test forcing tokenizers

* Add Differential Pipeline.

* Fix Styling Issue using ruff -fix

* Add details to Contributing.md

* Revert "Fix Styling Issue using ruff -fix"

This reverts commit d347de162d.

* Revert "Revert "Fix Styling Issue using ruff -fix""

This reverts commit ce7c3ff216.

* Revert README changes

* Restore README.md

* Update README.md

* Resolved Comments:

* Fix Readme based on review

* Fix formatting after make style

---------

Co-authored-by: Aryan <aryan@huggingface.co>

docs/source/en/_toctree.yml

@@ -190,6 +190,10 @@
   - local: conceptual/evaluation
     title: Evaluating Diffusion Models
   title: Conceptual Guides
+- sections:
+  - local: community_projects
+    title: Projects built with Diffusers
+  title: Community Projects
 - sections:
   - isExpanded: false
     sections:
@@ -219,62 +223,76 @@
     sections:
     - local: api/models/overview
       title: Overview
-    - local: api/models/unet
-      title: UNet1DModel
-    - local: api/models/unet2d
-      title: UNet2DModel
-    - local: api/models/unet2d-cond
-      title: UNet2DConditionModel
-    - local: api/models/unet3d-cond
-      title: UNet3DConditionModel
-    - local: api/models/unet-motion
-      title: UNetMotionModel
-    - local: api/models/uvit2d
-      title: UViT2DModel
-    - local: api/models/vq
-      title: VQModel
-    - local: api/models/autoencoderkl
-      title: AutoencoderKL
-    - local: api/models/asymmetricautoencoderkl
-      title: AsymmetricAutoencoderKL
-    - local: api/models/autoencoder_tiny
-      title: Tiny AutoEncoder
-    - local: api/models/autoencoder_oobleck
-      title: Oobleck AutoEncoder
-    - local: api/models/consistency_decoder_vae
-      title: ConsistencyDecoderVAE
-    - local: api/models/transformer2d
-      title: Transformer2DModel
-    - local: api/models/pixart_transformer2d
-      title: PixArtTransformer2DModel
-    - local: api/models/dit_transformer2d
-      title: DiTTransformer2DModel
-    - local: api/models/hunyuan_transformer2d
-      title: HunyuanDiT2DModel
-    - local: api/models/aura_flow_transformer2d
-      title: AuraFlowTransformer2DModel
-    - local: api/models/flux_transformer
-      title: FluxTransformer2DModel
-    - local: api/models/latte_transformer3d
-      title: LatteTransformer3DModel
-    - local: api/models/lumina_nextdit2d
-      title: LuminaNextDiT2DModel
-    - local: api/models/transformer_temporal
-      title: TransformerTemporalModel
-    - local: api/models/sd3_transformer2d
-      title: SD3Transformer2DModel
-    - local: api/models/stable_audio_transformer
-      title: StableAudioDiTModel
-    - local: api/models/prior_transformer
-      title: PriorTransformer
-    - local: api/models/controlnet
-      title: ControlNetModel
-    - local: api/models/controlnet_hunyuandit
-      title: HunyuanDiT2DControlNetModel
-    - local: api/models/controlnet_sd3
-      title: SD3ControlNetModel
-    - local: api/models/controlnet_sparsectrl
-      title: SparseControlNetModel
+    - sections:
+      - local: api/models/controlnet
+        title: ControlNetModel
+      - local: api/models/controlnet_hunyuandit
+        title: HunyuanDiT2DControlNetModel
+      - local: api/models/controlnet_sd3
+        title: SD3ControlNetModel
+      - local: api/models/controlnet_sparsectrl
+        title: SparseControlNetModel
+      title: ControlNets
+    - sections:
+      - local: api/models/aura_flow_transformer2d
+        title: AuraFlowTransformer2DModel
+      - local: api/models/cogvideox_transformer3d
+        title: CogVideoXTransformer3DModel
+      - local: api/models/dit_transformer2d
+        title: DiTTransformer2DModel
+      - local: api/models/flux_transformer
+        title: FluxTransformer2DModel
+      - local: api/models/hunyuan_transformer2d
+        title: HunyuanDiT2DModel
+      - local: api/models/latte_transformer3d
+        title: LatteTransformer3DModel
+      - local: api/models/lumina_nextdit2d
+        title: LuminaNextDiT2DModel
+      - local: api/models/pixart_transformer2d
+        title: PixArtTransformer2DModel
+      - local: api/models/prior_transformer
+        title: PriorTransformer
+      - local: api/models/sd3_transformer2d
+        title: SD3Transformer2DModel
+      - local: api/models/stable_audio_transformer
+        title: StableAudioDiTModel
+      - local: api/models/transformer2d
+        title: Transformer2DModel
+      - local: api/models/transformer_temporal
+        title: TransformerTemporalModel
+      title: Transformers
+    - sections:
+      - local: api/models/stable_cascade_unet
+        title: StableCascadeUNet
+      - local: api/models/unet
+        title: UNet1DModel
+      - local: api/models/unet2d
+        title: UNet2DModel
+      - local: api/models/unet2d-cond
+        title: UNet2DConditionModel
+      - local: api/models/unet3d-cond
+        title: UNet3DConditionModel
+      - local: api/models/unet-motion
+        title: UNetMotionModel
+      - local: api/models/uvit2d
+        title: UViT2DModel
+      title: UNets
+    - sections:
+      - local: api/models/autoencoderkl
+        title: AutoencoderKL
+      - local: api/models/autoencoderkl_cogvideox
+        title: AutoencoderKLCogVideoX
+      - local: api/models/asymmetricautoencoderkl
+        title: AsymmetricAutoencoderKL
+      - local: api/models/consistency_decoder_vae
+        title: ConsistencyDecoderVAE
+      - local: api/models/autoencoder_oobleck
+        title: Oobleck AutoEncoder
+      - local: api/models/autoencoder_tiny
+        title: Tiny AutoEncoder
+      - local: api/models/vq
+        title: VQModel
+      title: VAEs
     title: Models
   - isExpanded: false
     sections:
@@ -296,6 +314,8 @@
       title: AutoPipeline
     - local: api/pipelines/blip_diffusion
       title: BLIP-Diffusion
+    - local: api/pipelines/cogvideox
+      title: CogVideoX
     - local: api/pipelines/consistency_models
       title: Consistency Models
     - local: api/pipelines/controlnet

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

@@ -53,6 +53,7 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:
 - [`AutoencoderKLCogVideoX`]
 - [`ControlNetModel`]
 - [`SD3Transformer2DModel`]
+- [`FluxTransformer2DModel`]
 
 ## FromSingleFileMixin
 

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

@@ -11,18 +11,14 @@ specific language governing permissions and limitations under the License. -->
 
 # AutoencoderKLCogVideoX
 
-The 3D variational autoencoder (VAE) model with KL loss using CogVideoX.
+The 3D variational autoencoder (VAE) model with KL loss used in [CogVideoX](https://github.com/THUDM/CogVideo) was introduced in [CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://github.com/THUDM/CogVideo/blob/main/resources/CogVideoX.pdf) by Tsinghua University & ZhipuAI.
 
-## Loading from the original format
+The model can be loaded with the following code snippet.
 
-By default, the [`AutoencoderKLCogVideoX`] should be loaded with [`~ModelMixin.from_pretrained`], but it can also be loaded from the original format using [`FromOriginalModelMixin.from_single_file`] as follows:
-
-```py
+```python
 from diffusers import AutoencoderKLCogVideoX
 
-url = "THUDM/CogVideoX-2b"  # can also be a local file
-model = AutoencoderKLCogVideoX.from_single_file(url)
-
+vae = AutoencoderKLCogVideoX.from_pretrained("THUDM/CogVideoX-2b", subfolder="vae", torch_dtype=torch.float16).to("cuda")
 ```
 
 ## AutoencoderKLCogVideoX
@@ -32,38 +28,10 @@ model = AutoencoderKLCogVideoX.from_single_file(url)
     - encode
     - all
 
-## CogVideoXSafeConv3d
+## AutoencoderKLOutput
 
-[[autodoc]] CogVideoXSafeConv3d
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
 
-## CogVideoXCausalConv3d
+## DecoderOutput
 
-[[autodoc]] CogVideoXCausalConv3d
-
-## CogVideoXSpatialNorm3D
-
-[[autodoc]] CogVideoXSpatialNorm3D
-
-## CogVideoXResnetBlock3D
-
-[[autodoc]] CogVideoXResnetBlock3D
-
-## CogVideoXDownBlock3D
-
-[[autodoc]] CogVideoXDownBlock3D
-
-## CogVideoXMidBlock3D
-
-[[autodoc]] CogVideoXMidBlock3D
-
-## CogVideoXUpBlock3D
-
-[[autodoc]] CogVideoXUpBlock3D
-
-## CogVideoXEncoder3D
-
-[[autodoc]] CogVideoXEncoder3D
-
-## CogVideoXDecoder3D
-
-[[autodoc]] CogVideoXDecoder3D
+[[autodoc]] models.autoencoders.vae.DecoderOutput

docs/source/en/api/models/cogvideox_transformer3d.md

@@ -9,10 +9,22 @@ Unless required by applicable law or agreed to in writing, software distributed
 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. -->
 
-## CogVideoXTransformer3DModel
+# CogVideoXTransformer3DModel
 
-A Diffusion Transformer model for 3D data from [CogVideoX](https://github.com/THUDM/CogVideoX).
+A Diffusion Transformer model for 3D data from [CogVideoX](https://github.com/THUDM/CogVideo) was introduced in [CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://github.com/THUDM/CogVideo/blob/main/resources/CogVideoX.pdf) by Tsinghua University & ZhipuAI.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import CogVideoXTransformer3DModel
+
+vae = CogVideoXTransformer3DModel.from_pretrained("THUDM/CogVideoX-2b", subfolder="transformer", torch_dtype=torch.float16).to("cuda")
+```
 
 ## CogVideoXTransformer3DModel
 
 [[autodoc]] CogVideoXTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

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

@@ -0,0 +1,19 @@
+<!--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.
+-->
+
+# StableCascadeUNet
+
+A UNet model from the [Stable Cascade pipeline](../pipelines/stable_cascade.md).
+
+## StableCascadeUNet
+
+[[autodoc]] models.unets.unet_stable_cascade.StableCascadeUNet

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

@@ -10,18 +10,18 @@
 # 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. 
-
-## TODO: The paper is still being written.
+# limitations under the License.
 -->
 
 # CogVideoX
 
-[TODO]() from Tsinghua University & ZhipuAI.
+<!-- TODO: update paper with ArXiv link when ready. -->
+
+[CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://github.com/THUDM/CogVideo/blob/main/resources/CogVideoX.pdf) from Tsinghua University & ZhipuAI.
 
 The abstract from the paper is:
 
-The paper is still being written.
+*We introduce CogVideoX, a large-scale diffusion transformer model designed for generating videos based on text prompts. To efficently model video data, we propose to levearge a 3D Variational Autoencoder (VAE) to compresses videos along both spatial and temporal dimensions. To improve the text-video alignment, we propose an expert transformer with the expert adaptive LayerNorm to facilitate the deep fusion between the two modalities. By employing a progressive training technique, CogVideoX is adept at producing coherent, long-duration videos characterized by significant motion. In addition, we develop an effectively text-video data processing pipeline that includes various data preprocessing strategies and a video captioning method. It significantly helps enhance the performance of CogVideoX, improving both generation quality and semantic alignment. Results show that CogVideoX demonstrates state-of-the-art performance across both multiple machine metrics and human evaluations. The model weight of CogVideoX-2B is publicly available at https://github.com/THUDM/CogVideo.*
 
 <Tip>
 
@@ -29,7 +29,9 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.m
 
 </Tip>
 
-### Inference
+This pipeline was contributed by [zRzRzRzRzRzRzR](https://github.com/zRzRzRzRzRzRzR). The original codebase can be found [here](https://huggingface.co/THUDM). The original weights can be found under [hf.co/THUDM](https://huggingface.co/THUDM).
+
+## Inference
 
 Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
 
@@ -37,11 +39,20 @@ First, load the pipeline:
 
 ```python
 import torch
-from diffusers import LattePipeline
+from diffusers import CogVideoXPipeline
+from diffusers.utils import export_to_video
 
-pipeline = LattePipeline.from_pretrained(
-	"THUDM/CogVideoX-2b", torch_dtype=torch.float16
-).to("cuda")
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b").to("cuda")
+prompt = (
+    "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
+    "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
+    "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, "
+    "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. "
+    "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
+    "atmosphere of this unique musical performance."
+)
+video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+export_to_video(video, "output.mp4", fps=8)
 ```
 
 Then change the memory layout of the pipelines `transformer` and `vae` components to `torch.channels-last`:
@@ -76,4 +87,5 @@ With torch.compile(): Average inference time: TODO seconds.
   - __call__
 
 ## CogVideoXPipelineOutput
-[[autodoc]] pipelines.pipline_cogvideo.pipeline_output.CogVideoXPipelineOutput
+
+[[autodoc]] pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput

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

@@ -37,7 +37,7 @@ Both checkpoints have slightly difference usage which we detail below.
 
 ```python
 import torch
-from diffusers import  FluxPipeline
+from diffusers import FluxPipeline
 
 pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16)
 pipe.enable_model_cpu_offload()
@@ -61,7 +61,7 @@ out.save("image.png")
 
 ```python
 import torch
-from diffusers import  FluxPipeline
+from diffusers import FluxPipeline
 
 pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
 pipe.enable_model_cpu_offload()
@@ -77,8 +77,89 @@ out = pipe(
 out.save("image.png")
 ```
 
+## Running FP16 inference
+Flux can generate high-quality images with FP16 (i.e. to accelerate inference on Turing/Volta GPUs) but produces different outputs compared to FP32/BF16. The issue is that some activations in the text encoders have to be clipped when running in FP16, which affects the overall image. Forcing text encoders to run with FP32 inference thus removes this output difference. See [here](https://github.com/huggingface/diffusers/pull/9097#issuecomment-2272292516) for details.
+
+FP16 inference code:
+```python
+import torch
+from diffusers import FluxPipeline
+
+pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16) # can replace schnell with dev
+# to run on low vram GPUs (i.e. between 4 and 32 GB VRAM)
+pipe.enable_sequential_cpu_offload()
+pipe.vae.enable_slicing()
+pipe.vae.enable_tiling()
+
+pipe.to(torch.float16) # casting here instead of in the pipeline constructor because doing so in the constructor loads all models into CPU memory at once
+
+prompt = "A cat holding a sign that says hello world"
+out = pipe(
+    prompt=prompt,
+    guidance_scale=0.,
+    height=768,
+    width=1360,
+    num_inference_steps=4,
+    max_sequence_length=256,
+).images[0]
+out.save("image.png")
+```
+
+## Single File Loading for the `FluxTransformer2DModel`
+
+The `FluxTransformer2DModel` supports loading checkpoints in the original format shipped by Black Forest Labs. This is also useful when trying to load finetunes or quantized versions of the models that have been published by the community.
+
+<Tip>
+`FP8` inference can be brittle depending on the GPU type, CUDA version, and `torch` version that you are using. It is recommended that you use the `optimum-quanto` library in order to run FP8 inference on your machine.
+</Tip>
+
+The following example demonstrates how to run Flux with less than 16GB of VRAM.
+
+First install `optimum-quanto`
+
+```shell
+pip install optimum-quanto
+```
+
+Then run the following example
+
+```python
+import torch
+from diffusers import FluxTransformer2DModel, FluxPipeline
+from transformers import T5EncoderModel, CLIPTextModel
+from optimum.quanto import freeze, qfloat8, quantize
+
+bfl_repo = "black-forest-labs/FLUX.1-dev"
+dtype = torch.bfloat16
+
+transformer = FluxTransformer2DModel.from_single_file("https://huggingface.co/Kijai/flux-fp8/blob/main/flux1-dev-fp8.safetensors", torch_dtype=dtype)
+quantize(transformer, weights=qfloat8)
+freeze(transformer)
+
+text_encoder_2 = T5EncoderModel.from_pretrained(bfl_repo, subfolder="text_encoder_2", torch_dtype=dtype)
+quantize(text_encoder_2, weights=qfloat8)
+freeze(text_encoder_2)
+
+pipe = FluxPipeline.from_pretrained(bfl_repo, transformer=None, text_encoder_2=None, torch_dtype=dtype)
+pipe.transformer = transformer
+pipe.text_encoder_2 = text_encoder_2
+
+pipe.enable_model_cpu_offload()
+
+prompt = "A cat holding a sign that says hello world"
+image = pipe(
+    prompt,
+    guidance_scale=3.5,
+    output_type="pil",
+    num_inference_steps=20,
+    generator=torch.Generator("cpu").manual_seed(0)
+).images[0]
+
+image.save("flux-fp8-dev.png")
+```
+
 ## FluxPipeline
 
 [[autodoc]] FluxPipeline
 	- all
-	- __call__
+	- __call__

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

@@ -43,6 +43,11 @@ Since RegEx is supported as a way for matching layer identifiers, it is crucial
   - all
   - __call__
 
+## KolorsPAGPipeline
+[[autodoc]] KolorsPAGPipeline
+  - all
+  - __call__  
+
 ## StableDiffusionPAGPipeline
 [[autodoc]] StableDiffusionPAGPipeline
 	- all
@@ -74,6 +79,12 @@ Since RegEx is supported as a way for matching layer identifiers, it is crucial
 	- __call__
 
 
+## StableDiffusion3PAGPipeline
+[[autodoc]] StableDiffusion3PAGPipeline
+	- all
+	- __call__
+
+
 ## PixArtSigmaPAGPipeline
 [[autodoc]] PixArtSigmaPAGPipeline
 	- all

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

@@ -21,7 +21,7 @@ Stable Audio is trained on a corpus of around 48k audio recordings, where around
 The abstract of the paper is the following:
 *Open generative models are vitally important for the community, allowing for fine-tunes and serving as baselines when presenting new models. However, most current text-to-audio models are private and not accessible for artists and researchers to build upon. Here we describe the architecture and training process of a new open-weights text-to-audio model trained with Creative Commons data. Our evaluation shows that the model's performance is competitive with the state-of-the-art across various metrics. Notably, the reported FDopenl3 results (measuring the realism of the generations) showcase its potential for high-quality stereo sound synthesis at 44.1kHz.*
 
-This pipeline was contributed by [Yoach Lacombe](https://huggingface.co/ylacombe). The original codebase can be found at [Stability-AI/stable-audio-tool](https://github.com/Stability-AI/stable-audio-tool).
+This pipeline was contributed by [Yoach Lacombe](https://huggingface.co/ylacombe). The original codebase can be found at [Stability-AI/stable-audio-tools](https://github.com/Stability-AI/stable-audio-tools).
 
 ## Tips
 

docs/source/en/community_projects.md

@@ -0,0 +1,78 @@
+<!--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.
+-->
+
+# Community Projects
+
+Welcome to Community Projects. This space is dedicated to showcasing the incredible work and innovative applications created by our vibrant community using the `diffusers` library.
+
+This section aims to:
+
+- Highlight diverse and inspiring projects built with `diffusers`
+- Foster knowledge sharing within our community
+- Provide real-world examples of how `diffusers` can be leveraged
+
+Happy exploring, and thank you for being part of the Diffusers community!
+
+<table>
+    <tr>
+        <th>Project Name</th>
+        <th>Description</th>
+    </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/carson-katri/dream-textures"> dream-textures </a></td>
+    <td>Stable Diffusion built-in to Blender</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/megvii-research/HiDiffusion"> HiDiffusion </a></td>
+    <td>Increases the resolution and speed of your diffusion model by only adding a single line of code</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/lllyasviel/IC-Light"> IC-Light </a></td>
+    <td>IC-Light is a project to manipulate the illumination of images</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/InstantID/InstantID"> InstantID </a></td>
+    <td>InstantID : Zero-shot Identity-Preserving Generation in Seconds</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/Sanster/IOPaint"> IOPaint </a></td>
+    <td>Image inpainting tool powered by SOTA AI Model. Remove any unwanted object, defect, people from your pictures or erase and replace(powered by stable diffusion) any thing on your pictures.</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/bmaltais/kohya_ss"> Kohya </a></td>
+    <td>Gradio GUI for Kohya's Stable Diffusion trainers</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/magic-research/magic-animate"> MagicAnimate </a></td>
+    <td>MagicAnimate: Temporally Consistent Human Image Animation using Diffusion Model</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/levihsu/OOTDiffusion"> OOTDiffusion </a></td>
+    <td>Outfitting Fusion based Latent Diffusion for Controllable Virtual Try-on</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/vladmandic/automatic"> SD.Next </a></td>
+    <td>SD.Next: Advanced Implementation of Stable Diffusion and other Diffusion-based generative image models</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/ashawkey/stable-dreamfusion"> stable-dreamfusion </a></td>
+    <td>Text-to-3D & Image-to-3D & Mesh Exportation with NeRF + Diffusion</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/HVision-NKU/StoryDiffusion"> StoryDiffusion </a></td>
+    <td>StoryDiffusion can create a magic story by generating consistent images and videos.</td>
+  </tr>
+  <tr style="border-top: 2px solid black">
+    <td><a href="https://github.com/cumulo-autumn/StreamDiffusion"> StreamDiffusion </a></td>
+    <td>A Pipeline-Level Solution for Real-Time Interactive Generation</td>
+  </tr>
+</table>

docs/source/en/optimization/fp16.md

@@ -125,3 +125,5 @@ image
     <figcaption class="mt-2 text-center text-sm text-gray-500">distilled Stable Diffusion + Tiny AutoEncoder</figcaption>
   </div>
 </div>
+
+More tiny autoencoder models for other Stable Diffusion models, like Stable Diffusion 3, are available from [madebyollin](https://huggingface.co/madebyollin).

docs/source/en/training/instructpix2pix.md

@@ -14,7 +14,7 @@ specific language governing permissions and limitations under the License.
 
 [InstructPix2Pix](https://hf.co/papers/2211.09800) is a Stable Diffusion model trained to edit images from human-provided instructions. For example, your prompt can be "turn the clouds rainy" and the model will edit the input image accordingly. This model is conditioned on the text prompt (or editing instruction) and the input image.
 
-This guide will explore the [train_instruct_pix2pix.py](https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix.py) training script to help you become familiar with it, and how you can adapt it for your own use-case.
+This guide will explore the [train_instruct_pix2pix.py](https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix.py) training script to help you become familiar with it, and how you can adapt it for your own use case.
 
 Before running the script, make sure you install the library from source:
 
@@ -117,7 +117,7 @@ optimizer = optimizer_cls(
 )
 ```
 
-Next, the edited images and and edit instructions are [preprocessed](https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L624) and [tokenized](https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L610C24-L610C24). It is important the same image transformations are applied to the original and edited images.
+Next, the edited images and edit instructions are [preprocessed](https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L624) and [tokenized](https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L610C24-L610C24). It is important the same image transformations are applied to the original and edited images.
 
 ```py
 def preprocess_train(examples):
@@ -249,4 +249,4 @@ The SDXL training script is discussed in more detail in the [SDXL training](sdxl
 
 Congratulations on training your own InstructPix2Pix model! 🥳 To learn more about the model, it may be helpful to:
 
-- Read the [Instruction-tuning Stable Diffusion with InstructPix2Pix](https://huggingface.co/blog/instruction-tuning-sd) blog post to learn more about some experiments we've done with InstructPix2Pix, dataset preparation, and results for different instructions.
+- Read the [Instruction-tuning Stable Diffusion with InstructPix2Pix](https://huggingface.co/blog/instruction-tuning-sd) blog post to learn more about some experiments we've done with InstructPix2Pix, dataset preparation, and results for different instructions.

docs/source/en/tutorials/using_peft_for_inference.md

@@ -34,7 +34,7 @@ pipe_id = "stabilityai/stable-diffusion-xl-base-1.0"
 pipe = DiffusionPipeline.from_pretrained(pipe_id, torch_dtype=torch.float16).to("cuda")
 ```
 
-Next, load a [CiroN2022/toy-face](https://huggingface.co/CiroN2022/toy-face) adapter with the [`~diffusers.loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] method. With the 🤗 PEFT integration, you can assign a specific `adapter_name` to the checkpoint, which let's you easily switch between different LoRA checkpoints. Let's call this adapter `"toy"`.
+Next, load a [CiroN2022/toy-face](https://huggingface.co/CiroN2022/toy-face) adapter with the [`~diffusers.loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] method. With the 🤗 PEFT integration, you can assign a specific `adapter_name` to the checkpoint, which lets you easily switch between different LoRA checkpoints. Let's call this adapter `"toy"`.
 
 ```python
 pipe.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")

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

@@ -12,7 +12,7 @@ specific language governing permissions and limitations under the License.
 
 # Pipeline callbacks
 
-The denoising loop of a pipeline can be modified with custom defined functions using the `callback_on_step_end` parameter. The callback function is executed at the end of each step, and modifies the pipeline attributes and variables for the next step. This is really useful for *dynamically* adjusting certain pipeline attributes or modifying tensor variables. This versatility allows for interesting use-cases such as changing the prompt embeddings at each timestep, assigning different weights to the prompt embeddings, and editing the guidance scale. With callbacks, you can implement new features without modifying the underlying code!
+The denoising loop of a pipeline can be modified with custom defined functions using the `callback_on_step_end` parameter. The callback function is executed at the end of each step, and modifies the pipeline attributes and variables for the next step. This is really useful for *dynamically* adjusting certain pipeline attributes or modifying tensor variables. This versatility allows for interesting use cases such as changing the prompt embeddings at each timestep, assigning different weights to the prompt embeddings, and editing the guidance scale. With callbacks, you can implement new features without modifying the underlying code!
 
 > [!TIP]
 > 🤗 Diffusers currently only supports `callback_on_step_end`, but feel free to open a [feature request](https://github.com/huggingface/diffusers/issues/new/choose) if you have a cool use-case and require a callback function with a different execution point!
@@ -75,7 +75,7 @@ out.images[0].save("official_callback.png")
     <figcaption class="mt-2 text-center text-sm text-gray-500">without SDXLCFGCutoffCallback</figcaption>
   </div>
   <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/with_cfg_callback.png" alt="generated image of a a sports car at the road with cfg callback" />
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/with_cfg_callback.png" alt="generated image of a sports car at the road with cfg callback" />
     <figcaption class="mt-2 text-center text-sm text-gray-500">with SDXLCFGCutoffCallback</figcaption>
   </div>
 </div>

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

@@ -289,9 +289,9 @@ scheduler = DPMSolverMultistepScheduler.from_pretrained(pipe_id, subfolder="sche
 3. Load an image processor:
 
 ```python
-from transformers import CLIPFeatureExtractor
+from transformers import CLIPImageProcessor
 
-feature_extractor = CLIPFeatureExtractor.from_pretrained(pipe_id, subfolder="feature_extractor")
+feature_extractor = CLIPImageProcessor.from_pretrained(pipe_id, subfolder="feature_extractor")
 ```
 
 <Tip warning={true}>

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

@@ -212,14 +212,14 @@ TCD-LoRA is very versatile, and it can be combined with other adapter types like
 import torch
 import numpy as np
 from PIL import Image
-from transformers import DPTFeatureExtractor, DPTForDepthEstimation
+from transformers import DPTImageProcessor, DPTForDepthEstimation
 from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline
 from diffusers.utils import load_image, make_image_grid
 from scheduling_tcd import TCDScheduler
 
 device = "cuda"
 depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to(device)
-feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-hybrid-midas")
+feature_extractor = DPTImageProcessor.from_pretrained("Intel/dpt-hybrid-midas")
 
 def get_depth_map(image):
     image = feature_extractor(images=image, return_tensors="pt").pixel_values.to(device)

docs/source/ko/using-diffusers/loading.md

@@ -307,7 +307,7 @@ print(pipeline)
 
 위의 코드 출력 결과를 확인해보면, `pipeline`은 [`StableDiffusionPipeline`]의 인스턴스이며, 다음과 같이 총 7개의 컴포넌트로 구성된다는 것을 알 수 있습니다.
 
-- `"feature_extractor"`: [`~transformers.CLIPFeatureExtractor`]의 인스턴스
+- `"feature_extractor"`: [`~transformers.CLIPImageProcessor`]의 인스턴스
 - `"safety_checker"`: 유해한 컨텐츠를 스크리닝하기 위한 [컴포넌트](https://github.com/huggingface/diffusers/blob/e55687e1e15407f60f32242027b7bb8170e58266/src/diffusers/pipelines/stable_diffusion/safety_checker.py#L32)
 - `"scheduler"`: [`PNDMScheduler`]의 인스턴스
 - `"text_encoder"`: [`~transformers.CLIPTextModel`]의 인스턴스

docs/source/ko/using-diffusers/textual_inversion_inference.md

@@ -24,7 +24,7 @@ import PIL
 from PIL import Image
 
 from diffusers import StableDiffusionPipeline
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 
 def image_grid(imgs, rows, cols):

examples/community/README.md

@@ -71,6 +71,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | Stable Diffusion BoxDiff Pipeline | Training-free controlled generation with bounding boxes using [BoxDiff](https://github.com/showlab/BoxDiff) | [Stable Diffusion BoxDiff Pipeline](#stable-diffusion-boxdiff) | - | [Jingyang Zhang](https://github.com/zjysteven/) |
 |   FRESCO V2V Pipeline                                                                                                    | Implementation of [[CVPR 2024] FRESCO: Spatial-Temporal Correspondence for Zero-Shot Video Translation](https://arxiv.org/abs/2403.12962)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [FRESCO V2V Pipeline](#fresco)      | - |              [Yifan Zhou](https://github.com/SingleZombie) |
 | AnimateDiff IPEX Pipeline | Accelerate AnimateDiff inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [AnimateDiff on IPEX](#animatediff-on-ipex) | - | [Dan Li](https://github.com/ustcuna/) |
+| HunyuanDiT Differential Diffusion Pipeline | Applies [Differential Diffsuion](https://github.com/exx8/differential-diffusion) to [HunyuanDiT](https://github.com/huggingface/diffusers/pull/8240). | [HunyuanDiT with Differential Diffusion](#hunyuandit-with-differential-diffusion) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1v44a5fpzyr4Ffr4v2XBQ7BajzG874N4P?usp=sharing) | [Monjoy Choudhury](https://github.com/MnCSSJ4x) |
 
 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.
 
@@ -1435,9 +1436,9 @@ import requests
 import torch
 from diffusers import DiffusionPipeline
 from PIL import Image
-from transformers import CLIPFeatureExtractor, CLIPModel
+from transformers import CLIPImageProcessor, CLIPModel
 
-feature_extractor = CLIPFeatureExtractor.from_pretrained(
+feature_extractor = CLIPImageProcessor.from_pretrained(
     "laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
 )
 clip_model = CLIPModel.from_pretrained(
@@ -1646,7 +1647,6 @@ from diffusers import DiffusionPipeline
 scheduler = DDIMScheduler.from_pretrained("stabilityai/stable-diffusion-2-1",
                                             subfolder="scheduler")
 
-
 pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1",
                                             custom_pipeline="stable_diffusion_tensorrt_img2img",
                                             variant='fp16',
@@ -1661,7 +1661,6 @@ pipe = pipe.to("cuda")
 url = "https://pajoca.com/wp-content/uploads/2022/09/tekito-yamakawa-1.png"
 response = requests.get(url)
 input_image = Image.open(BytesIO(response.content)).convert("RGB")
-
 prompt = "photorealistic new zealand hills"
 image = pipe(prompt, image=input_image, strength=0.75,).images[0]
 image.save('tensorrt_img2img_new_zealand_hills.png')
@@ -2122,7 +2121,7 @@ import torch
 import open_clip
 from open_clip import SimpleTokenizer
 from diffusers import DiffusionPipeline
-from transformers import CLIPFeatureExtractor, CLIPModel
+from transformers import CLIPImageProcessor, CLIPModel
 
 
 def download_image(url):
@@ -2130,7 +2129,7 @@ def download_image(url):
     return PIL.Image.open(BytesIO(response.content)).convert("RGB")
 
 # Loading additional models
-feature_extractor = CLIPFeatureExtractor.from_pretrained(
+feature_extractor = CLIPImageProcessor.from_pretrained(
     "laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
 )
 clip_model = CLIPModel.from_pretrained(
@@ -4209,6 +4208,52 @@ print("Latency of AnimateDiffPipelineIpex--fp32", latency, "s for total", step,
 latency = elapsed_time(pipe4, num_inference_steps=step)
 print("Latency of AnimateDiffPipeline--fp32",latency, "s for total", step, "steps")
 ```
+### HunyuanDiT with Differential Diffusion
+
+#### Usage
+
+```python
+import torch
+from diffusers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import load_image
+from PIL import Image
+from torchvision import transforms
+
+from pipeline_hunyuandit_differential_img2img import (
+    HunyuanDiTDifferentialImg2ImgPipeline,
+)
+
+
+pipe = HunyuanDiTDifferentialImg2ImgPipeline.from_pretrained(
+    "Tencent-Hunyuan/HunyuanDiT-Diffusers", torch_dtype=torch.float16
+).to("cuda")
+
+
+source_image = load_image(
+    "https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/differential/20240329211129_4024911930.png"
+)
+map = load_image(
+    "https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/differential/gradient_mask_2.png"
+)
+prompt = "a green pear"
+negative_prompt = "blurry"
+
+image = pipe(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    image=source_image,
+    num_inference_steps=28,
+    guidance_scale=4.5,
+    strength=1.0,
+    map=map,
+).images[0]
+```
+
+| ![Gradient](https://github.com/user-attachments/assets/e38ce4d5-1ae6-4df0-ab43-adc1b45716b5) | ![Input](https://github.com/user-attachments/assets/9c95679c-e9d7-4f5a-90d6-560203acd6b3) | ![Output](https://github.com/user-attachments/assets/5313ff64-a0c4-418b-8b55-a38f1a5e7532) |
+| ------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------- |
+| Gradient                                                                                   | Input                                                                                   | Output                                                                                   |
+
+A colab notebook demonstrating all results can be found [here](https://colab.research.google.com/drive/1v44a5fpzyr4Ffr4v2XBQ7BajzG874N4P?usp=sharing). Depth Maps have also been added in the same colab.
 
 # Perturbed-Attention Guidance
 
@@ -4285,4 +4330,4 @@ grid_image.save(grid_dir + "sample.png")
 
 `pag_scale` : guidance scale of PAG (ex: 5.0)
 
-`pag_applied_layers_index` : index of the layer to apply perturbation (ex: ['m0'])
+`pag_applied_layers_index` : index of the layer to apply perturbation (ex: ['m0'])

examples/community/clip_guided_images_mixing_stable_diffusion.py

@@ -7,7 +7,7 @@ import PIL.Image
 import torch
 from torch.nn import functional as F
 from torchvision import transforms
-from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPModel, CLIPTextModel, CLIPTokenizer
 
 from diffusers import (
     AutoencoderKL,
@@ -86,7 +86,7 @@ class CLIPGuidedImagesMixingStableDiffusion(DiffusionPipeline, StableDiffusionMi
         tokenizer: CLIPTokenizer,
         unet: UNet2DConditionModel,
         scheduler: Union[PNDMScheduler, LMSDiscreteScheduler, DDIMScheduler, DPMSolverMultistepScheduler],
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         coca_model=None,
         coca_tokenizer=None,
         coca_transform=None,

examples/community/clip_guided_stable_diffusion_img2img.py

@@ -7,7 +7,7 @@ import torch
 from torch import nn
 from torch.nn import functional as F
 from torchvision import transforms
-from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPModel, CLIPTextModel, CLIPTokenizer
 
 from diffusers import (
     AutoencoderKL,
@@ -32,9 +32,9 @@ EXAMPLE_DOC_STRING = """
         import torch
         from diffusers import DiffusionPipeline
         from PIL import Image
-        from transformers import CLIPFeatureExtractor, CLIPModel
+        from transformers import CLIPImageProcessor, CLIPModel
 
-        feature_extractor = CLIPFeatureExtractor.from_pretrained(
+        feature_extractor = CLIPImageProcessor.from_pretrained(
             "laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
         )
         clip_model = CLIPModel.from_pretrained(
@@ -139,7 +139,7 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline, StableDiffusionMixin):
         tokenizer: CLIPTokenizer,
         unet: UNet2DConditionModel,
         scheduler: Union[PNDMScheduler, LMSDiscreteScheduler, DDIMScheduler, DPMSolverMultistepScheduler],
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
     ):
         super().__init__()
         self.register_modules(

examples/community/mixture_canvas.py

@@ -9,7 +9,7 @@ import torch
 from numpy import exp, pi, sqrt
 from torchvision.transforms.functional import resize
 from tqdm.auto import tqdm
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
@@ -275,7 +275,7 @@ class StableDiffusionCanvasPipeline(DiffusionPipeline, StableDiffusionMixin):
         unet: UNet2DConditionModel,
         scheduler: Union[DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler],
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
     ):
         super().__init__()
         self.register_modules(

examples/community/mixture_tiling.py

@@ -15,7 +15,7 @@ from diffusers.utils import logging
 
 try:
     from ligo.segments import segment
-    from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+    from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 except ImportError:
     raise ImportError("Please install transformers and ligo-segments to use the mixture pipeline")
 
@@ -144,7 +144,7 @@ class StableDiffusionTilingPipeline(DiffusionPipeline, StableDiffusionExtrasMixi
         unet: UNet2DConditionModel,
         scheduler: Union[DDIMScheduler, PNDMScheduler],
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
     ):
         super().__init__()
         self.register_modules(

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter.py

@@ -189,7 +189,7 @@ class StableDiffusionXLControlNetAdapterPipeline(
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py

@@ -332,7 +332,7 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
         requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`):
             Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config

examples/community/pipeline_zero1to3.py

@@ -9,7 +9,7 @@ import numpy as np
 import PIL.Image
 import torch
 from packaging import version
-from transformers import CLIPFeatureExtractor, CLIPVisionModelWithProjection
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
 
 # from ...configuration_utils import FrozenDict
 # from ...models import AutoencoderKL, UNet2DConditionModel
@@ -87,7 +87,7 @@ class Zero1to3StableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
         cc_projection ([`CCProjection`]):
             Projection layer to project the concated CLIP features and pose embeddings to the original CLIP feature size.
@@ -102,7 +102,7 @@ class Zero1to3StableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
         unet: UNet2DConditionModel,
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         cc_projection: CCProjection,
         requires_safety_checker: bool = True,
     ):

examples/community/regional_prompting_stable_diffusion.py

@@ -3,7 +3,7 @@ from typing import Dict, Optional
 
 import torch
 import torchvision.transforms.functional as FF
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from diffusers import StableDiffusionPipeline
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
@@ -69,7 +69,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
         unet: UNet2DConditionModel,
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         requires_safety_checker: bool = True,
     ):
         super().__init__(

examples/community/stable_diffusion_ipex.py

@@ -18,7 +18,7 @@ from typing import Any, Callable, Dict, List, Optional, Union
 import intel_extension_for_pytorch as ipex
 import torch
 from packaging import version
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from diffusers.configuration_utils import FrozenDict
 from diffusers.loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
@@ -86,7 +86,7 @@ class StableDiffusionIPEXPipeline(
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 
@@ -100,7 +100,7 @@ class StableDiffusionIPEXPipeline(
         unet: UNet2DConditionModel,
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         requires_safety_checker: bool = True,
     ):
         super().__init__()

examples/community/stable_diffusion_tensorrt_img2img.py

@@ -42,7 +42,7 @@ from polygraphy.backend.trt import (
     network_from_onnx_path,
     save_engine,
 )
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from diffusers import DiffusionPipeline
 from diffusers.configuration_utils import FrozenDict, deprecate
@@ -679,7 +679,7 @@ class TensorRTStableDiffusionImg2ImgPipeline(DiffusionPipeline):
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 
@@ -693,7 +693,7 @@ class TensorRTStableDiffusionImg2ImgPipeline(DiffusionPipeline):
         unet: UNet2DConditionModel,
         scheduler: DDIMScheduler,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         image_encoder: CLIPVisionModelWithProjection = None,
         requires_safety_checker: bool = True,
         stages=["clip", "unet", "vae", "vae_encoder"],

examples/community/stable_diffusion_tensorrt_inpaint.py

@@ -42,7 +42,7 @@ from polygraphy.backend.trt import (
     network_from_onnx_path,
     save_engine,
 )
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from diffusers import DiffusionPipeline
 from diffusers.configuration_utils import FrozenDict, deprecate
@@ -683,7 +683,7 @@ class TensorRTStableDiffusionInpaintPipeline(DiffusionPipeline):
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 
@@ -697,7 +697,7 @@ class TensorRTStableDiffusionInpaintPipeline(DiffusionPipeline):
         unet: UNet2DConditionModel,
         scheduler: DDIMScheduler,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         image_encoder: CLIPVisionModelWithProjection = None,
         requires_safety_checker: bool = True,
         stages=["clip", "unet", "vae", "vae_encoder"],

examples/community/stable_diffusion_tensorrt_txt2img.py

@@ -42,7 +42,7 @@ from polygraphy.backend.trt import (
     network_from_onnx_path,
     save_engine,
 )
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from diffusers import DiffusionPipeline
 from diffusers.configuration_utils import FrozenDict, deprecate
@@ -595,7 +595,7 @@ class TensorRTStableDiffusionPipeline(DiffusionPipeline):
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 
@@ -609,7 +609,7 @@ class TensorRTStableDiffusionPipeline(DiffusionPipeline):
         unet: UNet2DConditionModel,
         scheduler: DDIMScheduler,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         image_encoder: CLIPVisionModelWithProjection = None,
         requires_safety_checker: bool = True,
         stages=["clip", "unet", "vae"],

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -1271,7 +1271,7 @@ def main(args):
         lora_state_dict = StableDiffusion3Pipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")
+            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")

examples/research_projects/controlnet/train_controlnet_webdataset.py

@@ -43,7 +43,7 @@ from PIL import Image
 from torch.utils.data import default_collate
 from torchvision import transforms
 from tqdm.auto import tqdm
-from transformers import AutoTokenizer, DPTFeatureExtractor, DPTForDepthEstimation, PretrainedConfig
+from transformers import AutoTokenizer, DPTForDepthEstimation, DPTImageProcessor, PretrainedConfig
 from webdataset.tariterators import (
     base_plus_ext,
     tar_file_expander,
@@ -205,7 +205,7 @@ class Text2ImageDataset:
         pin_memory: bool = False,
         persistent_workers: bool = False,
         control_type: str = "canny",
-        feature_extractor: Optional[DPTFeatureExtractor] = None,
+        feature_extractor: Optional[DPTImageProcessor] = None,
     ):
         if not isinstance(train_shards_path_or_url, str):
             train_shards_path_or_url = [list(braceexpand(urls)) for urls in train_shards_path_or_url]
@@ -1011,7 +1011,7 @@ def main(args):
         controlnet = pre_controlnet
 
     if args.control_type == "depth":
-        feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-hybrid-midas")
+        feature_extractor = DPTImageProcessor.from_pretrained("Intel/dpt-hybrid-midas")
         depth_model = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas")
         depth_model.requires_grad_(False)
     else:

examples/research_projects/gligen/demo.ipynb

@@ -45,7 +45,7 @@
     "    UniPCMultistepScheduler,\n",
     "    EulerDiscreteScheduler,\n",
     ")\n",
-    "from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer\n",
+    "from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer\n",
     "# pretrained_model_name_or_path = 'masterful/gligen-1-4-generation-text-box'\n",
     "\n",
     "pretrained_model_name_or_path = '/root/data/zhizhonghuang/checkpoints/models--masterful--gligen-1-4-generation-text-box/snapshots/d2820dc1e9ba6ca082051ce79cfd3eb468ae2c83'\n",

examples/research_projects/rdm/pipeline_rdm.py

@@ -4,7 +4,7 @@ from typing import Callable, List, Optional, Union
 import torch
 from PIL import Image
 from retriever import Retriever, normalize_images, preprocess_images
-from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPModel, CLIPTokenizer
 
 from diffusers import (
     AutoencoderKL,
@@ -47,7 +47,7 @@ class RDMPipeline(DiffusionPipeline, StableDiffusionMixin):
         scheduler ([`SchedulerMixin`]):
             A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
             [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 
@@ -65,7 +65,7 @@ class RDMPipeline(DiffusionPipeline, StableDiffusionMixin):
             EulerAncestralDiscreteScheduler,
             DPMSolverMultistepScheduler,
         ],
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         retriever: Optional[Retriever] = None,
     ):
         super().__init__()

examples/research_projects/rdm/retriever.py

@@ -6,7 +6,7 @@ import numpy as np
 import torch
 from datasets import Dataset, load_dataset
 from PIL import Image
-from transformers import CLIPFeatureExtractor, CLIPModel, PretrainedConfig
+from transformers import CLIPImageProcessor, CLIPModel, PretrainedConfig
 
 from diffusers import logging
 
@@ -20,7 +20,7 @@ def normalize_images(images: List[Image.Image]):
     return images
 
 
-def preprocess_images(images: List[np.array], feature_extractor: CLIPFeatureExtractor) -> torch.Tensor:
+def preprocess_images(images: List[np.array], feature_extractor: CLIPImageProcessor) -> torch.Tensor:
     """
     Preprocesses a list of images into a batch of tensors.
 
@@ -95,14 +95,12 @@ class Index:
     def build_index(
         self,
         model=None,
-        feature_extractor: CLIPFeatureExtractor = None,
+        feature_extractor: CLIPImageProcessor = None,
         torch_dtype=torch.float32,
     ):
         if not self.index_initialized:
             model = model or CLIPModel.from_pretrained(self.config.clip_name_or_path).to(dtype=torch_dtype)
-            feature_extractor = feature_extractor or CLIPFeatureExtractor.from_pretrained(
-                self.config.clip_name_or_path
-            )
+            feature_extractor = feature_extractor or CLIPImageProcessor.from_pretrained(self.config.clip_name_or_path)
             self.dataset = get_dataset_with_emb_from_clip_model(
                 self.dataset,
                 model,
@@ -136,7 +134,7 @@ class Retriever:
         index: Index = None,
         dataset: Dataset = None,
         model=None,
-        feature_extractor: CLIPFeatureExtractor = None,
+        feature_extractor: CLIPImageProcessor = None,
     ):
         self.config = config
         self.index = index or self._build_index(config, dataset, model=model, feature_extractor=feature_extractor)
@@ -148,7 +146,7 @@ class Retriever:
         index: Index = None,
         dataset: Dataset = None,
         model=None,
-        feature_extractor: CLIPFeatureExtractor = None,
+        feature_extractor: CLIPImageProcessor = None,
         **kwargs,
     ):
         config = kwargs.pop("config", None) or IndexConfig.from_pretrained(retriever_name_or_path, **kwargs)
@@ -156,7 +154,7 @@ class Retriever:
 
     @staticmethod
     def _build_index(
-        config: IndexConfig, dataset: Dataset = None, model=None, feature_extractor: CLIPFeatureExtractor = None
+        config: IndexConfig, dataset: Dataset = None, model=None, feature_extractor: CLIPImageProcessor = None
     ):
         dataset = dataset or load_dataset(config.dataset_name)
         dataset = dataset[config.dataset_set]

examples/research_projects/sdxl_flax/sdxl_single_aot.py

@@ -18,7 +18,7 @@ cc.initialize_cache("/tmp/sdxl_cache")
 NUM_DEVICES = jax.device_count()
 
 # 1. Let's start by downloading the model and loading it into our pipeline class
-# Adhering to JAX's functional approach, the model's parameters are returned seperatetely and
+# Adhering to JAX's functional approach, the model's parameters are returned separately and
 # will have to be passed to the pipeline during inference
 pipeline, params = FlaxStableDiffusionXLPipeline.from_pretrained(
     "stabilityai/stable-diffusion-xl-base-1.0", revision="refs/pr/95", split_head_dim=True
@@ -69,7 +69,7 @@ def replicate_all(prompt_ids, neg_prompt_ids, seed):
 # to the function and tell JAX which are static arguments, that is, arguments that
 # are known at compile time and won't change. In our case, it is num_inference_steps,
 # height, width and return_latents.
-# Once the function is compiled, these parameters are ommited from future calls and
+# Once the function is compiled, these parameters are omitted from future calls and
 # cannot be changed without modifying the code and recompiling.
 def aot_compile(
     prompt=default_prompt,

examples/text_to_image/train_text_to_image.py

@@ -826,17 +826,22 @@ def main():
     )
 
     # Scheduler and math around the number of training steps.
-    overrode_max_train_steps = False
-    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * num_update_steps_per_epoch * accelerator.num_processes
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
     )
 
     # Prepare everything with our `accelerator`.
@@ -866,8 +871,14 @@ def main():
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps * accelerator.num_processes:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 

setup.py

@@ -135,6 +135,7 @@ _deps = [
     "transformers>=4.41.2",
     "urllib3<=2.0.0",
     "black",
+    "tokenizers==0.20.0rc1"
 ]
 
 # this is a lookup table with items like:

src/diffusers/__init__.py

@@ -12,6 +12,7 @@ from .utils import (
     is_note_seq_available,
     is_onnx_available,
     is_scipy_available,
+    is_sentencepiece_available,
     is_torch_available,
     is_torchsde_available,
     is_transformers_available,
@@ -250,8 +251,6 @@ else:
             "AuraFlowPipeline",
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
-            "ChatGLMModel",
-            "ChatGLMTokenizer",
             "CLIPImageProjection",
             "CogVideoXPipeline",
             "CycleDiffusionPipeline",
@@ -286,8 +285,6 @@ else:
             "KandinskyV22Pipeline",
             "KandinskyV22PriorEmb2EmbPipeline",
             "KandinskyV22PriorPipeline",
-            "KolorsImg2ImgPipeline",
-            "KolorsPipeline",
             "LatentConsistencyModelImg2ImgPipeline",
             "LatentConsistencyModelPipeline",
             "LattePipeline",
@@ -314,6 +311,7 @@ else:
             "StableDiffusion3ControlNetPipeline",
             "StableDiffusion3Img2ImgPipeline",
             "StableDiffusion3InpaintPipeline",
+            "StableDiffusion3PAGPipeline",
             "StableDiffusion3Pipeline",
             "StableDiffusionAdapterPipeline",
             "StableDiffusionAttendAndExcitePipeline",
@@ -391,6 +389,19 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["pipelines"].extend(["StableDiffusionKDiffusionPipeline", "StableDiffusionXLKDiffusionPipeline"])
 
+try:
+    if not (is_torch_available() and is_transformers_available() and is_sentencepiece_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_transformers_and_sentencepiece_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_transformers_and_sentencepiece_objects"] = [
+        name for name in dir(dummy_torch_and_transformers_and_sentencepiece_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["pipelines"].extend(["KolorsImg2ImgPipeline", "KolorsPAGPipeline", "KolorsPipeline"])
+
 try:
     if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
         raise OptionalDependencyNotAvailable()
@@ -679,8 +690,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDM2UNet2DConditionModel,
             AudioLDMPipeline,
             AuraFlowPipeline,
-            ChatGLMModel,
-            ChatGLMTokenizer,
             CLIPImageProjection,
             CogVideoXPipeline,
             CycleDiffusionPipeline,
@@ -715,8 +724,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             KandinskyV22Pipeline,
             KandinskyV22PriorEmb2EmbPipeline,
             KandinskyV22PriorPipeline,
-            KolorsImg2ImgPipeline,
-            KolorsPipeline,
             LatentConsistencyModelImg2ImgPipeline,
             LatentConsistencyModelPipeline,
             LattePipeline,
@@ -743,6 +750,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusion3ControlNetPipeline,
             StableDiffusion3Img2ImgPipeline,
             StableDiffusion3InpaintPipeline,
+            StableDiffusion3PAGPipeline,
             StableDiffusion3Pipeline,
             StableDiffusionAdapterPipeline,
             StableDiffusionAttendAndExcitePipeline,
@@ -814,6 +822,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     else:
         from .pipelines import StableDiffusionKDiffusionPipeline, StableDiffusionXLKDiffusionPipeline
 
+    try:
+        if not (is_torch_available() and is_transformers_available() and is_sentencepiece_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_transformers_and_sentencepiece_objects import *  # noqa F403
+    else:
+        from .pipelines import KolorsImg2ImgPipeline, KolorsPAGPipeline, KolorsPipeline
     try:
         if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
             raise OptionalDependencyNotAvailable()

src/diffusers/loaders/single_file_model.py

@@ -24,6 +24,7 @@ from .single_file_utils import (
     SingleFileComponentError,
     convert_animatediff_checkpoint_to_diffusers,
     convert_controlnet_checkpoint,
+    convert_flux_transformer_checkpoint_to_diffusers,
     convert_ldm_unet_checkpoint,
     convert_ldm_vae_checkpoint,
     convert_sd3_transformer_checkpoint_to_diffusers,
@@ -74,6 +75,13 @@ SINGLE_FILE_LOADABLE_CLASSES = {
     "MotionAdapter": {
         "checkpoint_mapping_fn": convert_animatediff_checkpoint_to_diffusers,
     },
+    "SparseControlNetModel": {
+        "checkpoint_mapping_fn": convert_animatediff_checkpoint_to_diffusers,
+    },
+    "FluxTransformer2DModel": {
+        "checkpoint_mapping_fn": convert_flux_transformer_checkpoint_to_diffusers,
+        "default_subfolder": "transformer",
+    },
 }
 
 

src/diffusers/loaders/single_file_utils.py

@@ -74,9 +74,12 @@ CHECKPOINT_KEY_NAMES = {
     "stable_cascade_stage_b": "down_blocks.1.0.channelwise.0.weight",
     "stable_cascade_stage_c": "clip_txt_mapper.weight",
     "sd3": "model.diffusion_model.joint_blocks.0.context_block.adaLN_modulation.1.bias",
-    "animatediff": "down_blocks.0.motion_modules.0.temporal_transformer.transformer_blocks.0.attention_blocks.1.pos_encoder.pe",
+    "animatediff": "down_blocks.0.motion_modules.0.temporal_transformer.transformer_blocks.0.attention_blocks.0.pos_encoder.pe",
     "animatediff_v2": "mid_block.motion_modules.0.temporal_transformer.norm.bias",
     "animatediff_sdxl_beta": "up_blocks.2.motion_modules.0.temporal_transformer.norm.weight",
+    "animatediff_scribble": "controlnet_cond_embedding.conv_in.weight",
+    "animatediff_rgb": "controlnet_cond_embedding.weight",
+    "flux": "double_blocks.0.img_attn.norm.key_norm.scale",
 }
 
 DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
@@ -110,6 +113,10 @@ DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
     "animatediff_v2": {"pretrained_model_name_or_path": "guoyww/animatediff-motion-adapter-v1-5-2"},
     "animatediff_v3": {"pretrained_model_name_or_path": "guoyww/animatediff-motion-adapter-v1-5-3"},
     "animatediff_sdxl_beta": {"pretrained_model_name_or_path": "guoyww/animatediff-motion-adapter-sdxl-beta"},
+    "animatediff_scribble": {"pretrained_model_name_or_path": "guoyww/animatediff-sparsectrl-scribble"},
+    "animatediff_rgb": {"pretrained_model_name_or_path": "guoyww/animatediff-sparsectrl-rgb"},
+    "flux-dev": {"pretrained_model_name_or_path": "black-forest-labs/FLUX.1-dev"},
+    "flux-schnell": {"pretrained_model_name_or_path": "black-forest-labs/FLUX.1-schnell"},
 }
 
 # Use to configure model sample size when original config is provided
@@ -491,7 +498,13 @@ def infer_diffusers_model_type(checkpoint):
         model_type = "sd3"
 
     elif CHECKPOINT_KEY_NAMES["animatediff"] in checkpoint:
-        if CHECKPOINT_KEY_NAMES["animatediff_v2"] in checkpoint:
+        if CHECKPOINT_KEY_NAMES["animatediff_scribble"] in checkpoint:
+            model_type = "animatediff_scribble"
+
+        elif CHECKPOINT_KEY_NAMES["animatediff_rgb"] in checkpoint:
+            model_type = "animatediff_rgb"
+
+        elif CHECKPOINT_KEY_NAMES["animatediff_v2"] in checkpoint:
             model_type = "animatediff_v2"
 
         elif checkpoint[CHECKPOINT_KEY_NAMES["animatediff_sdxl_beta"]].shape[-1] == 320:
@@ -503,6 +516,11 @@ def infer_diffusers_model_type(checkpoint):
         else:
             model_type = "animatediff_v3"
 
+    elif CHECKPOINT_KEY_NAMES["flux"] in checkpoint:
+        if "guidance_in.in_layer.bias" in checkpoint:
+            model_type = "flux-dev"
+        else:
+            model_type = "flux-schnell"
     else:
         model_type = "v1"
 
@@ -1859,3 +1877,195 @@ def convert_animatediff_checkpoint_to_diffusers(checkpoint, **kwargs):
             ] = v
 
     return converted_state_dict
+
+
+def convert_flux_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
+    converted_state_dict = {}
+
+    num_layers = list(set(int(k.split(".", 2)[1]) for k in checkpoint if "double_blocks." in k))[-1] + 1  # noqa: C401
+    num_single_layers = list(set(int(k.split(".", 2)[1]) for k in checkpoint if "single_blocks." in k))[-1] + 1  # noqa: C401
+    mlp_ratio = 4.0
+    inner_dim = 3072
+
+    # in SD3 original implementation of AdaLayerNormContinuous, it split linear projection output into shift, scale;
+    # while in diffusers it split into scale, shift. Here we swap the linear projection weights in order to be able to use diffusers implementation
+    def swap_scale_shift(weight):
+        shift, scale = weight.chunk(2, dim=0)
+        new_weight = torch.cat([scale, shift], dim=0)
+        return new_weight
+
+    ## time_text_embed.timestep_embedder <-  time_in
+    converted_state_dict["time_text_embed.timestep_embedder.linear_1.weight"] = checkpoint.pop(
+        "time_in.in_layer.weight"
+    )
+    converted_state_dict["time_text_embed.timestep_embedder.linear_1.bias"] = checkpoint.pop("time_in.in_layer.bias")
+    converted_state_dict["time_text_embed.timestep_embedder.linear_2.weight"] = checkpoint.pop(
+        "time_in.out_layer.weight"
+    )
+    converted_state_dict["time_text_embed.timestep_embedder.linear_2.bias"] = checkpoint.pop("time_in.out_layer.bias")
+
+    ## time_text_embed.text_embedder <- vector_in
+    converted_state_dict["time_text_embed.text_embedder.linear_1.weight"] = checkpoint.pop("vector_in.in_layer.weight")
+    converted_state_dict["time_text_embed.text_embedder.linear_1.bias"] = checkpoint.pop("vector_in.in_layer.bias")
+    converted_state_dict["time_text_embed.text_embedder.linear_2.weight"] = checkpoint.pop(
+        "vector_in.out_layer.weight"
+    )
+    converted_state_dict["time_text_embed.text_embedder.linear_2.bias"] = checkpoint.pop("vector_in.out_layer.bias")
+
+    # guidance
+    has_guidance = any("guidance" in k for k in checkpoint)
+    if has_guidance:
+        converted_state_dict["time_text_embed.guidance_embedder.linear_1.weight"] = checkpoint.pop(
+            "guidance_in.in_layer.weight"
+        )
+        converted_state_dict["time_text_embed.guidance_embedder.linear_1.bias"] = checkpoint.pop(
+            "guidance_in.in_layer.bias"
+        )
+        converted_state_dict["time_text_embed.guidance_embedder.linear_2.weight"] = checkpoint.pop(
+            "guidance_in.out_layer.weight"
+        )
+        converted_state_dict["time_text_embed.guidance_embedder.linear_2.bias"] = checkpoint.pop(
+            "guidance_in.out_layer.bias"
+        )
+
+    # context_embedder
+    converted_state_dict["context_embedder.weight"] = checkpoint.pop("txt_in.weight")
+    converted_state_dict["context_embedder.bias"] = checkpoint.pop("txt_in.bias")
+
+    # x_embedder
+    converted_state_dict["x_embedder.weight"] = checkpoint.pop("img_in.weight")
+    converted_state_dict["x_embedder.bias"] = checkpoint.pop("img_in.bias")
+
+    # double transformer blocks
+    for i in range(num_layers):
+        block_prefix = f"transformer_blocks.{i}."
+        # norms.
+        ## norm1
+        converted_state_dict[f"{block_prefix}norm1.linear.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.img_mod.lin.weight"
+        )
+        converted_state_dict[f"{block_prefix}norm1.linear.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.img_mod.lin.bias"
+        )
+        ## norm1_context
+        converted_state_dict[f"{block_prefix}norm1_context.linear.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mod.lin.weight"
+        )
+        converted_state_dict[f"{block_prefix}norm1_context.linear.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mod.lin.bias"
+        )
+        # Q, K, V
+        sample_q, sample_k, sample_v = torch.chunk(checkpoint.pop(f"double_blocks.{i}.img_attn.qkv.weight"), 3, dim=0)
+        context_q, context_k, context_v = torch.chunk(
+            checkpoint.pop(f"double_blocks.{i}.txt_attn.qkv.weight"), 3, dim=0
+        )
+        sample_q_bias, sample_k_bias, sample_v_bias = torch.chunk(
+            checkpoint.pop(f"double_blocks.{i}.img_attn.qkv.bias"), 3, dim=0
+        )
+        context_q_bias, context_k_bias, context_v_bias = torch.chunk(
+            checkpoint.pop(f"double_blocks.{i}.txt_attn.qkv.bias"), 3, dim=0
+        )
+        converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([sample_q])
+        converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([sample_q_bias])
+        converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([sample_k])
+        converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([sample_k_bias])
+        converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([sample_v])
+        converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([sample_v_bias])
+        converted_state_dict[f"{block_prefix}attn.add_q_proj.weight"] = torch.cat([context_q])
+        converted_state_dict[f"{block_prefix}attn.add_q_proj.bias"] = torch.cat([context_q_bias])
+        converted_state_dict[f"{block_prefix}attn.add_k_proj.weight"] = torch.cat([context_k])
+        converted_state_dict[f"{block_prefix}attn.add_k_proj.bias"] = torch.cat([context_k_bias])
+        converted_state_dict[f"{block_prefix}attn.add_v_proj.weight"] = torch.cat([context_v])
+        converted_state_dict[f"{block_prefix}attn.add_v_proj.bias"] = torch.cat([context_v_bias])
+        # qk_norm
+        converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.img_attn.norm.query_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.img_attn.norm.key_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_added_q.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_attn.norm.query_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_added_k.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_attn.norm.key_norm.scale"
+        )
+        # ff img_mlp
+        converted_state_dict[f"{block_prefix}ff.net.0.proj.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.img_mlp.0.weight"
+        )
+        converted_state_dict[f"{block_prefix}ff.net.0.proj.bias"] = checkpoint.pop(f"double_blocks.{i}.img_mlp.0.bias")
+        converted_state_dict[f"{block_prefix}ff.net.2.weight"] = checkpoint.pop(f"double_blocks.{i}.img_mlp.2.weight")
+        converted_state_dict[f"{block_prefix}ff.net.2.bias"] = checkpoint.pop(f"double_blocks.{i}.img_mlp.2.bias")
+        converted_state_dict[f"{block_prefix}ff_context.net.0.proj.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mlp.0.weight"
+        )
+        converted_state_dict[f"{block_prefix}ff_context.net.0.proj.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mlp.0.bias"
+        )
+        converted_state_dict[f"{block_prefix}ff_context.net.2.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mlp.2.weight"
+        )
+        converted_state_dict[f"{block_prefix}ff_context.net.2.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_mlp.2.bias"
+        )
+        # output projections.
+        converted_state_dict[f"{block_prefix}attn.to_out.0.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.img_attn.proj.weight"
+        )
+        converted_state_dict[f"{block_prefix}attn.to_out.0.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.img_attn.proj.bias"
+        )
+        converted_state_dict[f"{block_prefix}attn.to_add_out.weight"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_attn.proj.weight"
+        )
+        converted_state_dict[f"{block_prefix}attn.to_add_out.bias"] = checkpoint.pop(
+            f"double_blocks.{i}.txt_attn.proj.bias"
+        )
+
+    # single transfomer blocks
+    for i in range(num_single_layers):
+        block_prefix = f"single_transformer_blocks.{i}."
+        # norm.linear  <- single_blocks.0.modulation.lin
+        converted_state_dict[f"{block_prefix}norm.linear.weight"] = checkpoint.pop(
+            f"single_blocks.{i}.modulation.lin.weight"
+        )
+        converted_state_dict[f"{block_prefix}norm.linear.bias"] = checkpoint.pop(
+            f"single_blocks.{i}.modulation.lin.bias"
+        )
+        # Q, K, V, mlp
+        mlp_hidden_dim = int(inner_dim * mlp_ratio)
+        split_size = (inner_dim, inner_dim, inner_dim, mlp_hidden_dim)
+        q, k, v, mlp = torch.split(checkpoint.pop(f"single_blocks.{i}.linear1.weight"), split_size, dim=0)
+        q_bias, k_bias, v_bias, mlp_bias = torch.split(
+            checkpoint.pop(f"single_blocks.{i}.linear1.bias"), split_size, dim=0
+        )
+        converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([q])
+        converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([q_bias])
+        converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([k])
+        converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([k_bias])
+        converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([v])
+        converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([v_bias])
+        converted_state_dict[f"{block_prefix}proj_mlp.weight"] = torch.cat([mlp])
+        converted_state_dict[f"{block_prefix}proj_mlp.bias"] = torch.cat([mlp_bias])
+        # qk norm
+        converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = checkpoint.pop(
+            f"single_blocks.{i}.norm.query_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = checkpoint.pop(
+            f"single_blocks.{i}.norm.key_norm.scale"
+        )
+        # output projections.
+        converted_state_dict[f"{block_prefix}proj_out.weight"] = checkpoint.pop(f"single_blocks.{i}.linear2.weight")
+        converted_state_dict[f"{block_prefix}proj_out.bias"] = checkpoint.pop(f"single_blocks.{i}.linear2.bias")
+
+    converted_state_dict["proj_out.weight"] = checkpoint.pop("final_layer.linear.weight")
+    converted_state_dict["proj_out.bias"] = checkpoint.pop("final_layer.linear.bias")
+    converted_state_dict["norm_out.linear.weight"] = swap_scale_shift(
+        checkpoint.pop("final_layer.adaLN_modulation.1.weight")
+    )
+    converted_state_dict["norm_out.linear.bias"] = swap_scale_shift(
+        checkpoint.pop("final_layer.adaLN_modulation.1.bias")
+    )
+
+    return converted_state_dict

src/diffusers/models/attention.py

@@ -11,7 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import Any, Dict, Optional
+from typing import Any, Dict, List, Optional, Tuple
 
 import torch
 import torch.nn.functional as F
@@ -272,6 +272,17 @@ class BasicTransformerBlock(nn.Module):
         attention_out_bias: bool = True,
     ):
         super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.dropout = dropout
+        self.cross_attention_dim = cross_attention_dim
+        self.activation_fn = activation_fn
+        self.attention_bias = attention_bias
+        self.double_self_attention = double_self_attention
+        self.norm_elementwise_affine = norm_elementwise_affine
+        self.positional_embeddings = positional_embeddings
+        self.num_positional_embeddings = num_positional_embeddings
         self.only_cross_attention = only_cross_attention
 
         # We keep these boolean flags for backward-compatibility.
@@ -376,7 +387,7 @@ class BasicTransformerBlock(nn.Module):
                 "layer_norm",
             )
 
-        elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]:
+        elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm"]:
             self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
         elif norm_type == "layer_norm_i2vgen":
             self.norm3 = None
@@ -782,6 +793,319 @@ class SkipFFTransformerBlock(nn.Module):
         return hidden_states
 
 
+@maybe_allow_in_graph
+class FreeNoiseTransformerBlock(nn.Module):
+    r"""
+    A FreeNoise Transformer block.
+
+    Parameters:
+        dim (`int`):
+            The number of channels in the input and output.
+        num_attention_heads (`int`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`):
+            The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability to use.
+        cross_attention_dim (`int`, *optional*):
+            The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`):
+            Activation function to be used in feed-forward.
+        num_embeds_ada_norm (`int`, *optional*):
+            The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (`bool`, defaults to `False`):
+            Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, defaults to `False`):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, defaults to `False`):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, defaults to `False`):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` defaults to `False`):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+        ff_inner_dim (`int`, *optional*):
+            Hidden dimension of feed-forward MLP.
+        ff_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in feed-forward MLP.
+        attention_out_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in attention output project layer.
+        context_length (`int`, defaults to `16`):
+            The maximum number of frames that the FreeNoise block processes at once.
+        context_stride (`int`, defaults to `4`):
+            The number of frames to be skipped before starting to process a new batch of `context_length` frames.
+        weighting_scheme (`str`, defaults to `"pyramid"`):
+            The weighting scheme to use for weighting averaging of processed latent frames. As described in the
+            Equation 9. of the [FreeNoise](https://arxiv.org/abs/2310.15169) paper, "pyramid" is the default setting
+            used.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout: float = 0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+        context_length: int = 16,
+        context_stride: int = 4,
+        weighting_scheme: str = "pyramid",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.dropout = dropout
+        self.cross_attention_dim = cross_attention_dim
+        self.activation_fn = activation_fn
+        self.attention_bias = attention_bias
+        self.double_self_attention = double_self_attention
+        self.norm_elementwise_affine = norm_elementwise_affine
+        self.positional_embeddings = positional_embeddings
+        self.num_positional_embeddings = num_positional_embeddings
+        self.only_cross_attention = only_cross_attention
+
+        self.set_free_noise_properties(context_length, context_stride, weighting_scheme)
+
+        # We keep these boolean flags for backward-compatibility.
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        self.norm_type = norm_type
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+            )  # is self-attn if encoder_hidden_states is none
+
+        # 3. Feed-forward
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+        self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def _get_frame_indices(self, num_frames: int) -> List[Tuple[int, int]]:
+        frame_indices = []
+        for i in range(0, num_frames - self.context_length + 1, self.context_stride):
+            window_start = i
+            window_end = min(num_frames, i + self.context_length)
+            frame_indices.append((window_start, window_end))
+        return frame_indices
+
+    def _get_frame_weights(self, num_frames: int, weighting_scheme: str = "pyramid") -> List[float]:
+        if weighting_scheme == "pyramid":
+            if num_frames % 2 == 0:
+                # num_frames = 4 => [1, 2, 2, 1]
+                weights = list(range(1, num_frames // 2 + 1))
+                weights = weights + weights[::-1]
+            else:
+                # num_frames = 5 => [1, 2, 3, 2, 1]
+                weights = list(range(1, num_frames // 2 + 1))
+                weights = weights + [num_frames // 2 + 1] + weights[::-1]
+        else:
+            raise ValueError(f"Unsupported value for weighting_scheme={weighting_scheme}")
+
+        return weights
+
+    def set_free_noise_properties(
+        self, context_length: int, context_stride: int, weighting_scheme: str = "pyramid"
+    ) -> None:
+        self.context_length = context_length
+        self.context_stride = context_stride
+        self.weighting_scheme = weighting_scheme
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0) -> None:
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
+
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        # hidden_states: [B x H x W, F, C]
+        device = hidden_states.device
+        dtype = hidden_states.dtype
+
+        num_frames = hidden_states.size(1)
+        frame_indices = self._get_frame_indices(num_frames)
+        frame_weights = self._get_frame_weights(self.context_length, self.weighting_scheme)
+        frame_weights = torch.tensor(frame_weights, device=device, dtype=dtype).unsqueeze(0).unsqueeze(-1)
+        is_last_frame_batch_complete = frame_indices[-1][1] == num_frames
+
+        # Handle out-of-bounds case if num_frames isn't perfectly divisible by context_length
+        # For example, num_frames=25, context_length=16, context_stride=4, then we expect the ranges:
+        #    [(0, 16), (4, 20), (8, 24), (10, 26)]
+        if not is_last_frame_batch_complete:
+            if num_frames < self.context_length:
+                raise ValueError(f"Expected {num_frames=} to be greater or equal than {self.context_length=}")
+            last_frame_batch_length = num_frames - frame_indices[-1][1]
+            frame_indices.append((num_frames - self.context_length, num_frames))
+
+        num_times_accumulated = torch.zeros((1, num_frames, 1), device=device)
+        accumulated_values = torch.zeros_like(hidden_states)
+
+        for i, (frame_start, frame_end) in enumerate(frame_indices):
+            # The reason for slicing here is to ensure that if (frame_end - frame_start) is to handle
+            # cases like frame_indices=[(0, 16), (16, 20)], if the user provided a video with 19 frames, or
+            # essentially a non-multiple of `context_length`.
+            weights = torch.ones_like(num_times_accumulated[:, frame_start:frame_end])
+            weights *= frame_weights
+
+            hidden_states_chunk = hidden_states[:, frame_start:frame_end]
+
+            # Notice that normalization is always applied before the real computation in the following blocks.
+            # 1. Self-Attention
+            norm_hidden_states = self.norm1(hidden_states_chunk)
+
+            if self.pos_embed is not None:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn1(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+                attention_mask=attention_mask,
+                **cross_attention_kwargs,
+            )
+
+            hidden_states_chunk = attn_output + hidden_states_chunk
+            if hidden_states_chunk.ndim == 4:
+                hidden_states_chunk = hidden_states_chunk.squeeze(1)
+
+            # 2. Cross-Attention
+            if self.attn2 is not None:
+                norm_hidden_states = self.norm2(hidden_states_chunk)
+
+                if self.pos_embed is not None and self.norm_type != "ada_norm_single":
+                    norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+                attn_output = self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=encoder_attention_mask,
+                    **cross_attention_kwargs,
+                )
+                hidden_states_chunk = attn_output + hidden_states_chunk
+
+            if i == len(frame_indices) - 1 and not is_last_frame_batch_complete:
+                accumulated_values[:, -last_frame_batch_length:] += (
+                    hidden_states_chunk[:, -last_frame_batch_length:] * weights[:, -last_frame_batch_length:]
+                )
+                num_times_accumulated[:, -last_frame_batch_length:] += weights[:, -last_frame_batch_length]
+            else:
+                accumulated_values[:, frame_start:frame_end] += hidden_states_chunk * weights
+                num_times_accumulated[:, frame_start:frame_end] += weights
+
+        hidden_states = torch.where(
+            num_times_accumulated > 0, accumulated_values / num_times_accumulated, accumulated_values
+        ).to(dtype)
+
+        # 3. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
 class FeedForward(nn.Module):
     r"""
     A feed-forward layer.

src/diffusers/models/attention_processor.py

@@ -227,6 +227,7 @@ class Attention(nn.Module):
             self.to_k = None
             self.to_v = None
 
+        self.added_proj_bias = added_proj_bias
         if self.added_kv_proj_dim is not None:
             self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias)
             self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias)
@@ -698,12 +699,15 @@ class Attention(nn.Module):
             in_features = concatenated_weights.shape[1]
             out_features = concatenated_weights.shape[0]
 
-            self.to_added_qkv = nn.Linear(in_features, out_features, bias=True, device=device, dtype=dtype)
-            self.to_added_qkv.weight.copy_(concatenated_weights)
-            concatenated_bias = torch.cat(
-                [self.add_q_proj.bias.data, self.add_k_proj.bias.data, self.add_v_proj.bias.data]
+            self.to_added_qkv = nn.Linear(
+                in_features, out_features, bias=self.added_proj_bias, device=device, dtype=dtype
             )
-            self.to_added_qkv.bias.copy_(concatenated_bias)
+            self.to_added_qkv.weight.copy_(concatenated_weights)
+            if self.added_proj_bias:
+                concatenated_bias = torch.cat(
+                    [self.add_q_proj.bias.data, self.add_k_proj.bias.data, self.add_v_proj.bias.data]
+                )
+                self.to_added_qkv.bias.copy_(concatenated_bias)
 
         self.fused_projections = fuse
 
@@ -1102,6 +1106,326 @@ class JointAttnProcessor2_0:
         return hidden_states, encoder_hidden_states
 
 
+class PAGJointAttnProcessor2_0:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "PAGJointAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        input_ndim = hidden_states.ndim
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+        context_input_ndim = encoder_hidden_states.ndim
+        if context_input_ndim == 4:
+            batch_size, channel, height, width = encoder_hidden_states.shape
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        # store the length of image patch sequences to create a mask that prevents interaction between patches
+        # similar to making the self-attention map an identity matrix
+        identity_block_size = hidden_states.shape[1]
+
+        # chunk
+        hidden_states_org, hidden_states_ptb = hidden_states.chunk(2)
+        encoder_hidden_states_org, encoder_hidden_states_ptb = encoder_hidden_states.chunk(2)
+
+        ################## original path ##################
+        batch_size = encoder_hidden_states_org.shape[0]
+
+        # `sample` projections.
+        query_org = attn.to_q(hidden_states_org)
+        key_org = attn.to_k(hidden_states_org)
+        value_org = attn.to_v(hidden_states_org)
+
+        # `context` projections.
+        encoder_hidden_states_org_query_proj = attn.add_q_proj(encoder_hidden_states_org)
+        encoder_hidden_states_org_key_proj = attn.add_k_proj(encoder_hidden_states_org)
+        encoder_hidden_states_org_value_proj = attn.add_v_proj(encoder_hidden_states_org)
+
+        # attention
+        query_org = torch.cat([query_org, encoder_hidden_states_org_query_proj], dim=1)
+        key_org = torch.cat([key_org, encoder_hidden_states_org_key_proj], dim=1)
+        value_org = torch.cat([value_org, encoder_hidden_states_org_value_proj], dim=1)
+
+        inner_dim = key_org.shape[-1]
+        head_dim = inner_dim // attn.heads
+        query_org = query_org.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key_org = key_org.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value_org = value_org.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        hidden_states_org = F.scaled_dot_product_attention(
+            query_org, key_org, value_org, dropout_p=0.0, is_causal=False
+        )
+        hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states_org = hidden_states_org.to(query_org.dtype)
+
+        # Split the attention outputs.
+        hidden_states_org, encoder_hidden_states_org = (
+            hidden_states_org[:, : residual.shape[1]],
+            hidden_states_org[:, residual.shape[1] :],
+        )
+
+        # linear proj
+        hidden_states_org = attn.to_out[0](hidden_states_org)
+        # dropout
+        hidden_states_org = attn.to_out[1](hidden_states_org)
+        if not attn.context_pre_only:
+            encoder_hidden_states_org = attn.to_add_out(encoder_hidden_states_org)
+
+        if input_ndim == 4:
+            hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width)
+        if context_input_ndim == 4:
+            encoder_hidden_states_org = encoder_hidden_states_org.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        ################## perturbed path ##################
+
+        batch_size = encoder_hidden_states_ptb.shape[0]
+
+        # `sample` projections.
+        query_ptb = attn.to_q(hidden_states_ptb)
+        key_ptb = attn.to_k(hidden_states_ptb)
+        value_ptb = attn.to_v(hidden_states_ptb)
+
+        # `context` projections.
+        encoder_hidden_states_ptb_query_proj = attn.add_q_proj(encoder_hidden_states_ptb)
+        encoder_hidden_states_ptb_key_proj = attn.add_k_proj(encoder_hidden_states_ptb)
+        encoder_hidden_states_ptb_value_proj = attn.add_v_proj(encoder_hidden_states_ptb)
+
+        # attention
+        query_ptb = torch.cat([query_ptb, encoder_hidden_states_ptb_query_proj], dim=1)
+        key_ptb = torch.cat([key_ptb, encoder_hidden_states_ptb_key_proj], dim=1)
+        value_ptb = torch.cat([value_ptb, encoder_hidden_states_ptb_value_proj], dim=1)
+
+        inner_dim = key_ptb.shape[-1]
+        head_dim = inner_dim // attn.heads
+        query_ptb = query_ptb.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key_ptb = key_ptb.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value_ptb = value_ptb.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # create a full mask with all entries set to 0
+        seq_len = query_ptb.size(2)
+        full_mask = torch.zeros((seq_len, seq_len), device=query_ptb.device, dtype=query_ptb.dtype)
+
+        # set the attention value between image patches to -inf
+        full_mask[:identity_block_size, :identity_block_size] = float("-inf")
+
+        # set the diagonal of the attention value between image patches to 0
+        full_mask[:identity_block_size, :identity_block_size].fill_diagonal_(0)
+
+        # expand the mask to match the attention weights shape
+        full_mask = full_mask.unsqueeze(0).unsqueeze(0)  # Add batch and num_heads dimensions
+
+        hidden_states_ptb = F.scaled_dot_product_attention(
+            query_ptb, key_ptb, value_ptb, attn_mask=full_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states_ptb = hidden_states_ptb.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states_ptb = hidden_states_ptb.to(query_ptb.dtype)
+
+        # split the attention outputs.
+        hidden_states_ptb, encoder_hidden_states_ptb = (
+            hidden_states_ptb[:, : residual.shape[1]],
+            hidden_states_ptb[:, residual.shape[1] :],
+        )
+
+        # linear proj
+        hidden_states_ptb = attn.to_out[0](hidden_states_ptb)
+        # dropout
+        hidden_states_ptb = attn.to_out[1](hidden_states_ptb)
+        if not attn.context_pre_only:
+            encoder_hidden_states_ptb = attn.to_add_out(encoder_hidden_states_ptb)
+
+        if input_ndim == 4:
+            hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width)
+        if context_input_ndim == 4:
+            encoder_hidden_states_ptb = encoder_hidden_states_ptb.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        ################ concat ###############
+        hidden_states = torch.cat([hidden_states_org, hidden_states_ptb])
+        encoder_hidden_states = torch.cat([encoder_hidden_states_org, encoder_hidden_states_ptb])
+
+        return hidden_states, encoder_hidden_states
+
+
+class PAGCFGJointAttnProcessor2_0:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "PAGCFGJointAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        input_ndim = hidden_states.ndim
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+        context_input_ndim = encoder_hidden_states.ndim
+        if context_input_ndim == 4:
+            batch_size, channel, height, width = encoder_hidden_states.shape
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        identity_block_size = hidden_states.shape[
+            1
+        ]  # patch embeddings width * height (correspond to self-attention map width or height)
+
+        # chunk
+        hidden_states_uncond, hidden_states_org, hidden_states_ptb = hidden_states.chunk(3)
+        hidden_states_org = torch.cat([hidden_states_uncond, hidden_states_org])
+
+        (
+            encoder_hidden_states_uncond,
+            encoder_hidden_states_org,
+            encoder_hidden_states_ptb,
+        ) = encoder_hidden_states.chunk(3)
+        encoder_hidden_states_org = torch.cat([encoder_hidden_states_uncond, encoder_hidden_states_org])
+
+        ################## original path ##################
+        batch_size = encoder_hidden_states_org.shape[0]
+
+        # `sample` projections.
+        query_org = attn.to_q(hidden_states_org)
+        key_org = attn.to_k(hidden_states_org)
+        value_org = attn.to_v(hidden_states_org)
+
+        # `context` projections.
+        encoder_hidden_states_org_query_proj = attn.add_q_proj(encoder_hidden_states_org)
+        encoder_hidden_states_org_key_proj = attn.add_k_proj(encoder_hidden_states_org)
+        encoder_hidden_states_org_value_proj = attn.add_v_proj(encoder_hidden_states_org)
+
+        # attention
+        query_org = torch.cat([query_org, encoder_hidden_states_org_query_proj], dim=1)
+        key_org = torch.cat([key_org, encoder_hidden_states_org_key_proj], dim=1)
+        value_org = torch.cat([value_org, encoder_hidden_states_org_value_proj], dim=1)
+
+        inner_dim = key_org.shape[-1]
+        head_dim = inner_dim // attn.heads
+        query_org = query_org.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key_org = key_org.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value_org = value_org.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        hidden_states_org = F.scaled_dot_product_attention(
+            query_org, key_org, value_org, dropout_p=0.0, is_causal=False
+        )
+        hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states_org = hidden_states_org.to(query_org.dtype)
+
+        # Split the attention outputs.
+        hidden_states_org, encoder_hidden_states_org = (
+            hidden_states_org[:, : residual.shape[1]],
+            hidden_states_org[:, residual.shape[1] :],
+        )
+
+        # linear proj
+        hidden_states_org = attn.to_out[0](hidden_states_org)
+        # dropout
+        hidden_states_org = attn.to_out[1](hidden_states_org)
+        if not attn.context_pre_only:
+            encoder_hidden_states_org = attn.to_add_out(encoder_hidden_states_org)
+
+        if input_ndim == 4:
+            hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width)
+        if context_input_ndim == 4:
+            encoder_hidden_states_org = encoder_hidden_states_org.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        ################## perturbed path ##################
+
+        batch_size = encoder_hidden_states_ptb.shape[0]
+
+        # `sample` projections.
+        query_ptb = attn.to_q(hidden_states_ptb)
+        key_ptb = attn.to_k(hidden_states_ptb)
+        value_ptb = attn.to_v(hidden_states_ptb)
+
+        # `context` projections.
+        encoder_hidden_states_ptb_query_proj = attn.add_q_proj(encoder_hidden_states_ptb)
+        encoder_hidden_states_ptb_key_proj = attn.add_k_proj(encoder_hidden_states_ptb)
+        encoder_hidden_states_ptb_value_proj = attn.add_v_proj(encoder_hidden_states_ptb)
+
+        # attention
+        query_ptb = torch.cat([query_ptb, encoder_hidden_states_ptb_query_proj], dim=1)
+        key_ptb = torch.cat([key_ptb, encoder_hidden_states_ptb_key_proj], dim=1)
+        value_ptb = torch.cat([value_ptb, encoder_hidden_states_ptb_value_proj], dim=1)
+
+        inner_dim = key_ptb.shape[-1]
+        head_dim = inner_dim // attn.heads
+        query_ptb = query_ptb.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key_ptb = key_ptb.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value_ptb = value_ptb.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # create a full mask with all entries set to 0
+        seq_len = query_ptb.size(2)
+        full_mask = torch.zeros((seq_len, seq_len), device=query_ptb.device, dtype=query_ptb.dtype)
+
+        # set the attention value between image patches to -inf
+        full_mask[:identity_block_size, :identity_block_size] = float("-inf")
+
+        # set the diagonal of the attention value between image patches to 0
+        full_mask[:identity_block_size, :identity_block_size].fill_diagonal_(0)
+
+        # expand the mask to match the attention weights shape
+        full_mask = full_mask.unsqueeze(0).unsqueeze(0)  # Add batch and num_heads dimensions
+
+        hidden_states_ptb = F.scaled_dot_product_attention(
+            query_ptb, key_ptb, value_ptb, attn_mask=full_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states_ptb = hidden_states_ptb.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states_ptb = hidden_states_ptb.to(query_ptb.dtype)
+
+        # split the attention outputs.
+        hidden_states_ptb, encoder_hidden_states_ptb = (
+            hidden_states_ptb[:, : residual.shape[1]],
+            hidden_states_ptb[:, residual.shape[1] :],
+        )
+
+        # linear proj
+        hidden_states_ptb = attn.to_out[0](hidden_states_ptb)
+        # dropout
+        hidden_states_ptb = attn.to_out[1](hidden_states_ptb)
+        if not attn.context_pre_only:
+            encoder_hidden_states_ptb = attn.to_add_out(encoder_hidden_states_ptb)
+
+        if input_ndim == 4:
+            hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width)
+        if context_input_ndim == 4:
+            encoder_hidden_states_ptb = encoder_hidden_states_ptb.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        ################ concat ###############
+        hidden_states = torch.cat([hidden_states_org, hidden_states_ptb])
+        encoder_hidden_states = torch.cat([encoder_hidden_states_org, encoder_hidden_states_ptb])
+
+        return hidden_states, encoder_hidden_states
+
+
 class FusedJointAttnProcessor2_0:
     """Attention processor used typically in processing the SD3-like self-attention projections."""
 
@@ -1274,6 +1598,103 @@ class AuraFlowAttnProcessor2_0:
             return hidden_states
 
 
+class FusedAuraFlowAttnProcessor2_0:
+    """Attention processor used typically in processing Aura Flow with fused projections."""
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention") and is_torch_version("<", "2.1"):
+            raise ImportError(
+                "FusedAuraFlowAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to at least 2.1 or above as we use `scale` in `F.scaled_dot_product_attention()`. "
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        batch_size = hidden_states.shape[0]
+
+        # `sample` projections.
+        qkv = attn.to_qkv(hidden_states)
+        split_size = qkv.shape[-1] // 3
+        query, key, value = torch.split(qkv, split_size, dim=-1)
+
+        # `context` projections.
+        if encoder_hidden_states is not None:
+            encoder_qkv = attn.to_added_qkv(encoder_hidden_states)
+            split_size = encoder_qkv.shape[-1] // 3
+            (
+                encoder_hidden_states_query_proj,
+                encoder_hidden_states_key_proj,
+                encoder_hidden_states_value_proj,
+            ) = torch.split(encoder_qkv, split_size, dim=-1)
+
+        # Reshape.
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+        query = query.view(batch_size, -1, attn.heads, head_dim)
+        key = key.view(batch_size, -1, attn.heads, head_dim)
+        value = value.view(batch_size, -1, attn.heads, head_dim)
+
+        # Apply QK norm.
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Concatenate the projections.
+        if encoder_hidden_states is not None:
+            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            )
+            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(batch_size, -1, attn.heads, head_dim)
+            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            )
+
+            if attn.norm_added_q is not None:
+                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
+            if attn.norm_added_k is not None:
+                encoder_hidden_states_key_proj = attn.norm_added_q(encoder_hidden_states_key_proj)
+
+            query = torch.cat([encoder_hidden_states_query_proj, query], dim=1)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
+
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+
+        # Attention.
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, dropout_p=0.0, scale=attn.scale, is_causal=False
+        )
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # Split the attention outputs.
+        if encoder_hidden_states is not None:
+            hidden_states, encoder_hidden_states = (
+                hidden_states[:, encoder_hidden_states.shape[1] :],
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+            )
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+        if encoder_hidden_states is not None:
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        if encoder_hidden_states is not None:
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states
+
+
 # YiYi to-do: refactor rope related functions/classes
 def apply_rope(xq, xk, freqs_cis):
     xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -1,3 +1,18 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from typing import Optional, Tuple, Union
 
 import numpy as np
@@ -7,6 +22,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders.single_file_model import FromOriginalModelMixin
+from ...utils import logging
 from ...utils.accelerate_utils import apply_forward_hook
 from ..activations import get_activation
 from ..downsampling import CogVideoXDownsample3D
@@ -16,6 +32,9 @@ from ..upsampling import CogVideoXUpsample3D
 from .vae import DecoderOutput, DiagonalGaussianDistribution
 
 
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
 class CogVideoXSafeConv3d(nn.Conv3d):
     """
     A 3D convolution layer that splits the input tensor into smaller parts to avoid OOM in CogVideoX Model.
@@ -98,7 +117,7 @@ class CogVideoXCausalConv3d(nn.Module):
 
         self.conv_cache = None
 
-    def fake_cp_pass_from_previous_rank(self, inputs: torch.Tensor) -> torch.Tensor:
+    def fake_context_parallel_forward(self, inputs: torch.Tensor) -> torch.Tensor:
         dim = self.temporal_dim
         kernel_size = self.time_kernel_size
         if kernel_size == 1:
@@ -114,13 +133,15 @@ class CogVideoXCausalConv3d(nn.Module):
         inputs = inputs.transpose(0, dim).contiguous()
         return inputs
 
-    def forward(self, inputs: torch.Tensor, clear_fake_cp_cache: bool = True):
-        input_parallel = self.fake_cp_pass_from_previous_rank(inputs)
-
+    def _clear_fake_context_parallel_cache(self):
         del self.conv_cache
         self.conv_cache = None
-        if not clear_fake_cp_cache:
-            self.conv_cache = input_parallel[:, :, -self.time_kernel_size + 1 :].contiguous().detach().clone().cpu()
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        input_parallel = self.fake_context_parallel_forward(inputs)
+
+        self._clear_fake_context_parallel_cache()
+        self.conv_cache = input_parallel[:, :, -self.time_kernel_size + 1 :].contiguous().detach().clone().cpu()
 
         padding_2d = (self.width_pad, self.width_pad, self.height_pad, self.height_pad)
         input_parallel = F.pad(input_parallel, padding_2d, mode="constant", value=0)
@@ -148,9 +169,10 @@ class CogVideoXSpatialNorm3D(nn.Module):
         self,
         f_channels: int,
         zq_channels: int,
+        groups: int = 32,
     ):
         super().__init__()
-        self.norm_layer = nn.GroupNorm(num_channels=f_channels, num_groups=32, eps=1e-6, affine=True)
+        self.norm_layer = nn.GroupNorm(num_channels=f_channels, num_groups=groups, eps=1e-6, affine=True)
         self.conv_y = CogVideoXCausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
         self.conv_b = CogVideoXCausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
 
@@ -217,10 +239,12 @@ class CogVideoXResnetBlock3D(nn.Module):
             self.norm1 = CogVideoXSpatialNorm3D(
                 f_channels=in_channels,
                 zq_channels=spatial_norm_dim,
+                groups=groups,
             )
             self.norm2 = CogVideoXSpatialNorm3D(
                 f_channels=out_channels,
                 zq_channels=spatial_norm_dim,
+                groups=groups,
             )
 
         self.conv1 = CogVideoXCausalConv3d(
@@ -250,15 +274,16 @@ class CogVideoXResnetBlock3D(nn.Module):
         inputs: torch.Tensor,
         temb: Optional[torch.Tensor] = None,
         zq: Optional[torch.Tensor] = None,
-        clear_fake_cp_cache: bool = True,
     ) -> torch.Tensor:
         hidden_states = inputs
+
         if zq is not None:
             hidden_states = self.norm1(hidden_states, zq)
         else:
             hidden_states = self.norm1(hidden_states)
+
         hidden_states = self.nonlinearity(hidden_states)
-        hidden_states = self.conv1(hidden_states, clear_fake_cp_cache=clear_fake_cp_cache)
+        hidden_states = self.conv1(hidden_states)
 
         if temb is not None:
             hidden_states = hidden_states + self.temb_proj(self.nonlinearity(temb))[:, :, None, None, None]
@@ -270,16 +295,13 @@ class CogVideoXResnetBlock3D(nn.Module):
 
         hidden_states = self.nonlinearity(hidden_states)
         hidden_states = self.dropout(hidden_states)
-        hidden_states = self.conv2(hidden_states, clear_fake_cp_cache=clear_fake_cp_cache)
+        hidden_states = self.conv2(hidden_states)
 
         if self.in_channels != self.out_channels:
-            if self.use_conv_shortcut:
-                inputs = self.conv_shortcut(inputs, clear_fake_cp_cache=clear_fake_cp_cache)
-            else:
-                inputs = self.conv_shortcut(inputs)
+            inputs = self.conv_shortcut(inputs)
 
-        output_tensor = inputs + hidden_states
-        return output_tensor
+        hidden_states = hidden_states + inputs
+        return hidden_states
 
 
 class CogVideoXDownBlock3D(nn.Module):
@@ -355,7 +377,6 @@ class CogVideoXDownBlock3D(nn.Module):
         hidden_states: torch.Tensor,
         temb: Optional[torch.Tensor] = None,
         zq: Optional[torch.Tensor] = None,
-        clear_fake_cp_cache: bool = False,
     ) -> torch.Tensor:
         for resnet in self.resnets:
             if self.training and self.gradient_checkpointing:
@@ -367,10 +388,10 @@ class CogVideoXDownBlock3D(nn.Module):
                     return create_forward
 
                 hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(resnet), hidden_states, temb, zq, clear_fake_cp_cache
+                    create_custom_forward(resnet), hidden_states, temb, zq
                 )
             else:
-                hidden_states = resnet(hidden_states, temb, zq, clear_fake_cp_cache)
+                hidden_states = resnet(hidden_states, temb, zq)
 
         if self.downsamplers is not None:
             for downsampler in self.downsamplers:
@@ -435,7 +456,6 @@ class CogVideoXMidBlock3D(nn.Module):
         hidden_states: torch.Tensor,
         temb: Optional[torch.Tensor] = None,
         zq: Optional[torch.Tensor] = None,
-        clear_fake_cp_cache: bool = False,
     ) -> torch.Tensor:
         for resnet in self.resnets:
             if self.training and self.gradient_checkpointing:
@@ -447,10 +467,10 @@ class CogVideoXMidBlock3D(nn.Module):
                     return create_forward
 
                 hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(resnet), hidden_states, temb, zq, clear_fake_cp_cache
+                    create_custom_forward(resnet), hidden_states, temb, zq
                 )
             else:
-                hidden_states = resnet(hidden_states, temb, zq, clear_fake_cp_cache)
+                hidden_states = resnet(hidden_states, temb, zq)
 
         return hidden_states
 
@@ -522,12 +542,13 @@ class CogVideoXUpBlock3D(nn.Module):
                 ]
             )
 
+        self.gradient_checkpointing = False
+
     def forward(
         self,
         hidden_states: torch.Tensor,
         temb: Optional[torch.Tensor] = None,
         zq: Optional[torch.Tensor] = None,
-        clear_fake_cp_cache: bool = False,
     ) -> torch.Tensor:
         r"""Forward method of the `CogVideoXUpBlock3D` class."""
         for resnet in self.resnets:
@@ -540,10 +561,10 @@ class CogVideoXUpBlock3D(nn.Module):
                     return create_forward
 
                 hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(resnet), hidden_states, temb, zq, clear_fake_cp_cache
+                    create_custom_forward(resnet), hidden_states, temb, zq
                 )
             else:
-                hidden_states = resnet(hidden_states, temb, zq, clear_fake_cp_cache)
+                hidden_states = resnet(hidden_states, temb, zq)
 
         if self.upsamplers is not None:
             for upsampler in self.upsamplers:
@@ -651,11 +672,9 @@ class CogVideoXEncoder3D(nn.Module):
 
         self.gradient_checkpointing = False
 
-    def forward(
-        self, sample: torch.Tensor, temb: Optional[torch.Tensor] = None, clear_fake_cp_cache: bool = True
-    ) -> torch.Tensor:
+    def forward(self, sample: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
         r"""The forward method of the `CogVideoXEncoder3D` class."""
-        hidden_states = self.conv_in(sample, clear_fake_cp_cache=clear_fake_cp_cache)
+        hidden_states = self.conv_in(sample)
 
         if self.training and self.gradient_checkpointing:
 
@@ -668,25 +687,25 @@ class CogVideoXEncoder3D(nn.Module):
             # 1. Down
             for down_block in self.down_blocks:
                 hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(down_block), hidden_states, temb, None, clear_fake_cp_cache
+                    create_custom_forward(down_block), hidden_states, temb, None
                 )
 
             # 2. Mid
             hidden_states = torch.utils.checkpoint.checkpoint(
-                create_custom_forward(self.mid_block), hidden_states, temb, None, clear_fake_cp_cache
+                create_custom_forward(self.mid_block), hidden_states, temb, None
             )
         else:
             # 1. Down
             for down_block in self.down_blocks:
-                hidden_states = down_block(hidden_states, temb, None, clear_fake_cp_cache)
+                hidden_states = down_block(hidden_states, temb, None)
 
             # 2. Mid
-            hidden_states = self.mid_block(hidden_states, temb, None, clear_fake_cp_cache)
+            hidden_states = self.mid_block(hidden_states, temb, None)
 
         # 3. Post-process
         hidden_states = self.norm_out(hidden_states)
         hidden_states = self.conv_act(hidden_states)
-        hidden_states = self.conv_out(hidden_states, clear_fake_cp_cache=clear_fake_cp_cache)
+        hidden_states = self.conv_out(hidden_states)
         return hidden_states
 
 
@@ -788,7 +807,7 @@ class CogVideoXDecoder3D(nn.Module):
 
             self.up_blocks.append(up_block)
 
-        self.norm_out = CogVideoXSpatialNorm3D(reversed_block_out_channels[-1], in_channels)
+        self.norm_out = CogVideoXSpatialNorm3D(reversed_block_out_channels[-1], in_channels, groups=norm_num_groups)
         self.conv_act = nn.SiLU()
         self.conv_out = CogVideoXCausalConv3d(
             reversed_block_out_channels[-1], out_channels, kernel_size=3, pad_mode=pad_mode
@@ -796,11 +815,9 @@ class CogVideoXDecoder3D(nn.Module):
 
         self.gradient_checkpointing = False
 
-    def forward(
-        self, sample: torch.Tensor, temb: Optional[torch.Tensor] = None, clear_fake_cp_cache: bool = True
-    ) -> torch.Tensor:
+    def forward(self, sample: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
         r"""The forward method of the `CogVideoXDecoder3D` class."""
-        hidden_states = self.conv_in(sample, clear_fake_cp_cache=clear_fake_cp_cache)
+        hidden_states = self.conv_in(sample)
 
         if self.training and self.gradient_checkpointing:
 
@@ -812,32 +829,33 @@ class CogVideoXDecoder3D(nn.Module):
 
             # 1. Mid
             hidden_states = torch.utils.checkpoint.checkpoint(
-                create_custom_forward(self.mid_block), hidden_states, temb, sample, clear_fake_cp_cache
+                create_custom_forward(self.mid_block), hidden_states, temb, sample
             )
 
             # 2. Up
             for up_block in self.up_blocks:
                 hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(up_block), hidden_states, temb, sample, clear_fake_cp_cache
+                    create_custom_forward(up_block), hidden_states, temb, sample
                 )
         else:
             # 1. Mid
-            hidden_states = self.mid_block(hidden_states, temb, sample, clear_fake_cp_cache)
+            hidden_states = self.mid_block(hidden_states, temb, sample)
 
             # 2. Up
             for up_block in self.up_blocks:
-                hidden_states = up_block(hidden_states, temb, sample, clear_fake_cp_cache)
+                hidden_states = up_block(hidden_states, temb, sample)
 
         # 3. Post-process
         hidden_states = self.norm_out(hidden_states, sample)
         hidden_states = self.conv_act(hidden_states)
-        hidden_states = self.conv_out(hidden_states, clear_fake_cp_cache=clear_fake_cp_cache)
+        hidden_states = self.conv_out(hidden_states)
         return hidden_states
 
 
 class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
     r"""
-    A VAE model with KL loss for encodfing images into latents and decoding latent representations into images.
+    A VAE model with KL loss for encoding images into latents and decoding latent representations into images. Used in
+    [CogVideoX](https://github.com/THUDM/CogVideo).
 
     This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
     for all models (such as downloading or saving).
@@ -864,9 +882,6 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
             can be fine-tuned / trained to a lower range without loosing too much precision in which case
             `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
-        mid_block_add_attention (`bool`, *optional*, default to `True`):
-            If enabled, the mid_block of the Encoder and Decoder will have attention blocks. If set to false, the
-            mid_block will only have resnet blocks
     """
 
     _supports_gradient_checkpointing = True
@@ -904,7 +919,6 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         force_upcast: float = True,
         use_quant_conv: bool = False,
         use_post_quant_conv: bool = False,
-        mid_block_add_attention: bool = True,
     ):
         super().__init__()
 
@@ -949,9 +963,15 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         if isinstance(module, (CogVideoXEncoder3D, CogVideoXDecoder3D)):
             module.gradient_checkpointing = value
 
+    def clear_fake_context_parallel_cache(self):
+        for name, module in self.named_modules():
+            if isinstance(module, CogVideoXCausalConv3d):
+                logger.debug(f"Clearing fake Context Parallel cache for layer: {name}")
+                module._clear_fake_context_parallel_cache()
+
     @apply_forward_hook
     def encode(
-        self, x: torch.Tensor, return_dict: bool = True, fake_cp: bool = False
+        self, x: torch.Tensor, return_dict: bool = True
     ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
         """
         Encode a batch of images into latents.
@@ -960,14 +980,12 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             x (`torch.Tensor`): Input batch of images.
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
-            fake_cp (`bool`, *optional*, defaults to `True`):
-                If True, the fake context parallel will be used to reduce GPU memory consumption (Only 1 GPU work).
 
         Returns:
                 The latent representations of the encoded images. If `return_dict` is True, a
                 [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
         """
-        h = self.encoder(x, clear_fake_cp_cache=not fake_cp)
+        h = self.encoder(x)
         if self.quant_conv is not None:
             h = self.quant_conv(h)
         posterior = DiagonalGaussianDistribution(h)
@@ -976,9 +994,7 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         return AutoencoderKLOutput(latent_dist=posterior)
 
     @apply_forward_hook
-    def decode(
-        self, z: torch.FloatTensor, return_dict: bool = True, fake_cp: bool = False
-    ) -> Union[DecoderOutput, torch.FloatTensor]:
+    def decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
         """
         Decode a batch of images.
 
@@ -986,8 +1002,6 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             z (`torch.Tensor`): Input batch of latent vectors.
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
-            fake_cp (`bool`, *optional*, defaults to `True`):
-                If True, the fake context parallel will be used to reduce GPU memory consumption (Only 1 GPU work).
 
         Returns:
             [`~models.vae.DecoderOutput`] or `tuple`:
@@ -997,7 +1011,7 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         """
         if self.post_quant_conv is not None:
             z = self.post_quant_conv(z)
-        dec = self.decoder(z, clear_fake_cp_cache=not fake_cp)
+        dec = self.decoder(z)
         if not return_dict:
             return (dec,)
         return DecoderOutput(sample=dec)

src/diffusers/models/controlnet_sparsectrl.py

@@ -20,6 +20,7 @@ from torch import nn
 from torch.nn import functional as F
 
 from ..configuration_utils import ConfigMixin, register_to_config
+from ..loaders import FromOriginalModelMixin
 from ..utils import BaseOutput, logging
 from .attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
@@ -92,7 +93,7 @@ class SparseControlNetConditioningEmbedding(nn.Module):
         return embedding
 
 
-class SparseControlNetModel(ModelMixin, ConfigMixin):
+class SparseControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
     """
     A SparseControlNet model as described in [SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion
     Models](https://arxiv.org/abs/2311.16933).
@@ -314,6 +315,7 @@ class SparseControlNetModel(ModelMixin, ConfigMixin):
                     temporal_num_attention_heads=motion_num_attention_heads[i],
                     temporal_max_seq_length=motion_max_seq_length,
                     temporal_transformer_layers_per_block=temporal_transformer_layers_per_block[i],
+                    temporal_double_self_attention=False,
                 )
             elif down_block_type == "DownBlockMotion":
                 down_block = DownBlockMotion(
@@ -331,6 +333,7 @@ class SparseControlNetModel(ModelMixin, ConfigMixin):
                     temporal_num_attention_heads=motion_num_attention_heads[i],
                     temporal_max_seq_length=motion_max_seq_length,
                     temporal_transformer_layers_per_block=temporal_transformer_layers_per_block[i],
+                    temporal_double_self_attention=False,
                 )
             else:
                 raise ValueError(

src/diffusers/models/controlnet_xs.py

@@ -285,7 +285,7 @@ class ControlNetXSAdapter(ModelMixin, ConfigMixin):
         upcast_attention (`bool`, defaults to `True`):
             Whether the attention computation should always be upcasted.
         max_norm_num_groups (`int`, defaults to 32):
-            Maximum number of groups in group normal. The actual number will the the largest divisor of the respective
+            Maximum number of groups in group normal. The actual number will be the largest divisor of the respective
             channels, that is <= max_norm_num_groups.
     """
 

src/diffusers/models/modeling_utils.py

@@ -773,7 +773,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                     try:
                         accelerate.load_checkpoint_and_dispatch(
                             model,
-                            model_file if not is_sharded else sharded_ckpt_cached_folder,
+                            model_file if not is_sharded else index_file,
                             device_map,
                             max_memory=max_memory,
                             offload_folder=offload_folder,
@@ -803,7 +803,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                             model._temp_convert_self_to_deprecated_attention_blocks()
                             accelerate.load_checkpoint_and_dispatch(
                                 model,
-                                model_file if not is_sharded else sharded_ckpt_cached_folder,
+                                model_file if not is_sharded else index_file,
                                 device_map,
                                 max_memory=max_memory,
                                 offload_folder=offload_folder,

src/diffusers/models/normalization.py

@@ -34,7 +34,11 @@ class AdaLayerNorm(nn.Module):
 
     Parameters:
         embedding_dim (`int`): The size of each embedding vector.
-        num_embeddings (`int`): The size of the embeddings dictionary.
+        num_embeddings (`int`, *optional*): The size of the embeddings dictionary.
+        output_dim (`int`, *optional*):
+        norm_elementwise_affine (`bool`, defaults to `False):
+        norm_eps (`bool`, defaults to `False`):
+        chunk_dim (`int`, defaults to `0`):
     """
 
     def __init__(
@@ -49,14 +53,13 @@ class AdaLayerNorm(nn.Module):
         super().__init__()
 
         self.chunk_dim = chunk_dim
+        output_dim = output_dim or embedding_dim * 2
 
         if num_embeddings is not None:
             self.emb = nn.Embedding(num_embeddings, embedding_dim)
         else:
             self.emb = None
 
-        output_dim = output_dim or embedding_dim * 2
-
         self.silu = nn.SiLU()
         self.linear = nn.Linear(embedding_dim, output_dim)
         self.norm = nn.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine)
@@ -68,7 +71,10 @@ class AdaLayerNorm(nn.Module):
             temb = self.emb(timestep)
 
         temb = self.linear(self.silu(temb))
+
         if self.chunk_dim == 1:
+            # This is a bit weird why we have the order of "shift, scale" here and "scale, shift" in the
+            # other if-branch. This branch is specific to CogVideoX for now.
             shift, scale = temb.chunk(2, dim=1)
             shift = shift[:, None, :]
             scale = scale[:, None, :]

src/diffusers/models/transformers/auraflow_transformer_2d.py

@@ -22,7 +22,12 @@ import torch.nn.functional as F
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import is_torch_version, logging
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..attention_processor import Attention, AuraFlowAttnProcessor2_0
+from ..attention_processor import (
+    Attention,
+    AttentionProcessor,
+    AuraFlowAttnProcessor2_0,
+    FusedAuraFlowAttnProcessor2_0,
+)
 from ..embeddings import TimestepEmbedding, Timesteps
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -320,6 +325,106 @@ class AuraFlowTransformer2DModel(ModelMixin, ConfigMixin):
 
         self.gradient_checkpointing = False
 
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedAuraFlowAttnProcessor2_0
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedAuraFlowAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
     def _set_gradient_checkpointing(self, module, value=False):
         if hasattr(module, "gradient_checkpointing"):
             module.gradient_checkpointing = value

src/diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -34,37 +34,30 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 @maybe_allow_in_graph
 class CogVideoXBlock(nn.Module):
     r"""
-    Transformer block used in CogVideoX model. TODO: add link to CogVideoX upon release
+    Transformer block used in [CogVideoX](https://github.com/THUDM/CogVideo) model.
 
     Parameters:
         dim (`int`): The number of channels in the input and output.
         num_attention_heads (`int`): The number of heads to use for multi-head attention.
         attention_head_dim (`int`): The number of channels in each head.
         dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
-        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
         activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
-        num_embeds_ada_norm (:
-            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
         attention_bias (:
             obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
-        only_cross_attention (`bool`, *optional*):
-            Whether to use only cross-attention layers. In this case two cross attention layers are used.
-        double_self_attention (`bool`, *optional*):
-            Whether to use two self-attention layers. In this case no cross attention layers are used.
-        upcast_attention (`bool`, *optional*):
-            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
-        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+        qk_norm (`bool`, defaults to `True`):
+            Whether or not to use normalization after query and key projections in Attention.
+        norm_elementwise_affine (`bool`, defaults to `True`):
             Whether to use learnable elementwise affine parameters for normalization.
-        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
-            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
-        final_dropout (`bool` *optional*, defaults to False):
+        norm_eps (`float`, defaults to `1e-5`):
+            Epsilon value for normalization layers.
+        final_dropout (`bool` defaults to `False`):
             Whether to apply a final dropout after the last feed-forward layer.
-        attention_type (`str`, *optional*, defaults to `"default"`):
-            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
-        positional_embeddings (`str`, *optional*, defaults to `None`):
-            The type of positional embeddings to apply to.
-        num_positional_embeddings (`int`, *optional*, defaults to `None`):
-            The maximum number of positional embeddings to apply.
+        ff_inner_dim (`int`, *optional*, defaults to `None`):
+            Custom hidden dimension of Feed-forward layer. If not provided, `4 * dim` is used.
+        ff_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in Feed-forward layer.
+        attention_out_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in Attention output projection layer.
     """
 
     def __init__(
@@ -151,7 +144,7 @@ class CogVideoXBlock(nn.Module):
 
 class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
     """
-    A Transformer model for video-like data in CogVideoX. TODO: add link to CogVideoX upon release
+    A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).
 
     Parameters:
         num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.

src/diffusers/models/transformers/pixart_transformer_2d.py

@@ -19,7 +19,7 @@ from torch import nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import is_torch_version, logging
 from ..attention import BasicTransformerBlock
-from ..attention_processor import AttentionProcessor
+from ..attention_processor import Attention, AttentionProcessor, FusedAttnProcessor2_0
 from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -247,6 +247,46 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
         for name, module in self.named_children():
             fn_recursive_attn_processor(name, module, processor)
 
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
     def forward(
         self,
         hidden_states: torch.Tensor,

src/diffusers/models/transformers/transformer_flux.py

@@ -20,7 +20,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...loaders import PeftAdapterMixin
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
 from ...models.attention import FeedForward
 from ...models.attention_processor import Attention, FluxAttnProcessor2_0, FluxSingleAttnProcessor2_0
 from ...models.modeling_utils import ModelMixin
@@ -125,6 +125,8 @@ class FluxSingleTransformerBlock(nn.Module):
         gate = gate.unsqueeze(1)
         hidden_states = gate * self.proj_out(hidden_states)
         hidden_states = residual + hidden_states
+        if hidden_states.dtype == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
 
         return hidden_states
 
@@ -223,11 +225,13 @@ class FluxTransformerBlock(nn.Module):
 
         context_ff_output = self.ff_context(norm_encoder_hidden_states)
         encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+        if encoder_hidden_states.dtype == torch.float16:
+            encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
 
         return encoder_hidden_states, hidden_states
 
 
-class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
     """
     The Transformer model introduced in Flux.
 

src/diffusers/models/unets/unet_motion_model.py

@@ -233,6 +233,7 @@ class DownBlockMotion(nn.Module):
         temporal_cross_attention_dim: Optional[int] = None,
         temporal_max_seq_length: int = 32,
         temporal_transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        temporal_double_self_attention: bool = True,
     ):
         super().__init__()
         resnets = []
@@ -282,6 +283,7 @@ class DownBlockMotion(nn.Module):
                     positional_embeddings="sinusoidal",
                     num_positional_embeddings=temporal_max_seq_length,
                     attention_head_dim=out_channels // temporal_num_attention_heads[i],
+                    double_self_attention=temporal_double_self_attention,
                 )
             )
 
@@ -343,6 +345,7 @@ class DownBlockMotion(nn.Module):
 
             else:
                 hidden_states = resnet(hidden_states, temb)
+
             hidden_states = motion_module(hidden_states, num_frames=num_frames)
 
             output_states = output_states + (hidden_states,)
@@ -384,6 +387,7 @@ class CrossAttnDownBlockMotion(nn.Module):
         temporal_num_attention_heads: int = 8,
         temporal_max_seq_length: int = 32,
         temporal_transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        temporal_double_self_attention: bool = True,
     ):
         super().__init__()
         resnets = []
@@ -465,6 +469,7 @@ class CrossAttnDownBlockMotion(nn.Module):
                     positional_embeddings="sinusoidal",
                     num_positional_embeddings=temporal_max_seq_length,
                     attention_head_dim=out_channels // temporal_num_attention_heads,
+                    double_self_attention=temporal_double_self_attention,
                 )
             )
 
@@ -536,6 +541,7 @@ class CrossAttnDownBlockMotion(nn.Module):
                 )[0]
             else:
                 hidden_states = resnet(hidden_states, temb)
+
                 hidden_states = attn(
                     hidden_states,
                     encoder_hidden_states=encoder_hidden_states,
@@ -761,6 +767,7 @@ class CrossAttnUpBlockMotion(nn.Module):
                 )[0]
             else:
                 hidden_states = resnet(hidden_states, temb)
+
                 hidden_states = attn(
                     hidden_states,
                     encoder_hidden_states=encoder_hidden_states,
@@ -921,9 +928,9 @@ class UpBlockMotion(nn.Module):
                     hidden_states = torch.utils.checkpoint.checkpoint(
                         create_custom_forward(resnet), hidden_states, temb
                     )
-
             else:
                 hidden_states = resnet(hidden_states, temb)
+
             hidden_states = motion_module(hidden_states, num_frames=num_frames)
 
         if self.upsamplers is not None:
@@ -1923,7 +1930,6 @@ class UNetMotionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Peft
         for name, module in self.named_children():
             fn_recursive_attn_processor(name, module, processor)
 
-    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
     def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
         """
         Sets the attention processor to use [feed forward
@@ -1953,7 +1959,6 @@ class UNetMotionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Peft
         for module in self.children():
             fn_recursive_feed_forward(module, chunk_size, dim)
 
-    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.disable_forward_chunking
     def disable_forward_chunking(self) -> None:
         def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
             if hasattr(module, "set_chunk_feed_forward"):

src/diffusers/pipelines/__init__.py

@@ -10,6 +10,7 @@ from ..utils import (
     is_librosa_available,
     is_note_seq_available,
     is_onnx_available,
+    is_sentencepiece_available,
     is_torch_available,
     is_torch_npu_available,
     is_transformers_available,
@@ -146,7 +147,9 @@ else:
     _import_structure["pag"].extend(
         [
             "AnimateDiffPAGPipeline",
+            "KolorsPAGPipeline",
             "HunyuanDiTPAGPipeline",
+            "StableDiffusion3PAGPipeline",
             "StableDiffusionPAGPipeline",
             "StableDiffusionControlNetPAGPipeline",
             "StableDiffusionXLPAGPipeline",
@@ -206,12 +209,6 @@ else:
         "Kandinsky3Img2ImgPipeline",
         "Kandinsky3Pipeline",
     ]
-    _import_structure["kolors"] = [
-        "KolorsPipeline",
-        "KolorsImg2ImgPipeline",
-        "ChatGLMModel",
-        "ChatGLMTokenizer",
-    ]
     _import_structure["latent_consistency_models"] = [
         "LatentConsistencyModelImg2ImgPipeline",
         "LatentConsistencyModelPipeline",
@@ -351,6 +348,22 @@ else:
         "StableDiffusionKDiffusionPipeline",
         "StableDiffusionXLKDiffusionPipeline",
     ]
+
+try:
+    if not (is_torch_available() and is_transformers_available() and is_sentencepiece_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ..utils import (
+        dummy_torch_and_transformers_and_sentencepiece_objects,
+    )
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_sentencepiece_objects))
+else:
+    _import_structure["kolors"] = [
+        "KolorsPipeline",
+        "KolorsImg2ImgPipeline",
+    ]
+
 try:
     if not is_flax_available():
         raise OptionalDependencyNotAvailable()
@@ -509,12 +522,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             Kandinsky3Img2ImgPipeline,
             Kandinsky3Pipeline,
         )
-        from .kolors import (
-            ChatGLMModel,
-            ChatGLMTokenizer,
-            KolorsImg2ImgPipeline,
-            KolorsPipeline,
-        )
         from .latent_consistency_models import (
             LatentConsistencyModelImg2ImgPipeline,
             LatentConsistencyModelPipeline,
@@ -536,7 +543,9 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .pag import (
             AnimateDiffPAGPipeline,
             HunyuanDiTPAGPipeline,
+            KolorsPAGPipeline,
             PixArtSigmaPAGPipeline,
+            StableDiffusion3PAGPipeline,
             StableDiffusionControlNetPAGPipeline,
             StableDiffusionPAGPipeline,
             StableDiffusionXLControlNetPAGPipeline,
@@ -644,6 +653,17 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                 StableDiffusionXLKDiffusionPipeline,
             )
 
+        try:
+            if not (is_torch_available() and is_transformers_available() and is_sentencepiece_available()):
+                raise OptionalDependencyNotAvailable()
+        except OptionalDependencyNotAvailable:
+            from ..utils.dummy_torch_and_transformers_and_sentencepiece_objects import *
+        else:
+            from .kolors import (
+                KolorsImg2ImgPipeline,
+                KolorsPipeline,
+            )
+
         try:
             if not is_flax_available():
                 raise OptionalDependencyNotAvailable()

src/diffusers/pipelines/animatediff/pipeline_animatediff.py

@@ -42,6 +42,7 @@ from ...utils import (
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
 from ..free_init_utils import FreeInitMixin
+from ..free_noise_utils import AnimateDiffFreeNoiseMixin
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from .pipeline_output import AnimateDiffPipelineOutput
 
@@ -72,6 +73,7 @@ class AnimateDiffPipeline(
     IPAdapterMixin,
     StableDiffusionLoraLoaderMixin,
     FreeInitMixin,
+    AnimateDiffFreeNoiseMixin,
 ):
     r"""
     Pipeline for text-to-video generation.
@@ -394,15 +396,20 @@ class AnimateDiffPipeline(
 
         return ip_adapter_image_embeds
 
-    # Copied from diffusers.pipelines.text_to_video_synthesis/pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
-    def decode_latents(self, latents):
+    def decode_latents(self, latents, decode_chunk_size: int = 16):
         latents = 1 / self.vae.config.scaling_factor * latents
 
         batch_size, channels, num_frames, height, width = latents.shape
         latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
 
-        image = self.vae.decode(latents).sample
-        video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4)
+        video = []
+        for i in range(0, latents.shape[0], decode_chunk_size):
+            batch_latents = latents[i : i + decode_chunk_size]
+            batch_latents = self.vae.decode(batch_latents).sample
+            video.append(batch_latents)
+
+        video = torch.cat(video)
+        video = video[None, :].reshape((batch_size, num_frames, -1) + video.shape[2:]).permute(0, 2, 1, 3, 4)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
         video = video.float()
         return video
@@ -495,10 +502,21 @@ class AnimateDiffPipeline(
                     f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
                 )
 
-    # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
     def prepare_latents(
         self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
     ):
+        # If FreeNoise is enabled, generate latents as described in Equation (7) of [FreeNoise](https://arxiv.org/abs/2310.15169)
+        if self.free_noise_enabled:
+            latents = self._prepare_latents_free_noise(
+                batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents
+            )
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
         shape = (
             batch_size,
             num_channels_latents,
@@ -506,11 +524,6 @@ class AnimateDiffPipeline(
             height // self.vae_scale_factor,
             width // self.vae_scale_factor,
         )
-        if isinstance(generator, list) and len(generator) != batch_size:
-            raise ValueError(
-                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
-                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
-            )
 
         if latents is None:
             latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
@@ -569,6 +582,7 @@ class AnimateDiffPipeline(
         clip_skip: Optional[int] = None,
         callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        decode_chunk_size: int = 16,
         **kwargs,
     ):
         r"""
@@ -637,6 +651,8 @@ class AnimateDiffPipeline(
                 The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                 will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                 `._callback_tensor_inputs` attribute of your pipeline class.
+            decode_chunk_size (`int`, defaults to `16`):
+                The number of frames to decode at a time when calling `decode_latents` method.
 
         Examples:
 
@@ -808,7 +824,7 @@ class AnimateDiffPipeline(
         if output_type == "latent":
             video = latents
         else:
-            video_tensor = self.decode_latents(latents)
+            video_tensor = self.decode_latents(latents, decode_chunk_size)
             video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type)
 
         # 10. Offload all models

src/diffusers/pipelines/animatediff/pipeline_animatediff_controlnet.py

@@ -30,6 +30,7 @@ from ...utils.torch_utils import is_compiled_module, randn_tensor
 from ...video_processor import VideoProcessor
 from ..controlnet.multicontrolnet import MultiControlNetModel
 from ..free_init_utils import FreeInitMixin
+from ..free_noise_utils import AnimateDiffFreeNoiseMixin
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from .pipeline_output import AnimateDiffPipelineOutput
 
@@ -109,6 +110,7 @@ class AnimateDiffControlNetPipeline(
     IPAdapterMixin,
     StableDiffusionLoraLoaderMixin,
     FreeInitMixin,
+    AnimateDiffFreeNoiseMixin,
 ):
     r"""
     Pipeline for text-to-video generation with ControlNet guidance.
@@ -432,15 +434,16 @@ class AnimateDiffControlNetPipeline(
 
         return ip_adapter_image_embeds
 
-    def decode_latents(self, latents, decode_batch_size: int = 16):
+    # Copied from diffusers.pipelines.animatediff.pipeline_animatediff.AnimateDiffPipeline.decode_latents
+    def decode_latents(self, latents, decode_chunk_size: int = 16):
         latents = 1 / self.vae.config.scaling_factor * latents
 
         batch_size, channels, num_frames, height, width = latents.shape
         latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
 
         video = []
-        for i in range(0, latents.shape[0], decode_batch_size):
-            batch_latents = latents[i : i + decode_batch_size]
+        for i in range(0, latents.shape[0], decode_chunk_size):
+            batch_latents = latents[i : i + decode_chunk_size]
             batch_latents = self.vae.decode(batch_latents).sample
             video.append(batch_latents)
 
@@ -608,10 +611,22 @@ class AnimateDiffControlNetPipeline(
             if end > 1.0:
                 raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
 
-    # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
+    # Copied from diffusers.pipelines.animatediff.pipeline_animatediff.AnimateDiffPipeline.prepare_latents
     def prepare_latents(
         self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
     ):
+        # If FreeNoise is enabled, generate latents as described in Equation (7) of [FreeNoise](https://arxiv.org/abs/2310.15169)
+        if self.free_noise_enabled:
+            latents = self._prepare_latents_free_noise(
+                batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents
+            )
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
         shape = (
             batch_size,
             num_channels_latents,
@@ -619,11 +634,6 @@ class AnimateDiffControlNetPipeline(
             height // self.vae_scale_factor,
             width // self.vae_scale_factor,
         )
-        if isinstance(generator, list) and len(generator) != batch_size:
-            raise ValueError(
-                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
-                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
-            )
 
         if latents is None:
             latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
@@ -718,7 +728,7 @@ class AnimateDiffControlNetPipeline(
         clip_skip: Optional[int] = None,
         callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
-        decode_batch_size: int = 16,
+        decode_chunk_size: int = 16,
     ):
         r"""
         The call function to the pipeline for generation.
@@ -1054,7 +1064,7 @@ class AnimateDiffControlNetPipeline(
         if output_type == "latent":
             video = latents
         else:
-            video_tensor = self.decode_latents(latents, decode_batch_size)
+            video_tensor = self.decode_latents(latents, decode_chunk_size)
             video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type)
 
         # 10. Offload all models

src/diffusers/pipelines/animatediff/pipeline_animatediff_video2video.py

@@ -35,6 +35,7 @@ from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
 from ..free_init_utils import FreeInitMixin
+from ..free_noise_utils import AnimateDiffFreeNoiseMixin
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from .pipeline_output import AnimateDiffPipelineOutput
 
@@ -176,6 +177,7 @@ class AnimateDiffVideoToVideoPipeline(
     IPAdapterMixin,
     StableDiffusionLoraLoaderMixin,
     FreeInitMixin,
+    AnimateDiffFreeNoiseMixin,
 ):
     r"""
     Pipeline for video-to-video generation.
@@ -498,15 +500,29 @@ class AnimateDiffVideoToVideoPipeline(
 
         return ip_adapter_image_embeds
 
-    # Copied from diffusers.pipelines.text_to_video_synthesis/pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
-    def decode_latents(self, latents):
+    def encode_video(self, video, generator, decode_chunk_size: int = 16) -> torch.Tensor:
+        latents = []
+        for i in range(0, len(video), decode_chunk_size):
+            batch_video = video[i : i + decode_chunk_size]
+            batch_video = retrieve_latents(self.vae.encode(batch_video), generator=generator)
+            latents.append(batch_video)
+        return torch.cat(latents)
+
+    # Copied from diffusers.pipelines.animatediff.pipeline_animatediff.AnimateDiffPipeline.decode_latents
+    def decode_latents(self, latents, decode_chunk_size: int = 16):
         latents = 1 / self.vae.config.scaling_factor * latents
 
         batch_size, channels, num_frames, height, width = latents.shape
         latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
 
-        image = self.vae.decode(latents).sample
-        video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4)
+        video = []
+        for i in range(0, latents.shape[0], decode_chunk_size):
+            batch_latents = latents[i : i + decode_chunk_size]
+            batch_latents = self.vae.decode(batch_latents).sample
+            video.append(batch_latents)
+
+        video = torch.cat(video)
+        video = video[None, :].reshape((batch_size, num_frames, -1) + video.shape[2:]).permute(0, 2, 1, 3, 4)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
         video = video.float()
         return video
@@ -622,6 +638,7 @@ class AnimateDiffVideoToVideoPipeline(
         device,
         generator,
         latents=None,
+        decode_chunk_size: int = 16,
     ):
         if latents is None:
             num_frames = video.shape[1]
@@ -656,13 +673,11 @@ class AnimateDiffVideoToVideoPipeline(
                     )
 
                 init_latents = [
-                    retrieve_latents(self.vae.encode(video[i]), generator=generator[i]).unsqueeze(0)
+                    self.encode_video(video[i], generator[i], decode_chunk_size).unsqueeze(0)
                     for i in range(batch_size)
                 ]
             else:
-                init_latents = [
-                    retrieve_latents(self.vae.encode(vid), generator=generator).unsqueeze(0) for vid in video
-                ]
+                init_latents = [self.encode_video(vid, generator, decode_chunk_size).unsqueeze(0) for vid in video]
 
             init_latents = torch.cat(init_latents, dim=0)
 
@@ -747,6 +762,7 @@ class AnimateDiffVideoToVideoPipeline(
         clip_skip: Optional[int] = None,
         callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        decode_chunk_size: int = 16,
     ):
         r"""
         The call function to the pipeline for generation.
@@ -822,6 +838,8 @@ class AnimateDiffVideoToVideoPipeline(
                 The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                 will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                 `._callback_tensor_inputs` attribute of your pipeline class.
+            decode_chunk_size (`int`, defaults to `16`):
+                The number of frames to decode at a time when calling `decode_latents` method.
 
         Examples:
 
@@ -923,6 +941,7 @@ class AnimateDiffVideoToVideoPipeline(
             device=device,
             generator=generator,
             latents=latents,
+            decode_chunk_size=decode_chunk_size,
         )
 
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
@@ -990,7 +1009,7 @@ class AnimateDiffVideoToVideoPipeline(
         if output_type == "latent":
             video = latents
         else:
-            video_tensor = self.decode_latents(latents)
+            video_tensor = self.decode_latents(latents, decode_chunk_size)
             video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type)
 
         # 10. Offload all models

src/diffusers/pipelines/auto_pipeline.py

@@ -18,6 +18,7 @@ from collections import OrderedDict
 from huggingface_hub.utils import validate_hf_hub_args
 
 from ..configuration_utils import ConfigMixin
+from ..utils import is_sentencepiece_available
 from .aura_flow import AuraFlowPipeline
 from .controlnet import (
     StableDiffusionControlNetImg2ImgPipeline,
@@ -47,11 +48,11 @@ from .kandinsky2_2 import (
     KandinskyV22Pipeline,
 )
 from .kandinsky3 import Kandinsky3Img2ImgPipeline, Kandinsky3Pipeline
-from .kolors import KolorsImg2ImgPipeline, KolorsPipeline
 from .latent_consistency_models import LatentConsistencyModelImg2ImgPipeline, LatentConsistencyModelPipeline
 from .pag import (
     HunyuanDiTPAGPipeline,
     PixArtSigmaPAGPipeline,
+    StableDiffusion3PAGPipeline,
     StableDiffusionControlNetPAGPipeline,
     StableDiffusionPAGPipeline,
     StableDiffusionXLControlNetPAGPipeline,
@@ -84,6 +85,7 @@ AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("stable-diffusion", StableDiffusionPipeline),
         ("stable-diffusion-xl", StableDiffusionXLPipeline),
         ("stable-diffusion-3", StableDiffusion3Pipeline),
+        ("stable-diffusion-3-pag", StableDiffusion3PAGPipeline),
         ("if", IFPipeline),
         ("hunyuan", HunyuanDiTPipeline),
         ("hunyuan-pag", HunyuanDiTPAGPipeline),
@@ -103,7 +105,6 @@ AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGPipeline),
         ("pixart-sigma-pag", PixArtSigmaPAGPipeline),
         ("auraflow", AuraFlowPipeline),
-        ("kolors", KolorsPipeline),
         ("flux", FluxPipeline),
     ]
 )
@@ -121,7 +122,6 @@ AUTO_IMAGE2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetImg2ImgPipeline),
         ("stable-diffusion-xl-pag", StableDiffusionXLPAGImg2ImgPipeline),
         ("lcm", LatentConsistencyModelImg2ImgPipeline),
-        ("kolors", KolorsImg2ImgPipeline),
     ]
 )
 
@@ -160,6 +160,14 @@ _AUTO_INPAINT_DECODER_PIPELINES_MAPPING = OrderedDict(
     ]
 )
 
+if is_sentencepiece_available():
+    from .kolors import KolorsPipeline
+    from .pag import KolorsPAGPipeline
+
+    AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsPipeline
+    AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors-pag"] = KolorsPAGPipeline
+    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsPipeline
+
 SUPPORTED_TASKS_MAPPINGS = [
     AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
     AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,

src/diffusers/pipelines/cogvideo/pipeline_cogvideox.py

@@ -36,10 +36,11 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 EXAMPLE_DOC_STRING = """
     Examples:
         ```python
+        >>> import torch
         >>> from diffusers import CogVideoXPipeline
         >>> from diffusers.utils import export_to_video
 
-        >>> pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.bfloat16).to("cuda")
+        >>> pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16).to("cuda")
         >>> prompt = (
         ...     "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
         ...     "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
@@ -48,9 +49,7 @@ EXAMPLE_DOC_STRING = """
         ...     "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
         ...     "atmosphere of this unique musical performance."
         ... )
-        >>> video = pipe(
-        ...     "a polar bear dancing, high quality, realistic", guidance_scale=6, num_inference_steps=20
-        ... ).frames[0]
+        >>> video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
         >>> export_to_video(video, "output.mp4", fps=8)
         ```
 """
@@ -154,7 +153,7 @@ class CogVideoXPipeline(DiffusionPipeline):
             A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
     """
 
-    _optional_components = ["tokenizer", "text_encoder"]
+    _optional_components = []
     model_cpu_offload_seq = "text_encoder->transformer->vae"
 
     _callback_tensor_inputs = [
@@ -182,9 +181,6 @@ class CogVideoXPipeline(DiffusionPipeline):
         self.vae_scale_factor_temporal = (
             self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
         )
-        self.tokenizer_max_length = (
-            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 226
-        )
 
         self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
 
@@ -214,7 +210,7 @@ class CogVideoXPipeline(DiffusionPipeline):
         untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
 
         if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
-            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
             logger.warning(
                 "The following part of your input was truncated because `max_sequence_length` is set to "
                 f" {max_sequence_length} tokens: {removed_text}"
@@ -342,13 +338,13 @@ class CogVideoXPipeline(DiffusionPipeline):
 
         frames = []
         for i in range(num_seconds):
-            # Whether or not to clear fake context parallel cache
-            fake_cp = i + 1 < num_seconds
             start_frame, end_frame = (0, 3) if i == 0 else (2 * i + 1, 2 * i + 3)
 
-            current_frames = self.vae.decode(latents[:, :, start_frame:end_frame], fake_cp=fake_cp).sample
+            current_frames = self.vae.decode(latents[:, :, start_frame:end_frame]).sample
             frames.append(current_frames)
 
+        self.vae.clear_fake_context_parallel_cache()
+
         frames = torch.cat(frames, dim=2)
         return frames
 
@@ -447,6 +443,7 @@ class CogVideoXPipeline(DiffusionPipeline):
         num_inference_steps: int = 50,
         timesteps: Optional[List[int]] = None,
         guidance_scale: float = 6,
+        use_dynamic_cfg: bool = False,
         num_videos_per_prompt: int = 1,
         eta: float = 0.0,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
@@ -459,7 +456,7 @@ class CogVideoXPipeline(DiffusionPipeline):
             Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
         ] = None,
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
-        use_dynamic_cfg: bool = False,
+        max_sequence_length: int = 226,
     ) -> Union[CogVideoXPipelineOutput, Tuple]:
         """
         Function invoked when calling the pipeline for generation.
@@ -525,6 +522,9 @@ class CogVideoXPipeline(DiffusionPipeline):
                 The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                 will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                 `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, defaults to `226`):
+                Maximum sequence length in encoded prompt. Must be consistent with
+                `self.transformer.config.max_text_seq_length` otherwise may lead to poor results.
 
         Examples:
 
@@ -581,6 +581,7 @@ class CogVideoXPipeline(DiffusionPipeline):
             num_videos_per_prompt=num_videos_per_prompt,
             prompt_embeds=prompt_embeds,
             negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
             device=device,
         )
         if do_classifier_free_guidance:
@@ -671,7 +672,7 @@ class CogVideoXPipeline(DiffusionPipeline):
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
 
-        if not output_type == "latents":
+        if not output_type == "latent":
             video = self.decode_latents(latents, num_frames // fps)
             video = self.video_processor.postprocess_video(video=video, output_type=output_type)
         else:

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py

@@ -76,13 +76,13 @@ EXAMPLE_DOC_STRING = """
         >>> import numpy as np
         >>> from PIL import Image
 
-        >>> from transformers import DPTFeatureExtractor, DPTForDepthEstimation
+        >>> from transformers import DPTImageProcessor, DPTForDepthEstimation
         >>> from diffusers import ControlNetModel, StableDiffusionXLControlNetImg2ImgPipeline, AutoencoderKL
         >>> from diffusers.utils import load_image
 
 
         >>> depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to("cuda")
-        >>> feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-hybrid-midas")
+        >>> feature_extractor = DPTImageProcessor.from_pretrained("Intel/dpt-hybrid-midas")
         >>> controlnet = ControlNetModel.from_pretrained(
         ...     "diffusers/controlnet-depth-sdxl-1.0-small",
         ...     variant="fp16",

src/diffusers/pipelines/controlnet/pipeline_flax_controlnet.py

@@ -23,7 +23,7 @@ from flax.core.frozen_dict import FrozenDict
 from flax.jax_utils import unreplicate
 from flax.training.common_utils import shard
 from PIL import Image
-from transformers import CLIPFeatureExtractor, CLIPTokenizer, FlaxCLIPTextModel
+from transformers import CLIPImageProcessor, CLIPTokenizer, FlaxCLIPTextModel
 
 from ...models import FlaxAutoencoderKL, FlaxControlNetModel, FlaxUNet2DConditionModel
 from ...schedulers import (
@@ -149,7 +149,7 @@ class FlaxStableDiffusionControlNetPipeline(FlaxDiffusionPipeline):
             FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler
         ],
         safety_checker: FlaxStableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         dtype: jnp.dtype = jnp.float32,
     ):
         super().__init__()

src/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py

@@ -16,7 +16,7 @@ import inspect
 from typing import Any, Callable, Dict, List, Optional, Union
 
 import torch
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from ....image_processor import VaeImageProcessor
 from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
@@ -66,8 +66,8 @@ class StableDiffusionModelEditingPipeline(
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
             about a model's potential harms.
-        feature_extractor ([`~transformers.CLIPFeatureExtractor`]):
-            A `CLIPFeatureExtractor` to extract features from generated images; used as inputs to the `safety_checker`.
+        feature_extractor ([`~transformers.CLIPImageProcessor`]):
+            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
         with_to_k ([`bool`]):
             Whether to edit the key projection matrices along with the value projection matrices.
         with_augs ([`list`]):
@@ -86,7 +86,7 @@ class StableDiffusionModelEditingPipeline(
         unet: UNet2DConditionModel,
         scheduler: SchedulerMixin,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         requires_safety_checker: bool = True,
         with_to_k: bool = True,
         with_augs: list = AUGS_CONST,

src/diffusers/pipelines/free_noise_utils.py

@@ -0,0 +1,236 @@
+# 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 Optional, Union
+
+import torch
+
+from ..models.attention import BasicTransformerBlock, FreeNoiseTransformerBlock
+from ..models.unets.unet_motion_model import (
+    CrossAttnDownBlockMotion,
+    DownBlockMotion,
+    UpBlockMotion,
+)
+from ..utils import logging
+from ..utils.torch_utils import randn_tensor
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class AnimateDiffFreeNoiseMixin:
+    r"""Mixin class for [FreeNoise](https://arxiv.org/abs/2310.15169)."""
+
+    def _enable_free_noise_in_block(self, block: Union[CrossAttnDownBlockMotion, DownBlockMotion, UpBlockMotion]):
+        r"""Helper function to enable FreeNoise in transformer blocks."""
+
+        for motion_module in block.motion_modules:
+            num_transformer_blocks = len(motion_module.transformer_blocks)
+
+            for i in range(num_transformer_blocks):
+                if isinstance(motion_module.transformer_blocks[i], FreeNoiseTransformerBlock):
+                    motion_module.transformer_blocks[i].set_free_noise_properties(
+                        self._free_noise_context_length,
+                        self._free_noise_context_stride,
+                        self._free_noise_weighting_scheme,
+                    )
+                else:
+                    assert isinstance(motion_module.transformer_blocks[i], BasicTransformerBlock)
+                    basic_transfomer_block = motion_module.transformer_blocks[i]
+
+                    motion_module.transformer_blocks[i] = FreeNoiseTransformerBlock(
+                        dim=basic_transfomer_block.dim,
+                        num_attention_heads=basic_transfomer_block.num_attention_heads,
+                        attention_head_dim=basic_transfomer_block.attention_head_dim,
+                        dropout=basic_transfomer_block.dropout,
+                        cross_attention_dim=basic_transfomer_block.cross_attention_dim,
+                        activation_fn=basic_transfomer_block.activation_fn,
+                        attention_bias=basic_transfomer_block.attention_bias,
+                        only_cross_attention=basic_transfomer_block.only_cross_attention,
+                        double_self_attention=basic_transfomer_block.double_self_attention,
+                        positional_embeddings=basic_transfomer_block.positional_embeddings,
+                        num_positional_embeddings=basic_transfomer_block.num_positional_embeddings,
+                        context_length=self._free_noise_context_length,
+                        context_stride=self._free_noise_context_stride,
+                        weighting_scheme=self._free_noise_weighting_scheme,
+                    ).to(device=self.device, dtype=self.dtype)
+
+                    motion_module.transformer_blocks[i].load_state_dict(
+                        basic_transfomer_block.state_dict(), strict=True
+                    )
+
+    def _disable_free_noise_in_block(self, block: Union[CrossAttnDownBlockMotion, DownBlockMotion, UpBlockMotion]):
+        r"""Helper function to disable FreeNoise in transformer blocks."""
+
+        for motion_module in block.motion_modules:
+            num_transformer_blocks = len(motion_module.transformer_blocks)
+
+            for i in range(num_transformer_blocks):
+                if isinstance(motion_module.transformer_blocks[i], FreeNoiseTransformerBlock):
+                    free_noise_transfomer_block = motion_module.transformer_blocks[i]
+
+                    motion_module.transformer_blocks[i] = BasicTransformerBlock(
+                        dim=free_noise_transfomer_block.dim,
+                        num_attention_heads=free_noise_transfomer_block.num_attention_heads,
+                        attention_head_dim=free_noise_transfomer_block.attention_head_dim,
+                        dropout=free_noise_transfomer_block.dropout,
+                        cross_attention_dim=free_noise_transfomer_block.cross_attention_dim,
+                        activation_fn=free_noise_transfomer_block.activation_fn,
+                        attention_bias=free_noise_transfomer_block.attention_bias,
+                        only_cross_attention=free_noise_transfomer_block.only_cross_attention,
+                        double_self_attention=free_noise_transfomer_block.double_self_attention,
+                        positional_embeddings=free_noise_transfomer_block.positional_embeddings,
+                        num_positional_embeddings=free_noise_transfomer_block.num_positional_embeddings,
+                    ).to(device=self.device, dtype=self.dtype)
+
+                    motion_module.transformer_blocks[i].load_state_dict(
+                        free_noise_transfomer_block.state_dict(), strict=True
+                    )
+
+    def _prepare_latents_free_noise(
+        self,
+        batch_size: int,
+        num_channels_latents: int,
+        num_frames: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.Tensor] = None,
+    ):
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        context_num_frames = (
+            self._free_noise_context_length if self._free_noise_context_length == "repeat_context" else num_frames
+        )
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            context_num_frames,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+            if self._free_noise_noise_type == "random":
+                return latents
+        else:
+            if latents.size(2) == num_frames:
+                return latents
+            elif latents.size(2) != self._free_noise_context_length:
+                raise ValueError(
+                    f"You have passed `latents` as a parameter to FreeNoise. The expected number of frames is either {num_frames} or {self._free_noise_context_length}, but found {latents.size(2)}"
+                )
+            latents = latents.to(device)
+
+        if self._free_noise_noise_type == "shuffle_context":
+            for i in range(self._free_noise_context_length, num_frames, self._free_noise_context_stride):
+                # ensure window is within bounds
+                window_start = max(0, i - self._free_noise_context_length)
+                window_end = min(num_frames, window_start + self._free_noise_context_stride)
+                window_length = window_end - window_start
+
+                if window_length == 0:
+                    break
+
+                indices = torch.LongTensor(list(range(window_start, window_end)))
+                shuffled_indices = indices[torch.randperm(window_length, generator=generator)]
+
+                current_start = i
+                current_end = min(num_frames, current_start + window_length)
+                if current_end == current_start + window_length:
+                    # batch of frames perfectly fits the window
+                    latents[:, :, current_start:current_end] = latents[:, :, shuffled_indices]
+                else:
+                    # handle the case where the last batch of frames does not fit perfectly with the window
+                    prefix_length = current_end - current_start
+                    shuffled_indices = shuffled_indices[:prefix_length]
+                    latents[:, :, current_start:current_end] = latents[:, :, shuffled_indices]
+
+        elif self._free_noise_noise_type == "repeat_context":
+            num_repeats = (num_frames + self._free_noise_context_length - 1) // self._free_noise_context_length
+            latents = torch.cat([latents] * num_repeats, dim=2)
+
+        latents = latents[:, :, :num_frames]
+        return latents
+
+    def enable_free_noise(
+        self,
+        context_length: Optional[int] = 16,
+        context_stride: int = 4,
+        weighting_scheme: str = "pyramid",
+        noise_type: str = "shuffle_context",
+    ) -> None:
+        r"""
+        Enable long video generation using FreeNoise.
+
+        Args:
+            context_length (`int`, defaults to `16`, *optional*):
+                The number of video frames to process at once. It's recommended to set this to the maximum frames the
+                Motion Adapter was trained with (usually 16/24/32). If `None`, the default value from the motion
+                adapter config is used.
+            context_stride (`int`, *optional*):
+                Long videos are generated by processing many frames. FreeNoise processes these frames in sliding
+                windows of size `context_length`. Context stride allows you to specify how many frames to skip between
+                each window. For example, a context length of 16 and context stride of 4 would process 24 frames as:
+                    [0, 15], [4, 19], [8, 23] (0-based indexing)
+            weighting_scheme (`str`, defaults to `pyramid`):
+                Weighting scheme for averaging latents after accumulation in FreeNoise blocks. The following weighting
+                schemes are supported currently:
+                    - "pyramid"
+                        Peforms weighted averaging with a pyramid like weight pattern: [1, 2, 3, 2, 1].
+            noise_type (`str`, defaults to "shuffle_context"):
+                TODO
+        """
+
+        allowed_weighting_scheme = ["pyramid"]
+        allowed_noise_type = ["shuffle_context", "repeat_context", "random"]
+
+        if context_length > self.motion_adapter.config.motion_max_seq_length:
+            logger.warning(
+                f"You have set {context_length=} which is greater than {self.motion_adapter.config.motion_max_seq_length=}. This can lead to bad generation results."
+            )
+        if weighting_scheme not in allowed_weighting_scheme:
+            raise ValueError(
+                f"The parameter `weighting_scheme` must be one of {allowed_weighting_scheme}, but got {weighting_scheme=}"
+            )
+        if noise_type not in allowed_noise_type:
+            raise ValueError(f"The parameter `noise_type` must be one of {allowed_noise_type}, but got {noise_type=}")
+
+        self._free_noise_context_length = context_length or self.motion_adapter.config.motion_max_seq_length
+        self._free_noise_context_stride = context_stride
+        self._free_noise_weighting_scheme = weighting_scheme
+        self._free_noise_noise_type = noise_type
+
+        blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks]
+        for block in blocks:
+            self._enable_free_noise_in_block(block)
+
+    def disable_free_noise(self) -> None:
+        self._free_noise_context_length = None
+
+        blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks]
+        for block in blocks:
+            self._disable_free_noise_in_block(block)
+
+    @property
+    def free_noise_enabled(self):
+        return hasattr(self, "_free_noise_context_length") and self._free_noise_context_length is not None

src/diffusers/pipelines/kolors/__init__.py

@@ -5,6 +5,7 @@ from ...utils import (
     OptionalDependencyNotAvailable,
     _LazyModule,
     get_objects_from_module,
+    is_sentencepiece_available,
     is_torch_available,
     is_transformers_available,
 )
@@ -14,12 +15,12 @@ _dummy_objects = {}
 _import_structure = {}
 
 try:
-    if not (is_transformers_available() and is_torch_available()):
+    if not (is_transformers_available() and is_torch_available()) and is_sentencepiece_available():
         raise OptionalDependencyNotAvailable()
 except OptionalDependencyNotAvailable:
-    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+    from ...utils import dummy_torch_and_transformers_and_sentencepiece_objects  # noqa F403
 
-    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_sentencepiece_objects))
 else:
     _import_structure["pipeline_kolors"] = ["KolorsPipeline"]
     _import_structure["pipeline_kolors_img2img"] = ["KolorsImg2ImgPipeline"]
@@ -28,10 +29,10 @@ else:
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
-        if not (is_transformers_available() and is_torch_available()):
+        if not (is_transformers_available() and is_torch_available()) and is_sentencepiece_available():
             raise OptionalDependencyNotAvailable()
     except OptionalDependencyNotAvailable:
-        from ...utils.dummy_torch_and_transformers_objects import *
+        from ...utils.dummy_torch_and_transformers_and_sentencepiece_objects import *
 
     else:
         from .pipeline_kolors import KolorsPipeline

src/diffusers/pipelines/kolors/tokenizer.py

@@ -143,10 +143,18 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
     def unk_token(self) -> str:
         return "<unk>"
 
+    @unk_token.setter
+    def unk_token(self, value: str):
+        self._unk_token = value
+
     @property
     def pad_token(self) -> str:
         return "<unk>"
 
+    @pad_token.setter
+    def pad_token(self, value: str):
+        self._pad_token = value
+
     @property
     def pad_token_id(self):
         return self.get_command("<pad>")
@@ -155,6 +163,10 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
     def eos_token(self) -> str:
         return "</s>"
 
+    @eos_token.setter
+    def eos_token(self, value: str):
+        self._eos_token = value
+
     @property
     def eos_token_id(self):
         return self.get_command("<eos>")

src/diffusers/pipelines/pag/__init__.py

@@ -25,8 +25,10 @@ else:
     _import_structure["pipeline_pag_controlnet_sd"] = ["StableDiffusionControlNetPAGPipeline"]
     _import_structure["pipeline_pag_controlnet_sd_xl"] = ["StableDiffusionXLControlNetPAGPipeline"]
     _import_structure["pipeline_pag_hunyuandit"] = ["HunyuanDiTPAGPipeline"]
+    _import_structure["pipeline_pag_kolors"] = ["KolorsPAGPipeline"]
     _import_structure["pipeline_pag_pixart_sigma"] = ["PixArtSigmaPAGPipeline"]
     _import_structure["pipeline_pag_sd"] = ["StableDiffusionPAGPipeline"]
+    _import_structure["pipeline_pag_sd_3"] = ["StableDiffusion3PAGPipeline"]
     _import_structure["pipeline_pag_sd_animatediff"] = ["AnimateDiffPAGPipeline"]
     _import_structure["pipeline_pag_sd_xl"] = ["StableDiffusionXLPAGPipeline"]
     _import_structure["pipeline_pag_sd_xl_img2img"] = ["StableDiffusionXLPAGImg2ImgPipeline"]
@@ -43,8 +45,10 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .pipeline_pag_controlnet_sd import StableDiffusionControlNetPAGPipeline
         from .pipeline_pag_controlnet_sd_xl import StableDiffusionXLControlNetPAGPipeline
         from .pipeline_pag_hunyuandit import HunyuanDiTPAGPipeline
+        from .pipeline_pag_kolors import KolorsPAGPipeline
         from .pipeline_pag_pixart_sigma import PixArtSigmaPAGPipeline
         from .pipeline_pag_sd import StableDiffusionPAGPipeline
+        from .pipeline_pag_sd_3 import StableDiffusion3PAGPipeline
         from .pipeline_pag_sd_animatediff import AnimateDiffPAGPipeline
         from .pipeline_pag_sd_xl import StableDiffusionXLPAGPipeline
         from .pipeline_pag_sd_xl_img2img import StableDiffusionXLPAGImg2ImgPipeline

src/diffusers/pipelines/pag/pipeline_pag_sd_3.py

@@ -0,0 +1,985 @@
+# Copyright 2024 Stability AI and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import (
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    T5EncoderModel,
+    T5TokenizerFast,
+)
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import FromSingleFileMixin, SD3LoraLoaderMixin
+from ...models.attention_processor import PAGCFGJointAttnProcessor2_0, PAGJointAttnProcessor2_0
+from ...models.autoencoders import AutoencoderKL
+from ...models.transformers import SD3Transformer2DModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import (
+    USE_PEFT_BACKEND,
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from ..stable_diffusion_3.pipeline_output import StableDiffusion3PipelineOutput
+from .pag_utils import PAGMixin
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import AutoPipelineForText2Image
+
+        >>> pipe = AutoPipelineForText2Image.from_pretrained(
+        ...     "stabilityai/stable-diffusion-3-medium-diffusers",
+        ...     torch_dtype=torch.float16,
+        ...     enable_pag=True,
+        ...     pag_applied_layers=["blocks.13"],
+        ... )
+        >>> pipe.to("cuda")
+        >>> prompt = "A cat holding a sign that says hello world"
+        >>> image = pipe(prompt, guidance_scale=5.0, pag_scale=0.7).images[0]
+        >>> image.save("sd3_pag.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class StableDiffusion3PAGPipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, PAGMixin):
+    r"""
+    [PAG pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/pag) for text-to-image generation
+    using Stable Diffusion 3.
+
+    Args:
+        transformer ([`SD3Transformer2DModel`]):
+            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModelWithProjection`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant,
+            with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size`
+            as its dimension.
+        text_encoder_2 ([`CLIPTextModelWithProjection`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
+        text_encoder_3 ([`T5EncoderModel`]):
+            Frozen text-encoder. Stable Diffusion 3 uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
+            [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_3 (`T5TokenizerFast`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"]
+
+    def __init__(
+        self,
+        transformer: SD3Transformer2DModel,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer_2: CLIPTokenizer,
+        text_encoder_3: T5EncoderModel,
+        tokenizer_3: T5TokenizerFast,
+        pag_applied_layers: Union[str, List[str]] = "blocks.1",  # 1st transformer block
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            text_encoder_3=text_encoder_3,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            tokenizer_3=tokenizer_3,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
+        )
+        self.default_sample_size = (
+            self.transformer.config.sample_size
+            if hasattr(self, "transformer") and self.transformer is not None
+            else 128
+        )
+
+        self.set_pag_applied_layers(
+            pag_applied_layers, pag_attn_processors=(PAGCFGJointAttnProcessor2_0(), PAGJointAttnProcessor2_0())
+        )
+
+    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_images_per_prompt: int = 1,
+        max_sequence_length: int = 256,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        if self.text_encoder_3 is None:
+            return torch.zeros(
+                (
+                    batch_size * num_images_per_prompt,
+                    self.tokenizer_max_length,
+                    self.transformer.config.joint_attention_dim,
+                ),
+                device=device,
+                dtype=dtype,
+            )
+
+        text_inputs = self.tokenizer_3(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0]
+
+        dtype = self.text_encoder_3.dtype
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds
+    def _get_clip_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        clip_skip: Optional[int] = None,
+        clip_model_index: int = 0,
+    ):
+        device = device or self._execution_device
+
+        clip_tokenizers = [self.tokenizer, self.tokenizer_2]
+        clip_text_encoders = [self.text_encoder, self.text_encoder_2]
+
+        tokenizer = clip_tokenizers[clip_model_index]
+        text_encoder = clip_text_encoders[clip_model_index]
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer_max_length} tokens: {removed_text}"
+            )
+        prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+        pooled_prompt_embeds = prompt_embeds[0]
+
+        if clip_skip is None:
+            prompt_embeds = prompt_embeds.hidden_states[-2]
+        else:
+            prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1)
+
+        return prompt_embeds, pooled_prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_2: Union[str, List[str]],
+        prompt_3: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt_3: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        clip_skip: Optional[int] = None,
+        max_sequence_length: int = 256,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in all text-encoders
+            prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
+                used in all text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders.
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
+                `text_encoder_3`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder, lora_scale)
+            if self.text_encoder_2 is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            prompt_3 = prompt_3 or prompt
+            prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3
+
+            prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=0,
+            )
+            prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                prompt=prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=1,
+            )
+            clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1)
+
+            t5_prompt_embed = self._get_t5_prompt_embeds(
+                prompt=prompt_3,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+            clip_prompt_embeds = torch.nn.functional.pad(
+                clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1])
+            )
+
+            prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2)
+            pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+            negative_prompt_3 = negative_prompt_3 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+            negative_prompt_3 = (
+                batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3
+            )
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds(
+                negative_prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=0,
+            )
+            negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                negative_prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=1,
+            )
+            negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1)
+
+            t5_negative_prompt_embed = self._get_t5_prompt_embeds(
+                prompt=negative_prompt_3,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+            negative_clip_prompt_embeds = torch.nn.functional.pad(
+                negative_clip_prompt_embeds,
+                (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]),
+            )
+
+            negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2)
+            negative_pooled_prompt_embeds = torch.cat(
+                [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        prompt_3,
+        height,
+        width,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        negative_prompt_3=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        max_sequence_length=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_3 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+        elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)):
+            raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_3 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if max_sequence_length is not None and max_sequence_length > 512:
+            raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")
+
+    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.prepare_latents
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def joint_attention_kwargs(self):
+        return self._joint_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        prompt_3: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 28,
+        timesteps: List[int] = None,
+        guidance_scale: float = 7.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt_3: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 256,
+        pag_scale: float = 3.0,
+        pag_adaptive_scale: float = 0.0,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                will be used instead
+            prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
+                will be used instead
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used instead
+            negative_prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
+                `text_encoder_3`. If not defined, `negative_prompt` is used instead
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`.
+            pag_scale (`float`, *optional*, defaults to 3.0):
+                The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention
+                guidance will not be used.
+            pag_adaptive_scale (`float`, *optional*, defaults to 0.0):
+                The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is
+                used.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            prompt_3,
+            height,
+            width,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            negative_prompt_3=negative_prompt_3,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._clip_skip = clip_skip
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+        self._pag_scale = pag_scale
+        self._pag_adaptive_scale = pag_adaptive_scale  #
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        lora_scale = (
+            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_3=prompt_3,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            negative_prompt_3=negative_prompt_3,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            device=device,
+            clip_skip=self.clip_skip,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+
+        if self.do_perturbed_attention_guidance:
+            prompt_embeds = self._prepare_perturbed_attention_guidance(
+                prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance
+            )
+            pooled_prompt_embeds = self._prepare_perturbed_attention_guidance(
+                pooled_prompt_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance
+            )
+        elif self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        if self.do_perturbed_attention_guidance:
+            original_attn_proc = self.transformer.attn_processors
+            self._set_pag_attn_processor(
+                pag_applied_layers=self.pag_applied_layers,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+            )
+
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance, perturbed-attention guidance, or both
+                latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0]))
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    pooled_projections=pooled_prompt_embeds,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_perturbed_attention_guidance:
+                    noise_pred = self._apply_perturbed_attention_guidance(
+                        noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t
+                    )
+
+                elif self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            image = latents
+
+        else:
+            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if self.do_perturbed_attention_guidance:
+            self.transformer.set_attn_processor(original_attn_proc)
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusion3PipelineOutput(images=image)

src/diffusers/pipelines/pag/pipeline_pag_sd_animatediff.py

@@ -35,6 +35,7 @@ from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
 from ..animatediff.pipeline_output import AnimateDiffPipelineOutput
 from ..free_init_utils import FreeInitMixin
+from ..free_noise_utils import AnimateDiffFreeNoiseMixin
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from .pag_utils import PAGMixin
 
@@ -83,6 +84,7 @@ class AnimateDiffPAGPipeline(
     IPAdapterMixin,
     StableDiffusionLoraLoaderMixin,
     FreeInitMixin,
+    AnimateDiffFreeNoiseMixin,
     PAGMixin,
 ):
     r"""
@@ -404,15 +406,21 @@ class AnimateDiffPAGPipeline(
 
         return ip_adapter_image_embeds
 
-    # Copied from diffusers.pipelines.text_to_video_synthesis/pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
-    def decode_latents(self, latents):
+    # Copied from diffusers.pipelines.animatediff.pipeline_animatediff.AnimateDiffPipeline.decode_latents
+    def decode_latents(self, latents, decode_chunk_size: int = 16):
         latents = 1 / self.vae.config.scaling_factor * latents
 
         batch_size, channels, num_frames, height, width = latents.shape
         latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
 
-        image = self.vae.decode(latents).sample
-        video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4)
+        video = []
+        for i in range(0, latents.shape[0], decode_chunk_size):
+            batch_latents = latents[i : i + decode_chunk_size]
+            batch_latents = self.vae.decode(batch_latents).sample
+            video.append(batch_latents)
+
+        video = torch.cat(video)
+        video = video[None, :].reshape((batch_size, num_frames, -1) + video.shape[2:]).permute(0, 2, 1, 3, 4)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
         video = video.float()
         return video
@@ -499,10 +507,22 @@ class AnimateDiffPAGPipeline(
                     f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
                 )
 
-    # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
+    # Copied from diffusers.pipelines.animatediff.pipeline_animatediff.AnimateDiffPipeline.prepare_latents
     def prepare_latents(
         self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
     ):
+        # If FreeNoise is enabled, generate latents as described in Equation (7) of [FreeNoise](https://arxiv.org/abs/2310.15169)
+        if self.free_noise_enabled:
+            latents = self._prepare_latents_free_noise(
+                batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents
+            )
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
         shape = (
             batch_size,
             num_channels_latents,
@@ -510,11 +530,6 @@ class AnimateDiffPAGPipeline(
             height // self.vae_scale_factor,
             width // self.vae_scale_factor,
         )
-        if isinstance(generator, list) and len(generator) != batch_size:
-            raise ValueError(
-                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
-                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
-            )
 
         if latents is None:
             latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
@@ -573,6 +588,7 @@ class AnimateDiffPAGPipeline(
         clip_skip: Optional[int] = None,
         callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        decode_chunk_size: int = 16,
         pag_scale: float = 3.0,
         pag_adaptive_scale: float = 0.0,
     ):
@@ -831,7 +847,7 @@ class AnimateDiffPAGPipeline(
         if output_type == "latent":
             video = latents
         else:
-            video_tensor = self.decode_latents(latents)
+            video_tensor = self.decode_latents(latents, decode_chunk_size)
             video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type)
 
         # 10. Offload all models

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py

@@ -20,7 +20,7 @@ import numpy as np
 import PIL.Image
 import torch
 from packaging import version
-from transformers import CLIPTextModel, CLIPTokenizer, DPTFeatureExtractor, DPTForDepthEstimation
+from transformers import CLIPTextModel, CLIPTokenizer, DPTForDepthEstimation, DPTImageProcessor
 
 from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
@@ -111,7 +111,7 @@ class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoader
         unet: UNet2DConditionModel,
         scheduler: KarrasDiffusionSchedulers,
         depth_estimator: DPTForDepthEstimation,
-        feature_extractor: DPTFeatureExtractor,
+        feature_extractor: DPTImageProcessor,
     ):
         super().__init__()
 

src/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py

@@ -18,7 +18,7 @@ from typing import Any, Callable, Dict, List, Optional, Union
 
 import PIL.Image
 import torch
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from ...image_processor import VaeImageProcessor
 from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
@@ -138,7 +138,7 @@ class StableDiffusionGLIGENPipeline(DiffusionPipeline, StableDiffusionMixin):
         unet: UNet2DConditionModel,
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         requires_safety_checker: bool = True,
     ):
         super().__init__()

src/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py

@@ -19,7 +19,7 @@ from typing import Any, Callable, Dict, List, Optional, Union
 import PIL.Image
 import torch
 from transformers import (
-    CLIPFeatureExtractor,
+    CLIPImageProcessor,
     CLIPProcessor,
     CLIPTextModel,
     CLIPTokenizer,
@@ -193,7 +193,7 @@ class StableDiffusionGLIGENTextImagePipeline(DiffusionPipeline, StableDiffusionM
         unet: UNet2DConditionModel,
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         requires_safety_checker: bool = True,
     ):
         super().__init__()

src/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py

@@ -19,7 +19,7 @@ from typing import Any, Callable, Dict, List, Optional, Union
 import numpy as np
 import PIL.Image
 import torch
-from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from ...image_processor import VaeImageProcessor
 from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
@@ -209,7 +209,7 @@ class StableDiffusionAdapterPipeline(DiffusionPipeline, StableDiffusionMixin):
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 
@@ -225,7 +225,7 @@ class StableDiffusionAdapterPipeline(DiffusionPipeline, StableDiffusionMixin):
         adapter: Union[T2IAdapter, MultiAdapter, List[T2IAdapter]],
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPFeatureExtractor,
+        feature_extractor: CLIPImageProcessor,
         requires_safety_checker: bool = True,
     ):
         super().__init__()

src/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py

@@ -237,7 +237,7 @@ class StableDiffusionXLAdapterPipeline(
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
             Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
-        feature_extractor ([`CLIPFeatureExtractor`]):
+        feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 

src/diffusers/schedulers/scheduling_ddim_cogvideox.py

@@ -29,8 +29,8 @@ from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
 
 
 @dataclass
-# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->CogVideoXDDIM
-class CogVideoXDDIMSchedulerOutput(BaseOutput):
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM
+class DDIMSchedulerOutput(BaseOutput):
     """
     Output class for the scheduler's `step` function output.
 
@@ -312,7 +312,7 @@ class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin):
         generator=None,
         variance_noise: Optional[torch.Tensor] = None,
         return_dict: bool = True,
-    ) -> Union[CogVideoXDDIMSchedulerOutput, Tuple]:
+    ) -> Union[DDIMSchedulerOutput, Tuple]:
         """
         Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
         process from the learned model outputs (most often the predicted noise).
@@ -337,13 +337,12 @@ class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin):
                 Alternative to generating noise with `generator` by directly providing the noise for the variance
                 itself. Useful for methods such as [`CycleDiffusion`].
             return_dict (`bool`, *optional*, defaults to `True`):
-                Whether or not to return a [`~schedulers.scheduling_ddim_cogvideox.CogVideoXDDIMSchedulerOutput`] or
-                `tuple`.
+                Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`.
 
         Returns:
-            [`~schedulers.scheduling_ddim_cogvideox.CogVideoXDDIMSchedulerOutput`] or `tuple`:
-                If return_dict is `True`, [`~schedulers.scheduling_ddim_cogvideox.CogVideoXDDIMSchedulerOutput`] is
-                returned, otherwise a tuple is returned where the first element is the sample tensor.
+            [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a
+                tuple is returned where the first element is the sample tensor.
 
         """
         if self.num_inference_steps is None:
@@ -397,7 +396,7 @@ class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin):
         if not return_dict:
             return (prev_sample,)
 
-        return CogVideoXDDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
+        return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
 
     # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
     def add_noise(

src/diffusers/schedulers/scheduling_dpm_cogvideox.py

@@ -30,8 +30,8 @@ from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
 
 
 @dataclass
-# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->CogVideoX
-class CogVideoXDPMSchedulerOutput(BaseOutput):
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM
+class DDIMSchedulerOutput(BaseOutput):
     """
     Output class for the scheduler's `step` function output.
 
@@ -93,7 +93,6 @@ def betas_for_alpha_bar(
     return torch.tensor(betas, dtype=torch.float32)
 
 
-# Copied from diffusers.schedulers.scheduling_ddim_cogvideox.rescale_zero_terminal_snr
 def rescale_zero_terminal_snr(alphas_cumprod):
     """
     Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
@@ -232,6 +231,16 @@ class CogVideoXDPMScheduler(SchedulerMixin, ConfigMixin):
         self.num_inference_steps = None
         self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
 
+    def _get_variance(self, timestep, prev_timestep):
+        alpha_prod_t = self.alphas_cumprod[timestep]
+        alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+        beta_prod_t = 1 - alpha_prod_t
+        beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+        variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+
+        return variance
+
     def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor:
         """
         Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
@@ -330,7 +339,7 @@ class CogVideoXDPMScheduler(SchedulerMixin, ConfigMixin):
         generator=None,
         variance_noise: Optional[torch.Tensor] = None,
         return_dict: bool = False,
-    ) -> Union[CogVideoXDPMSchedulerOutput, Tuple]:
+    ) -> Union[DDIMSchedulerOutput, Tuple]:
         """
         Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
         process from the learned model outputs (most often the predicted noise).
@@ -355,13 +364,12 @@ class CogVideoXDPMScheduler(SchedulerMixin, ConfigMixin):
                 Alternative to generating noise with `generator` by directly providing the noise for the variance
                 itself. Useful for methods such as [`CycleDiffusion`].
             return_dict (`bool`, *optional*, defaults to `True`):
-                Whether or not to return a [`~schedulers.scheduling_dpm_cogvideox.CogVideoXDPMDPMSchedulerOutput`] or
-                `tuple`.
+                Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`.
 
         Returns:
-            [`~schedulers.scheduling_dpm_cogvideox.CogVideoXDPMSchedulerOutput`] or `tuple`:
-                If return_dict is `True`, [`~schedulers.scheduling_dpm_cogvideox.CogVideoXDPMSchedulerOutput`] is
-                returned, otherwise a tuple is returned where the first element is the sample tensor.
+            [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a
+                tuple is returned where the first element is the sample tensor.
 
         """
         if self.num_inference_steps is None:
@@ -428,7 +436,7 @@ class CogVideoXDPMScheduler(SchedulerMixin, ConfigMixin):
         if not return_dict:
             return (prev_sample, pred_original_sample)
 
-        return CogVideoXDPMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
+        return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
 
     # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
     def add_noise(

src/diffusers/utils/__init__.py

@@ -78,6 +78,7 @@ from .import_utils import (
     is_peft_version,
     is_safetensors_available,
     is_scipy_available,
+    is_sentencepiece_available,
     is_tensorboard_available,
     is_timm_available,
     is_torch_available,

src/diffusers/utils/dummy_torch_and_transformers_and_sentencepiece_objects.py

@@ -0,0 +1,47 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class KolorsImg2ImgPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers", "sentencepiece"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers", "sentencepiece"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers", "sentencepiece"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers", "sentencepiece"])
+
+
+class KolorsPAGPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers", "sentencepiece"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers", "sentencepiece"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers", "sentencepiece"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers", "sentencepiece"])
+
+
+class KolorsPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers", "sentencepiece"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers", "sentencepiece"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers", "sentencepiece"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers", "sentencepiece"])

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -242,36 +242,6 @@ class AuraFlowPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
-class ChatGLMModel(metaclass=DummyObject):
-    _backends = ["torch", "transformers"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["torch", "transformers"])
-
-    @classmethod
-    def from_config(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-
-class ChatGLMTokenizer(metaclass=DummyObject):
-    _backends = ["torch", "transformers"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["torch", "transformers"])
-
-    @classmethod
-    def from_config(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-
 class CLIPImageProjection(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 
@@ -782,36 +752,6 @@ class KandinskyV22PriorPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
-class KolorsImg2ImgPipeline(metaclass=DummyObject):
-    _backends = ["torch", "transformers"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["torch", "transformers"])
-
-    @classmethod
-    def from_config(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-
-class KolorsPipeline(metaclass=DummyObject):
-    _backends = ["torch", "transformers"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["torch", "transformers"])
-
-    @classmethod
-    def from_config(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-
 class LatentConsistencyModelImg2ImgPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 
@@ -1202,6 +1142,21 @@ class StableDiffusion3InpaintPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class StableDiffusion3PAGPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class StableDiffusion3Pipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

src/diffusers/utils/import_utils.py

@@ -294,6 +294,13 @@ try:
 except importlib_metadata.PackageNotFoundError:
     _torchvision_available = False
 
+_sentencepiece_available = importlib.util.find_spec("sentencepiece") is not None
+try:
+    _sentencepiece_version = importlib_metadata.version("sentencepiece")
+    logger.info(f"Successfully imported sentencepiece version {_sentencepiece_version}")
+except importlib_metadata.PackageNotFoundError:
+    _sentencepiece_available = False
+
 _matplotlib_available = importlib.util.find_spec("matplotlib") is not None
 try:
     _matplotlib_version = importlib_metadata.version("matplotlib")
@@ -436,6 +443,10 @@ def is_google_colab():
     return _is_google_colab
 
 
+def is_sentencepiece_available():
+    return _sentencepiece_available
+
+
 # docstyle-ignore
 FLAX_IMPORT_ERROR = """
 {0} requires the FLAX library but it was not found in your environment. Checkout the instructions on the
@@ -553,6 +564,12 @@ SAFETENSORS_IMPORT_ERROR = """
 {0} requires the safetensors library but it was not found in your environment. You can install it with pip: `pip install safetensors`
 """
 
+# docstyle-ignore
+SENTENCEPIECE_IMPORT_ERROR = """
+{0} requires the sentencepiece library but it was not found in your environment. You can install it with pip: `pip install sentencepiece`
+"""
+
+
 # docstyle-ignore
 BITSANDBYTES_IMPORT_ERROR = """
 {0} requires the bitsandbytes library but it was not found in your environment. You can install it with pip: `pip install bitsandbytes`
@@ -581,6 +598,7 @@ BACKENDS_MAPPING = OrderedDict(
         ("peft", (is_peft_available, PEFT_IMPORT_ERROR)),
         ("safetensors", (is_safetensors_available, SAFETENSORS_IMPORT_ERROR)),
         ("bitsandbytes", (is_bitsandbytes_available, BITSANDBYTES_IMPORT_ERROR)),
+        ("sentencepiece", (is_sentencepiece_available, SENTENCEPIECE_IMPORT_ERROR)),
     ]
 )
 

tests/models/transformers/test_models_transformer_hunyuan_dit.py

@@ -0,0 +1,113 @@
+# 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 unittest
+
+import torch
+
+from diffusers import HunyuanDiT2DModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    torch_device,
+)
+
+from ..test_modeling_common import ModelTesterMixin
+
+
+enable_full_determinism()
+
+
+class HunyuanDiTTests(ModelTesterMixin, unittest.TestCase):
+    model_class = HunyuanDiT2DModel
+    main_input_name = "hidden_states"
+
+    @property
+    def dummy_input(self):
+        batch_size = 2
+        num_channels = 4
+        height = width = 8
+        embedding_dim = 8
+        sequence_length = 4
+        sequence_length_t5 = 4
+
+        hidden_states = torch.randn((batch_size, num_channels, height, width)).to(torch_device)
+        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
+        text_embedding_mask = torch.ones(size=(batch_size, sequence_length)).to(torch_device)
+        encoder_hidden_states_t5 = torch.randn((batch_size, sequence_length_t5, embedding_dim)).to(torch_device)
+        text_embedding_mask_t5 = torch.ones(size=(batch_size, sequence_length_t5)).to(torch_device)
+        timestep = torch.randint(0, 1000, size=(batch_size,), dtype=encoder_hidden_states.dtype).to(torch_device)
+
+        original_size = [1024, 1024]
+        target_size = [16, 16]
+        crops_coords_top_left = [0, 0]
+        add_time_ids = list(original_size + target_size + crops_coords_top_left)
+        add_time_ids = torch.tensor([add_time_ids, add_time_ids], dtype=encoder_hidden_states.dtype).to(torch_device)
+        style = torch.zeros(size=(batch_size,), dtype=int).to(torch_device)
+        image_rotary_emb = [
+            torch.ones(size=(1, 8), dtype=encoder_hidden_states.dtype),
+            torch.zeros(size=(1, 8), dtype=encoder_hidden_states.dtype),
+        ]
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "text_embedding_mask": text_embedding_mask,
+            "encoder_hidden_states_t5": encoder_hidden_states_t5,
+            "text_embedding_mask_t5": text_embedding_mask_t5,
+            "timestep": timestep,
+            "image_meta_size": add_time_ids,
+            "style": style,
+            "image_rotary_emb": image_rotary_emb,
+        }
+
+    @property
+    def input_shape(self):
+        return (4, 8, 8)
+
+    @property
+    def output_shape(self):
+        return (8, 8, 8)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
+            "sample_size": 8,
+            "patch_size": 2,
+            "in_channels": 4,
+            "num_layers": 1,
+            "attention_head_dim": 8,
+            "num_attention_heads": 2,
+            "cross_attention_dim": 8,
+            "cross_attention_dim_t5": 8,
+            "pooled_projection_dim": 4,
+            "hidden_size": 16,
+            "text_len": 4,
+            "text_len_t5": 4,
+            "activation_fn": "gelu-approximate",
+        }
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict
+
+    def test_output(self):
+        super().test_output(
+            expected_output_shape=(self.dummy_input[self.main_input_name].shape[0],) + self.output_shape
+        )
+
+    @unittest.skip("HunyuanDIT use a custom processor HunyuanAttnProcessor2_0")
+    def test_set_xformers_attn_processor_for_determinism(self):
+        pass
+
+    @unittest.skip("HunyuanDIT use a custom processor HunyuanAttnProcessor2_0")
+    def test_set_attn_processor_for_determinism(self):
+        pass

tests/models/unets/test_models_unet_2d_condition.py

@@ -1121,6 +1121,18 @@ class UNet2DConditionModelTests(ModelTesterMixin, UNetTesterMixin, unittest.Test
         assert loaded_model
         assert new_output.sample.shape == (4, 4, 16, 16)
 
+    @require_torch_gpu
+    def test_load_sharded_checkpoint_with_variant_from_hub(self):
+        _, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        loaded_model = self.model_class.from_pretrained(
+            "hf-internal-testing/unet2d-sharded-with-variant-dummy", variant="fp16"
+        )
+        loaded_model = loaded_model.to(torch_device)
+        new_output = loaded_model(**inputs_dict)
+
+        assert loaded_model
+        assert new_output.sample.shape == (4, 4, 16, 16)
+
     @require_peft_backend
     def test_lora(self):
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

tests/pipelines/animatediff/test_animatediff.py

@@ -17,6 +17,7 @@ from diffusers import (
     UNet2DConditionModel,
     UNetMotionModel,
 )
+from diffusers.models.attention import FreeNoiseTransformerBlock
 from diffusers.utils import is_xformers_available, logging
 from diffusers.utils.testing_utils import numpy_cosine_similarity_distance, require_torch_gpu, slow, torch_device
 
@@ -401,6 +402,64 @@ class AnimateDiffPipelineFastTests(
                 "Enabling of FreeInit should lead to results different from the default pipeline results",
             )
 
+    def test_free_noise_blocks(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        pipe.enable_free_noise()
+        for block in pipe.unet.down_blocks:
+            for motion_module in block.motion_modules:
+                for transformer_block in motion_module.transformer_blocks:
+                    self.assertTrue(
+                        isinstance(transformer_block, FreeNoiseTransformerBlock),
+                        "Motion module transformer blocks must be an instance of `FreeNoiseTransformerBlock` after enabling FreeNoise.",
+                    )
+
+        pipe.disable_free_noise()
+        for block in pipe.unet.down_blocks:
+            for motion_module in block.motion_modules:
+                for transformer_block in motion_module.transformer_blocks:
+                    self.assertFalse(
+                        isinstance(transformer_block, FreeNoiseTransformerBlock),
+                        "Motion module transformer blocks must not be an instance of `FreeNoiseTransformerBlock` after disabling FreeNoise.",
+                    )
+
+    def test_free_noise(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        inputs_normal = self.get_dummy_inputs(torch_device)
+        frames_normal = pipe(**inputs_normal).frames[0]
+
+        for context_length in [8, 9]:
+            for context_stride in [4, 6]:
+                pipe.enable_free_noise(context_length, context_stride)
+
+                inputs_enable_free_noise = self.get_dummy_inputs(torch_device)
+                frames_enable_free_noise = pipe(**inputs_enable_free_noise).frames[0]
+
+                pipe.disable_free_noise()
+
+                inputs_disable_free_noise = self.get_dummy_inputs(torch_device)
+                frames_disable_free_noise = pipe(**inputs_disable_free_noise).frames[0]
+
+                sum_enabled = np.abs(to_np(frames_normal) - to_np(frames_enable_free_noise)).sum()
+                max_diff_disabled = np.abs(to_np(frames_normal) - to_np(frames_disable_free_noise)).max()
+                self.assertGreater(
+                    sum_enabled,
+                    1e1,
+                    "Enabling of FreeNoise should lead to results different from the default pipeline results",
+                )
+                self.assertLess(
+                    max_diff_disabled,
+                    1e-4,
+                    "Disabling of FreeNoise should lead to results similar to the default pipeline results",
+                )
+
     @unittest.skipIf(
         torch_device != "cuda" or not is_xformers_available(),
         reason="XFormers attention is only available with CUDA and `xformers` installed",

tests/pipelines/animatediff/test_animatediff_controlnet.py

@@ -18,6 +18,7 @@ from diffusers import (
     UNet2DConditionModel,
     UNetMotionModel,
 )
+from diffusers.models.attention import FreeNoiseTransformerBlock
 from diffusers.utils import logging
 from diffusers.utils.testing_utils import torch_device
 
@@ -409,6 +410,64 @@ class AnimateDiffControlNetPipelineFastTests(
                 "Enabling of FreeInit should lead to results different from the default pipeline results",
             )
 
+    def test_free_noise_blocks(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffControlNetPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        pipe.enable_free_noise()
+        for block in pipe.unet.down_blocks:
+            for motion_module in block.motion_modules:
+                for transformer_block in motion_module.transformer_blocks:
+                    self.assertTrue(
+                        isinstance(transformer_block, FreeNoiseTransformerBlock),
+                        "Motion module transformer blocks must be an instance of `FreeNoiseTransformerBlock` after enabling FreeNoise.",
+                    )
+
+        pipe.disable_free_noise()
+        for block in pipe.unet.down_blocks:
+            for motion_module in block.motion_modules:
+                for transformer_block in motion_module.transformer_blocks:
+                    self.assertFalse(
+                        isinstance(transformer_block, FreeNoiseTransformerBlock),
+                        "Motion module transformer blocks must not be an instance of `FreeNoiseTransformerBlock` after disabling FreeNoise.",
+                    )
+
+    def test_free_noise(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffControlNetPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        inputs_normal = self.get_dummy_inputs(torch_device, num_frames=16)
+        frames_normal = pipe(**inputs_normal).frames[0]
+
+        for context_length in [8, 9]:
+            for context_stride in [4, 6]:
+                pipe.enable_free_noise(context_length, context_stride)
+
+                inputs_enable_free_noise = self.get_dummy_inputs(torch_device, num_frames=16)
+                frames_enable_free_noise = pipe(**inputs_enable_free_noise).frames[0]
+
+                pipe.disable_free_noise()
+
+                inputs_disable_free_noise = self.get_dummy_inputs(torch_device, num_frames=16)
+                frames_disable_free_noise = pipe(**inputs_disable_free_noise).frames[0]
+
+                sum_enabled = np.abs(to_np(frames_normal) - to_np(frames_enable_free_noise)).sum()
+                max_diff_disabled = np.abs(to_np(frames_normal) - to_np(frames_disable_free_noise)).max()
+                self.assertGreater(
+                    sum_enabled,
+                    1e1,
+                    "Enabling of FreeNoise should lead to results different from the default pipeline results",
+                )
+                self.assertLess(
+                    max_diff_disabled,
+                    1e-4,
+                    "Disabling of FreeNoise should lead to results similar to the default pipeline results",
+                )
+
     def test_vae_slicing(self, video_count=2):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()

tests/pipelines/animatediff/test_animatediff_video2video.py

@@ -17,6 +17,7 @@ from diffusers import (
     UNet2DConditionModel,
     UNetMotionModel,
 )
+from diffusers.models.attention import FreeNoiseTransformerBlock
 from diffusers.utils import is_xformers_available, logging
 from diffusers.utils.testing_utils import torch_device
 
@@ -114,7 +115,7 @@ class AnimateDiffVideoToVideoPipelineFastTests(
         }
         return components
 
-    def get_dummy_inputs(self, device, seed=0):
+    def get_dummy_inputs(self, device, seed=0, num_frames: int = 2):
         if str(device).startswith("mps"):
             generator = torch.manual_seed(seed)
         else:
@@ -122,8 +123,7 @@ class AnimateDiffVideoToVideoPipelineFastTests(
 
         video_height = 32
         video_width = 32
-        video_num_frames = 2
-        video = [Image.new("RGB", (video_width, video_height))] * video_num_frames
+        video = [Image.new("RGB", (video_width, video_height))] * num_frames
 
         inputs = {
             "video": video,
@@ -428,3 +428,66 @@ class AnimateDiffVideoToVideoPipelineFastTests(
                 1e1,
                 "Enabling of FreeInit should lead to results different from the default pipeline results",
             )
+
+    def test_free_noise_blocks(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffVideoToVideoPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        pipe.enable_free_noise()
+        for block in pipe.unet.down_blocks:
+            for motion_module in block.motion_modules:
+                for transformer_block in motion_module.transformer_blocks:
+                    self.assertTrue(
+                        isinstance(transformer_block, FreeNoiseTransformerBlock),
+                        "Motion module transformer blocks must be an instance of `FreeNoiseTransformerBlock` after enabling FreeNoise.",
+                    )
+
+        pipe.disable_free_noise()
+        for block in pipe.unet.down_blocks:
+            for motion_module in block.motion_modules:
+                for transformer_block in motion_module.transformer_blocks:
+                    self.assertFalse(
+                        isinstance(transformer_block, FreeNoiseTransformerBlock),
+                        "Motion module transformer blocks must not be an instance of `FreeNoiseTransformerBlock` after disabling FreeNoise.",
+                    )
+
+    def test_free_noise(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffVideoToVideoPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        inputs_normal = self.get_dummy_inputs(torch_device, num_frames=16)
+        inputs_normal["num_inference_steps"] = 2
+        inputs_normal["strength"] = 0.5
+        frames_normal = pipe(**inputs_normal).frames[0]
+
+        for context_length in [8, 9]:
+            for context_stride in [4, 6]:
+                pipe.enable_free_noise(context_length, context_stride)
+
+                inputs_enable_free_noise = self.get_dummy_inputs(torch_device, num_frames=16)
+                inputs_enable_free_noise["num_inference_steps"] = 2
+                inputs_enable_free_noise["strength"] = 0.5
+                frames_enable_free_noise = pipe(**inputs_enable_free_noise).frames[0]
+
+                pipe.disable_free_noise()
+                inputs_disable_free_noise = self.get_dummy_inputs(torch_device, num_frames=16)
+                inputs_disable_free_noise["num_inference_steps"] = 2
+                inputs_disable_free_noise["strength"] = 0.5
+                frames_disable_free_noise = pipe(**inputs_disable_free_noise).frames[0]
+
+                sum_enabled = np.abs(to_np(frames_normal) - to_np(frames_enable_free_noise)).sum()
+                max_diff_disabled = np.abs(to_np(frames_normal) - to_np(frames_disable_free_noise)).max()
+                self.assertGreater(
+                    sum_enabled,
+                    1e1,
+                    "Enabling of FreeNoise should lead to results different from the default pipeline results",
+                )
+                self.assertLess(
+                    max_diff_disabled,
+                    1e-4,
+                    "Disabling of FreeNoise should lead to results similar to the default pipeline results",
+                )

tests/pipelines/aura_flow/test_pipeline_aura_flow.py

@@ -9,7 +9,11 @@ from diffusers.utils.testing_utils import (
     torch_device,
 )
 
-from ..test_pipelines_common import PipelineTesterMixin
+from ..test_pipelines_common import (
+    PipelineTesterMixin,
+    check_qkv_fusion_matches_attn_procs_length,
+    check_qkv_fusion_processors_exist,
+)
 
 
 class AuraFlowPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
@@ -119,3 +123,43 @@ class AuraFlowPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         # Attention slicing needs to implemented differently for this because how single DiT and MMDiT
         # blocks interfere with each other.
         return
+
+    def test_fused_qkv_projections(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        original_image_slice = image[0, -3:, -3:, -1]
+
+        # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
+        # to the pipeline level.
+        pipe.transformer.fuse_qkv_projections()
+        assert check_qkv_fusion_processors_exist(
+            pipe.transformer
+        ), "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        assert check_qkv_fusion_matches_attn_procs_length(
+            pipe.transformer, pipe.transformer.original_attn_processors
+        ), "Something wrong with the attention processors concerning the fused QKV projections."
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice_fused = image[0, -3:, -3:, -1]
+
+        pipe.transformer.unfuse_qkv_projections()
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice_disabled = image[0, -3:, -3:, -1]
+
+        assert np.allclose(
+            original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
+        ), "Fusion of QKV projections shouldn't affect the outputs."
+        assert np.allclose(
+            image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
+        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
+        assert np.allclose(
+            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
+        ), "Original outputs should match when fused QKV projections are disabled."

tests/pipelines/cogvideox/test_cogvideox.py

@@ -14,14 +14,13 @@
 
 import gc
 import inspect
-import tempfile
 import unittest
 
 import numpy as np
 import torch
 from transformers import AutoTokenizer, T5EncoderModel
 
-from diffusers import AutoencoderKL, CogVideoXPipeline, CogVideoXTransformer3DModel, DDIMScheduler
+from diffusers import AutoencoderKLCogVideoX, CogVideoXPipeline, CogVideoXTransformer3DModel, DDIMScheduler
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
     numpy_cosine_similarity_distance,
@@ -43,41 +42,71 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
     image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
     image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
-
-    required_optional_params = PipelineTesterMixin.required_optional_params
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
 
     def get_dummy_components(self):
         torch.manual_seed(0)
         transformer = CogVideoXTransformer3DModel(
-            sample_size=8,
-            num_layers=1,
-            patch_size=2,
+            # Product of num_attention_heads * attention_head_dim must be divisible by 16 for 3D positional embeddings
+            # But, since we are using tiny-random-t5 here, we need the internal dim of CogVideoXTransformer3DModel
+            # to be 32. The internal dim is product of num_attention_heads and attention_head_dim
+            num_attention_heads=4,
             attention_head_dim=8,
-            num_attention_heads=3,
-            caption_channels=32,
             in_channels=4,
-            cross_attention_dim=24,
-            out_channels=8,
-            attention_bias=True,
-            activation_fn="gelu-approximate",
-            num_embeds_ada_norm=1000,
-            norm_type="ada_norm_single",
-            norm_elementwise_affine=False,
-            norm_eps=1e-6,
+            out_channels=4,
+            time_embed_dim=2,
+            text_embed_dim=32,  # Must match with tiny-random-t5
+            num_layers=1,
+            sample_width=16,  # latent width: 2 -> final width: 16
+            sample_height=16,  # latent height: 2 -> final height: 16
+            sample_frames=9,  # latent frames: (9 - 1) / 4 + 1 = 3 -> final frames: 9
+            patch_size=2,
+            temporal_compression_ratio=4,
+            max_text_seq_length=16,
         )
-        torch.manual_seed(0)
-        vae = AutoencoderKL()
 
+        torch.manual_seed(0)
+        vae = AutoencoderKLCogVideoX(
+            in_channels=3,
+            out_channels=3,
+            down_block_types=(
+                "CogVideoXDownBlock3D",
+                "CogVideoXDownBlock3D",
+                "CogVideoXDownBlock3D",
+                "CogVideoXDownBlock3D",
+            ),
+            up_block_types=(
+                "CogVideoXUpBlock3D",
+                "CogVideoXUpBlock3D",
+                "CogVideoXUpBlock3D",
+                "CogVideoXUpBlock3D",
+            ),
+            block_out_channels=(8, 8, 8, 8),
+            latent_channels=4,
+            layers_per_block=1,
+            norm_num_groups=2,
+            temporal_compression_ratio=4,
+        )
+
+        torch.manual_seed(0)
         scheduler = DDIMScheduler()
         text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
-
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {
-            "transformer": transformer.eval(),
-            "vae": vae.eval(),
+            "transformer": transformer,
+            "vae": vae,
             "scheduler": scheduler,
-            "text_encoder": text_encoder.eval(),
+            "text_encoder": text_encoder,
             "tokenizer": tokenizer,
         }
         return components
@@ -88,16 +117,22 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         else:
             generator = torch.Generator(device=device).manual_seed(seed)
         inputs = {
-            "prompt": "A painting of a squirrel eating a burger",
-            "negative_prompt": "low quality",
+            "prompt": "dance monkey",
+            "negative_prompt": "",
             "generator": generator,
             "num_inference_steps": 2,
-            "guidance_scale": 5.0,
-            "height": 8,
-            "width": 8,
-            "video_length": 1,
+            "guidance_scale": 6.0,
+            # Cannot reduce because convolution kernel becomes bigger than sample
+            "height": 16,
+            "width": 16,
+            # TODO(aryan): improve this
+            # Cannot make this lower due to assert condition in pipeline at the moment.
+            # The reason why 8 can't be used here is due to how context-parallel cache works where the first
+            # second of video is decoded from latent frames (0, 3) instead of [(0, 2), (2, 3)]. If 8 is used,
+            # the number of output frames that you get are 5.
+            "num_frames": 8,
+            "max_sequence_length": 16,
             "output_type": "pt",
-            "clean_caption": False,
         }
         return inputs
 
@@ -113,8 +148,8 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         video = pipe(**inputs).frames
         generated_video = video[0]
 
-        self.assertEqual(generated_video.shape, (1, 3, 8, 8))
-        expected_video = torch.randn(1, 3, 8, 8)
+        self.assertEqual(generated_video.shape, (9, 3, 16, 16))
+        expected_video = torch.randn(9, 3, 16, 16)
         max_diff = np.abs(generated_video - expected_video).max()
         self.assertLessEqual(max_diff, 1e10)
 
@@ -180,76 +215,41 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-3)
 
-    def test_attention_slicing_forward_pass(self):
-        pass
-
-    def test_save_load_optional_components(self):
-        if not hasattr(self.pipeline_class, "_optional_components"):
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
             return
 
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
-
         for component in pipe.components.values():
             if hasattr(component, "set_default_attn_processor"):
                 component.set_default_attn_processor()
         pipe.to(torch_device)
         pipe.set_progress_bar_config(disable=None)
 
-        inputs = self.get_dummy_inputs(torch_device)
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
 
-        prompt = inputs["prompt"]
-        generator = inputs["generator"]
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
 
-        (
-            prompt_embeds,
-            negative_prompt_embeds,
-        ) = pipe.encode_prompt(prompt)
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
 
-        # inputs with prompt converted to embeddings
-        inputs = {
-            "prompt_embeds": prompt_embeds,
-            "negative_prompt": None,
-            "negative_prompt_embeds": negative_prompt_embeds,
-            "generator": generator,
-            "num_inference_steps": 2,
-            "guidance_scale": 5.0,
-            "height": 8,
-            "width": 8,
-            "video_length": 1,
-            "mask_feature": False,
-            "output_type": "pt",
-            "clean_caption": False,
-        }
-
-        # set all optional components to None
-        for optional_component in pipe._optional_components:
-            setattr(pipe, optional_component, None)
-
-        output = pipe(**inputs)[0]
-
-        with tempfile.TemporaryDirectory() as tmpdir:
-            pipe.save_pretrained(tmpdir, safe_serialization=False)
-            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir)
-            pipe_loaded.to(torch_device)
-
-            for component in pipe_loaded.components.values():
-                if hasattr(component, "set_default_attn_processor"):
-                    component.set_default_attn_processor()
-
-            pipe_loaded.set_progress_bar_config(disable=None)
-
-        for optional_component in pipe._optional_components:
-            self.assertTrue(
-                getattr(pipe_loaded, optional_component) is None,
-                f"`{optional_component}` did not stay set to None after loading.",
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
             )
 
-        output_loaded = pipe_loaded(**inputs)[0]
-
-        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
-        self.assertLess(max_diff, 1.0)
-
 
 @slow
 @require_torch_gpu
@@ -269,21 +269,22 @@ class CogVideoXPipelineIntegrationTests(unittest.TestCase):
     def test_cogvideox(self):
         generator = torch.Generator("cpu").manual_seed(0)
 
-        pipe = CogVideoXPipeline.from_pretrained("THUDM/cogvideox-2b", torch_dtype=torch.float16)
+        pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16)
         pipe.enable_model_cpu_offload()
         prompt = self.prompt
 
         videos = pipe(
             prompt=prompt,
-            height=512,
-            width=512,
+            height=480,
+            width=720,
+            num_frames=16,
             generator=generator,
             num_inference_steps=2,
-            clean_caption=False,
+            output_type="pt",
         ).frames
 
         video = videos[0]
-        expected_video = torch.randn(1, 512, 512, 3).numpy()
+        expected_video = torch.randn(1, 16, 480, 720, 3).numpy()
 
-        max_diff = numpy_cosine_similarity_distance(video.fCogVideoXn(), expected_video)
-        assert max_diff < 1e-3, f"Max diff is too high. got {video.fCogVideoXn()}"
+        max_diff = numpy_cosine_similarity_distance(video, expected_video)
+        assert max_diff < 1e-3, f"Max diff is too high. got {video}"

tests/pipelines/kolors/test_kolors.py

@@ -20,12 +20,11 @@ import torch
 
 from diffusers import (
     AutoencoderKL,
-    ChatGLMModel,
-    ChatGLMTokenizer,
     EulerDiscreteScheduler,
     KolorsPipeline,
     UNet2DConditionModel,
 )
+from diffusers.pipelines.kolors import ChatGLMModel, ChatGLMTokenizer
 from diffusers.utils.testing_utils import enable_full_determinism
 
 from ..pipeline_params import (
@@ -134,23 +133,11 @@ class KolorsPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         max_diff = np.abs(image_slice.flatten() - expected_slice).max()
         self.assertLessEqual(max_diff, 1e-3)
 
-    # throws AttributeError: property 'eos_token' of 'ChatGLMTokenizer' object has no setter
-    # not sure if it is worth to fix it before integrating it to transformers
     def test_save_load_optional_components(self):
-        # TODO (Alvaro) need to fix later
-        pass
+        super().test_save_load_optional_components(expected_max_difference=2e-4)
 
-    # throws AttributeError: property 'eos_token' of 'ChatGLMTokenizer' object has no setter
-    # not sure if it is worth to fix it before integrating it to transformers
     def test_save_load_float16(self):
-        # TODO (Alvaro) need to fix later
-        pass
-
-    # throws AttributeError: property 'eos_token' of 'ChatGLMTokenizer' object has no setter
-    # not sure if it is worth to fix it before integrating it to transformers
-    def test_save_load_local(self):
-        # TODO (Alvaro) need to fix later
-        pass
+        super().test_save_load_float16(expected_max_diff=2e-1)
 
     def test_inference_batch_single_identical(self):
-        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=5e-4)
+        self._test_inference_batch_single_identical(expected_max_diff=5e-4)

tests/pipelines/kolors/test_kolors_img2img.py

@@ -0,0 +1,152 @@
+# 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 random
+import unittest
+
+import numpy as np
+import torch
+
+from diffusers import (
+    AutoencoderKL,
+    EulerDiscreteScheduler,
+    KolorsImg2ImgPipeline,
+    UNet2DConditionModel,
+)
+from diffusers.pipelines.kolors import ChatGLMModel, ChatGLMTokenizer
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    floats_tensor,
+)
+
+from ..pipeline_params import (
+    TEXT_TO_IMAGE_BATCH_PARAMS,
+    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
+    TEXT_TO_IMAGE_IMAGE_PARAMS,
+    TEXT_TO_IMAGE_PARAMS,
+)
+from ..test_pipelines_common import PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+class KolorsPipelineImg2ImgFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = KolorsImg2ImgPipeline
+    params = TEXT_TO_IMAGE_PARAMS
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union({"add_text_embeds", "add_time_ids"})
+
+    # Copied from tests.pipelines.kolors.test_kolors.KolorsPipelineFastTests.get_dummy_components
+    def get_dummy_components(self, time_cond_proj_dim=None):
+        torch.manual_seed(0)
+        unet = UNet2DConditionModel(
+            block_out_channels=(2, 4),
+            layers_per_block=2,
+            time_cond_proj_dim=time_cond_proj_dim,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            # specific config below
+            attention_head_dim=(2, 4),
+            use_linear_projection=True,
+            addition_embed_type="text_time",
+            addition_time_embed_dim=8,
+            transformer_layers_per_block=(1, 2),
+            projection_class_embeddings_input_dim=56,
+            cross_attention_dim=8,
+            norm_num_groups=1,
+        )
+        scheduler = EulerDiscreteScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            steps_offset=1,
+            beta_schedule="scaled_linear",
+            timestep_spacing="leading",
+        )
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+            sample_size=128,
+        )
+        torch.manual_seed(0)
+        text_encoder = ChatGLMModel.from_pretrained("hf-internal-testing/tiny-random-chatglm3-6b")
+        tokenizer = ChatGLMTokenizer.from_pretrained("hf-internal-testing/tiny-random-chatglm3-6b")
+
+        components = {
+            "unet": unet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "image_encoder": None,
+            "feature_extractor": None,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        image = floats_tensor((1, 3, 64, 64), rng=random.Random(seed)).to(device)
+        image = image / 2 + 0.5
+
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "image": image,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "output_type": "np",
+            "strength": 0.8,
+        }
+
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        self.assertEqual(image.shape, (1, 64, 64, 3))
+        expected_slice = np.array(
+            [0.54823864, 0.43654007, 0.4886489, 0.63072854, 0.53641886, 0.4896852, 0.62123513, 0.5621531, 0.42809626]
+        )
+        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
+        self.assertLessEqual(max_diff, 1e-3)
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=3e-3)
+
+    def test_float16_inference(self):
+        super().test_float16_inference(expected_max_diff=7e-2)

tests/pipelines/pag/test_pag_animatediff.py

@@ -17,6 +17,7 @@ from diffusers import (
     UNet2DConditionModel,
     UNetMotionModel,
 )
+from diffusers.models.attention import FreeNoiseTransformerBlock
 from diffusers.utils import is_xformers_available
 from diffusers.utils.testing_utils import torch_device
 
@@ -347,6 +348,64 @@ class AnimateDiffPAGPipelineFastTests(
                 "Enabling of FreeInit should lead to results different from the default pipeline results",
             )
 
+    def test_free_noise_blocks(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffPAGPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        pipe.enable_free_noise()
+        for block in pipe.unet.down_blocks:
+            for motion_module in block.motion_modules:
+                for transformer_block in motion_module.transformer_blocks:
+                    self.assertTrue(
+                        isinstance(transformer_block, FreeNoiseTransformerBlock),
+                        "Motion module transformer blocks must be an instance of `FreeNoiseTransformerBlock` after enabling FreeNoise.",
+                    )
+
+        pipe.disable_free_noise()
+        for block in pipe.unet.down_blocks:
+            for motion_module in block.motion_modules:
+                for transformer_block in motion_module.transformer_blocks:
+                    self.assertFalse(
+                        isinstance(transformer_block, FreeNoiseTransformerBlock),
+                        "Motion module transformer blocks must not be an instance of `FreeNoiseTransformerBlock` after disabling FreeNoise.",
+                    )
+
+    def test_free_noise(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffPAGPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        inputs_normal = self.get_dummy_inputs(torch_device)
+        frames_normal = pipe(**inputs_normal).frames[0]
+
+        for context_length in [8, 9]:
+            for context_stride in [4, 6]:
+                pipe.enable_free_noise(context_length, context_stride)
+
+                inputs_enable_free_noise = self.get_dummy_inputs(torch_device)
+                frames_enable_free_noise = pipe(**inputs_enable_free_noise).frames[0]
+
+                pipe.disable_free_noise()
+
+                inputs_disable_free_noise = self.get_dummy_inputs(torch_device)
+                frames_disable_free_noise = pipe(**inputs_disable_free_noise).frames[0]
+
+                sum_enabled = np.abs(to_np(frames_normal) - to_np(frames_enable_free_noise)).sum()
+                max_diff_disabled = np.abs(to_np(frames_normal) - to_np(frames_disable_free_noise)).max()
+                self.assertGreater(
+                    sum_enabled,
+                    1e1,
+                    "Enabling of FreeNoise should lead to results different from the default pipeline results",
+                )
+                self.assertLess(
+                    max_diff_disabled,
+                    1e-4,
+                    "Disabling of FreeNoise should lead to results similar to the default pipeline results",
+                )
+
     @unittest.skipIf(
         torch_device != "cuda" or not is_xformers_available(),
         reason="XFormers attention is only available with CUDA and `xformers` installed",

tests/pipelines/pag/test_pag_kolors.py

@@ -0,0 +1,252 @@
+# 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 inspect
+import unittest
+
+import numpy as np
+import torch
+
+from diffusers import (
+    AutoencoderKL,
+    EulerDiscreteScheduler,
+    KolorsPAGPipeline,
+    KolorsPipeline,
+    UNet2DConditionModel,
+)
+from diffusers.pipelines.kolors import ChatGLMModel, ChatGLMTokenizer
+from diffusers.utils.testing_utils import enable_full_determinism
+
+from ..pipeline_params import (
+    TEXT_TO_IMAGE_BATCH_PARAMS,
+    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
+    TEXT_TO_IMAGE_IMAGE_PARAMS,
+    TEXT_TO_IMAGE_PARAMS,
+)
+from ..test_pipelines_common import (
+    PipelineFromPipeTesterMixin,
+    PipelineTesterMixin,
+)
+
+
+enable_full_determinism()
+
+
+class KolorsPAGPipelineFastTests(
+    PipelineTesterMixin,
+    PipelineFromPipeTesterMixin,
+    unittest.TestCase,
+):
+    pipeline_class = KolorsPAGPipeline
+    params = TEXT_TO_IMAGE_PARAMS.union({"pag_scale", "pag_adaptive_scale"})
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union({"add_text_embeds", "add_time_ids"})
+
+    # Copied from tests.pipelines.kolors.test_kolors.KolorsPipelineFastTests.get_dummy_components
+    def get_dummy_components(self, time_cond_proj_dim=None):
+        torch.manual_seed(0)
+        unet = UNet2DConditionModel(
+            block_out_channels=(2, 4),
+            layers_per_block=2,
+            time_cond_proj_dim=time_cond_proj_dim,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            # specific config below
+            attention_head_dim=(2, 4),
+            use_linear_projection=True,
+            addition_embed_type="text_time",
+            addition_time_embed_dim=8,
+            transformer_layers_per_block=(1, 2),
+            projection_class_embeddings_input_dim=56,
+            cross_attention_dim=8,
+            norm_num_groups=1,
+        )
+        scheduler = EulerDiscreteScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            steps_offset=1,
+            beta_schedule="scaled_linear",
+            timestep_spacing="leading",
+        )
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+            sample_size=128,
+        )
+        torch.manual_seed(0)
+        text_encoder = ChatGLMModel.from_pretrained("hf-internal-testing/tiny-random-chatglm3-6b")
+        tokenizer = ChatGLMTokenizer.from_pretrained("hf-internal-testing/tiny-random-chatglm3-6b")
+
+        components = {
+            "unet": unet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "image_encoder": None,
+            "feature_extractor": None,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "pag_scale": 0.9,
+            "output_type": "np",
+        }
+        return inputs
+
+    def test_pag_disable_enable(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        # base  pipeline (expect same output when pag is disabled)
+        pipe_sd = KolorsPipeline(**components)
+        pipe_sd = pipe_sd.to(device)
+        pipe_sd.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        del inputs["pag_scale"]
+        assert (
+            "pag_scale" not in inspect.signature(pipe_sd.__call__).parameters
+        ), f"`pag_scale` should not be a call parameter of the base pipeline {pipe_sd.__class__.__name__}."
+        out = pipe_sd(**inputs).images[0, -3:, -3:, -1]
+
+        # pag disabled with pag_scale=0.0
+        pipe_pag = self.pipeline_class(**components)
+        pipe_pag = pipe_pag.to(device)
+        pipe_pag.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["pag_scale"] = 0.0
+        out_pag_disabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]
+
+        # pag enabled
+        pipe_pag = self.pipeline_class(**components, pag_applied_layers=["mid", "up", "down"])
+        pipe_pag = pipe_pag.to(device)
+        pipe_pag.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        out_pag_enabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]
+
+        assert np.abs(out.flatten() - out_pag_disabled.flatten()).max() < 1e-3
+        assert np.abs(out.flatten() - out_pag_enabled.flatten()).max() > 1e-3
+
+    def test_pag_applied_layers(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        # base pipeline
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        # pag_applied_layers = ["mid","up","down"] should apply to all self-attention layers
+        all_self_attn_layers = [k for k in pipe.unet.attn_processors.keys() if "attn1" in k]
+        original_attn_procs = pipe.unet.attn_processors
+        pag_layers = ["mid", "down", "up"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(all_self_attn_layers)
+
+        all_self_attn_mid_layers = [
+            "mid_block.attentions.0.transformer_blocks.0.attn1.processor",
+            "mid_block.attentions.0.transformer_blocks.1.attn1.processor",
+        ]
+        pipe.unet.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["mid"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(all_self_attn_mid_layers)
+
+        pipe.unet.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["mid_block"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(all_self_attn_mid_layers)
+
+        pipe.unet.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["mid_block.attentions.0"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(all_self_attn_mid_layers)
+
+        # pag_applied_layers = ["mid.block_0.attentions_1"] does not exist in the model
+        pipe.unet.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["mid_block.attentions.1"]
+        with self.assertRaises(ValueError):
+            pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+
+        # pag_applied_layers = "down" should apply to all self-attention layers in down_blocks
+        pipe.unet.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["down"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert len(pipe.pag_attn_processors) == 4
+
+        pipe.unet.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["down_blocks.0"]
+        with self.assertRaises(ValueError):
+            pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+
+        pipe.unet.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["down_blocks.1"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert len(pipe.pag_attn_processors) == 4
+
+        pipe.unet.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["down_blocks.1.attentions.1"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert len(pipe.pag_attn_processors) == 2
+
+    def test_pag_inference(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        pipe_pag = self.pipeline_class(**components, pag_applied_layers=["mid", "up", "down"])
+        pipe_pag = pipe_pag.to(device)
+        pipe_pag.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe_pag(**inputs).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        assert image.shape == (
+            1,
+            64,
+            64,
+            3,
+        ), f"the shape of the output image should be (1, 64, 64, 3) but got {image.shape}"
+        expected_slice = np.array(
+            [0.26030684, 0.43192005, 0.4042826, 0.4189067, 0.5181305, 0.3832534, 0.472135, 0.4145031, 0.43726248]
+        )
+
+        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
+        self.assertLessEqual(max_diff, 1e-3)
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(expected_max_diff=3e-3)

tests/pipelines/pag/test_pag_sd3.py

@@ -0,0 +1,295 @@
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
+
+from diffusers import (
+    AutoencoderKL,
+    FlowMatchEulerDiscreteScheduler,
+    SD3Transformer2DModel,
+    StableDiffusion3PAGPipeline,
+    StableDiffusion3Pipeline,
+)
+from diffusers.utils.testing_utils import (
+    torch_device,
+)
+
+from ..test_pipelines_common import (
+    PipelineTesterMixin,
+    check_qkv_fusion_matches_attn_procs_length,
+    check_qkv_fusion_processors_exist,
+)
+
+
+class StableDiffusion3PAGPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
+    pipeline_class = StableDiffusion3PAGPipeline
+    params = frozenset(
+        [
+            "prompt",
+            "height",
+            "width",
+            "guidance_scale",
+            "negative_prompt",
+            "prompt_embeds",
+            "negative_prompt_embeds",
+        ]
+    )
+    batch_params = frozenset(["prompt", "negative_prompt"])
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = SD3Transformer2DModel(
+            sample_size=32,
+            patch_size=1,
+            in_channels=4,
+            num_layers=2,
+            attention_head_dim=8,
+            num_attention_heads=4,
+            caption_projection_dim=32,
+            joint_attention_dim=32,
+            pooled_projection_dim=64,
+            out_channels=4,
+        )
+        clip_text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            hidden_act="gelu",
+            projection_dim=32,
+        )
+
+        torch.manual_seed(0)
+        text_encoder = CLIPTextModelWithProjection(clip_text_encoder_config)
+
+        torch.manual_seed(0)
+        text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)
+
+        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+        tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+        tokenizer_3 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            sample_size=32,
+            in_channels=3,
+            out_channels=3,
+            block_out_channels=(4,),
+            layers_per_block=1,
+            latent_channels=4,
+            norm_num_groups=1,
+            use_quant_conv=False,
+            use_post_quant_conv=False,
+            shift_factor=0.0609,
+            scaling_factor=1.5035,
+        )
+
+        scheduler = FlowMatchEulerDiscreteScheduler()
+
+        return {
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "text_encoder_2": text_encoder_2,
+            "text_encoder_3": text_encoder_3,
+            "tokenizer": tokenizer,
+            "tokenizer_2": tokenizer_2,
+            "tokenizer_3": tokenizer_3,
+            "transformer": transformer,
+            "vae": vae,
+        }
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device="cpu").manual_seed(seed)
+
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "output_type": "np",
+            "pag_scale": 0.0,
+        }
+        return inputs
+
+    def test_stable_diffusion_3_different_prompts(self):
+        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        output_same_prompt = pipe(**inputs).images[0]
+
+        inputs = self.get_dummy_inputs(torch_device)
+        inputs["prompt_2"] = "a different prompt"
+        inputs["prompt_3"] = "another different prompt"
+        output_different_prompts = pipe(**inputs).images[0]
+
+        max_diff = np.abs(output_same_prompt - output_different_prompts).max()
+
+        # Outputs should be different here
+        assert max_diff > 1e-2
+
+    def test_stable_diffusion_3_different_negative_prompts(self):
+        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        output_same_prompt = pipe(**inputs).images[0]
+
+        inputs = self.get_dummy_inputs(torch_device)
+        inputs["negative_prompt_2"] = "deformed"
+        inputs["negative_prompt_3"] = "blurry"
+        output_different_prompts = pipe(**inputs).images[0]
+
+        max_diff = np.abs(output_same_prompt - output_different_prompts).max()
+
+        # Outputs should be different here
+        assert max_diff > 1e-2
+
+    def test_stable_diffusion_3_prompt_embeds(self):
+        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)
+        inputs = self.get_dummy_inputs(torch_device)
+
+        output_with_prompt = pipe(**inputs).images[0]
+
+        inputs = self.get_dummy_inputs(torch_device)
+        prompt = inputs.pop("prompt")
+
+        do_classifier_free_guidance = inputs["guidance_scale"] > 1
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = pipe.encode_prompt(
+            prompt,
+            prompt_2=None,
+            prompt_3=None,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            device=torch_device,
+        )
+        output_with_embeds = pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            **inputs,
+        ).images[0]
+
+        max_diff = np.abs(output_with_prompt - output_with_embeds).max()
+        assert max_diff < 1e-4
+
+    def test_fused_qkv_projections(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        original_image_slice = image[0, -3:, -3:, -1]
+
+        # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
+        # to the pipeline level.
+        pipe.transformer.fuse_qkv_projections()
+        assert check_qkv_fusion_processors_exist(
+            pipe.transformer
+        ), "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        assert check_qkv_fusion_matches_attn_procs_length(
+            pipe.transformer, pipe.transformer.original_attn_processors
+        ), "Something wrong with the attention processors concerning the fused QKV projections."
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice_fused = image[0, -3:, -3:, -1]
+
+        pipe.transformer.unfuse_qkv_projections()
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice_disabled = image[0, -3:, -3:, -1]
+
+        assert np.allclose(
+            original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
+        ), "Fusion of QKV projections shouldn't affect the outputs."
+        assert np.allclose(
+            image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
+        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
+        assert np.allclose(
+            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
+        ), "Original outputs should match when fused QKV projections are disabled."
+
+    def test_pag_disable_enable(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        # base pipeline (expect same output when pag is disabled)
+        pipe_sd = StableDiffusion3Pipeline(**components)
+        pipe_sd = pipe_sd.to(device)
+        pipe_sd.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        del inputs["pag_scale"]
+        assert (
+            "pag_scale" not in inspect.signature(pipe_sd.__call__).parameters
+        ), f"`pag_scale` should not be a call parameter of the base pipeline {pipe_sd.__class__.__name__}."
+        out = pipe_sd(**inputs).images[0, -3:, -3:, -1]
+
+        components = self.get_dummy_components()
+
+        # pag disabled with pag_scale=0.0
+        pipe_pag = self.pipeline_class(**components)
+        pipe_pag = pipe_pag.to(device)
+        pipe_pag.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["pag_scale"] = 0.0
+        out_pag_disabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]
+
+        assert np.abs(out.flatten() - out_pag_disabled.flatten()).max() < 1e-3
+
+    def test_pag_applied_layers(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        # base pipeline
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        all_self_attn_layers = [k for k in pipe.transformer.attn_processors.keys() if "attn" in k]
+        original_attn_procs = pipe.transformer.attn_processors
+        pag_layers = ["blocks.0", "blocks.1"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(all_self_attn_layers)
+
+        # blocks.0
+        block_0_self_attn = ["transformer_blocks.0.attn.processor"]
+        pipe.transformer.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["blocks.0"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(block_0_self_attn)
+
+        pipe.transformer.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["blocks.0.attn"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(block_0_self_attn)
+
+        pipe.transformer.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["blocks.(0|1)"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert (len(pipe.pag_attn_processors)) == 2
+
+        pipe.transformer.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["blocks.0", r"blocks\.1"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert len(pipe.pag_attn_processors) == 2

tests/pipelines/pixart_sigma/test_pixart.py

@@ -36,7 +36,12 @@ from diffusers.utils.testing_utils import (
 )
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
-from ..test_pipelines_common import PipelineTesterMixin, to_np
+from ..test_pipelines_common import (
+    PipelineTesterMixin,
+    check_qkv_fusion_matches_attn_procs_length,
+    check_qkv_fusion_processors_exist,
+    to_np,
+)
 
 
 enable_full_determinism()
@@ -308,6 +313,46 @@ class PixArtSigmaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(expected_max_diff=1e-3)
 
+    def test_fused_qkv_projections(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        original_image_slice = image[0, -3:, -3:, -1]
+
+        # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
+        # to the pipeline level.
+        pipe.transformer.fuse_qkv_projections()
+        assert check_qkv_fusion_processors_exist(
+            pipe.transformer
+        ), "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        assert check_qkv_fusion_matches_attn_procs_length(
+            pipe.transformer, pipe.transformer.original_attn_processors
+        ), "Something wrong with the attention processors concerning the fused QKV projections."
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice_fused = image[0, -3:, -3:, -1]
+
+        pipe.transformer.unfuse_qkv_projections()
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice_disabled = image[0, -3:, -3:, -1]
+
+        assert np.allclose(
+            original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
+        ), "Fusion of QKV projections shouldn't affect the outputs."
+        assert np.allclose(
+            image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
+        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
+        assert np.allclose(
+            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
+        ), "Original outputs should match when fused QKV projections are disabled."
+
 
 @slow
 @require_torch_gpu

tests/pipelines/stable_diffusion_2/test_stable_diffusion_depth.py

@@ -26,8 +26,8 @@ from transformers import (
     CLIPTextModel,
     CLIPTokenizer,
     DPTConfig,
-    DPTFeatureExtractor,
     DPTForDepthEstimation,
+    DPTImageProcessor,
 )
 
 from diffusers import (
@@ -145,9 +145,7 @@ class StableDiffusionDepth2ImgPipelineFastTests(
             backbone_featmap_shape=[1, 384, 24, 24],
         )
         depth_estimator = DPTForDepthEstimation(depth_estimator_config).eval()
-        feature_extractor = DPTFeatureExtractor.from_pretrained(
-            "hf-internal-testing/tiny-random-DPTForDepthEstimation"
-        )
+        feature_extractor = DPTImageProcessor.from_pretrained("hf-internal-testing/tiny-random-DPTForDepthEstimation")
 
         components = {
             "unet": unet,

tests/pipelines/test_pipelines_common.py

@@ -26,6 +26,7 @@ from diffusers import (
     ConsistencyDecoderVAE,
     DDIMScheduler,
     DiffusionPipeline,
+    KolorsPipeline,
     StableDiffusionPipeline,
     StableDiffusionXLPipeline,
     UNet2DConditionModel,
@@ -656,6 +657,8 @@ class PipelineFromPipeTesterMixin:
     def original_pipeline_class(self):
         if "xl" in self.pipeline_class.__name__.lower():
             original_pipeline_class = StableDiffusionXLPipeline
+        elif "kolors" in self.pipeline_class.__name__.lower():
+            original_pipeline_class = KolorsPipeline
         else:
             original_pipeline_class = StableDiffusionPipeline
 
@@ -681,6 +684,9 @@ class PipelineFromPipeTesterMixin:
         elif self.original_pipeline_class == StableDiffusionXLPipeline:
             original_repo = "hf-internal-testing/tiny-stable-diffusion-xl-pipe"
             original_kwargs = {"requires_aesthetics_score": True, "force_zeros_for_empty_prompt": False}
+        elif self.original_pipeline_class == KolorsPipeline:
+            original_repo = "hf-internal-testing/tiny-kolors-pipe"
+            original_kwargs = {"force_zeros_for_empty_prompt": False}
         else:
             raise ValueError(
                 "original_pipeline_class must be either StableDiffusionPipeline or StableDiffusionXLPipeline"