute.")
+        if len(data) != self.num_conditions:
+            raise ValueError(
+                f"Expected {self.num_conditions} data items, but got {len(data)}. Please check the input data."
+            )
+        forward_inputs = {getattr(d, self._identifier_key): d.noise_pred for d in data}
+        return self.forward(**forward_inputs)
+
+    def forward(self, *args, **kwargs) -> Any:
+        raise NotImplementedError("BaseGuidance::forward must be implemented in subclasses.")
+
+    @property
+    def is_conditional(self) -> bool:
+        raise NotImplementedError("BaseGuidance::is_conditional must be implemented in subclasses.")
+    
+    @property
+    def is_unconditional(self) -> bool:
+        return not self.is_conditional
+    
+    @property
+    def num_conditions(self) -> int:
+        raise NotImplementedError("BaseGuidance::num_conditions must be implemented in subclasses.")
+    
+    @classmethod
+    def _prepare_batch(cls, input_fields: Dict[str, Union[str, Tuple[str, str]]], data: "BlockState", tuple_index: int, identifier: str) -> "BlockState":
+        """
+        Prepares a batch of data for the guidance technique. This method is used in the `prepare_inputs` method of
+        the `BaseGuidance` class. It prepares the batch based on the provided tuple index.
+
+        Args:
+            input_fields (`Dict[str, Union[str, Tuple[str, str]]]`):
+                A dictionary where the keys are the names of the fields that will be used to store the data once
+                it is prepared with `prepare_inputs`. The values can be either a string or a tuple of length 2,
+                which is used to look up the required data provided for preparation.
+                If a string is provided, it will be used as the conditional data (or unconditional if used with
+                a guidance method that requires it). If a tuple of length 2 is provided, the first element must
+                be the conditional data identifier and the second element must be the unconditional data identifier
+                or None.
+            data (`BlockState`):
+                The input data to be prepared.
+            tuple_index (`int`):
+                The index to use when accessing input fields that are tuples.
+        
+        Returns:
+            `BlockState`: The prepared batch of data.
+        """
+        from ..modular_pipelines.modular_pipeline import BlockState
+
+        if input_fields is None:
+            raise ValueError("Input fields cannot be None. Please pass `input_fields` to `prepare_inputs` or call `set_input_fields` before preparing inputs.")
+        data_batch = {}
+        for key, value in input_fields.items():
+            try:
+                if isinstance(value, str):
+                    data_batch[key] = getattr(data, value)
+                elif isinstance(value, tuple):
+                    data_batch[key] = getattr(data, value[tuple_index])
+                else:
+                    # We've already checked that value is a string or a tuple of strings with length 2
+                    pass
+            except AttributeError:
+                logger.debug(f"`data` does not have attribute(s) {value}, skipping.")
+        data_batch[cls._identifier_key] = identifier
+        return BlockState(**data_batch)
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    r"""
+    Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on
+    Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
+    Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+    Args:
+        noise_cfg (`torch.Tensor`):
+            The predicted noise tensor for the guided diffusion process.
+        noise_pred_text (`torch.Tensor`):
+            The predicted noise tensor for the text-guided diffusion process.
+        guidance_rescale (`float`, *optional*, defaults to 0.0):
+            A rescale factor applied to the noise predictions.
+    Returns:
+        noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor.
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg