update

* refactor context parallel cache; update torch compile time benchmark

* add tiling support

* make style

* remove num_frames % 8 == 0 requirement

* update default num_frames to original value

* add explanations + refactor

* update torch compile example

* update docs

* update

* clean up if-statements

* address review comments

* add test for vae tiling

* update docs

* update docs

* update docstrings

* add modeling test for cogvideox transformer

* make style

README.md

@@ -202,6 +202,7 @@ Also, say 👋 in our public Discord channel <a href="https://discord.gg/G7tWnz9
 
 - https://github.com/microsoft/TaskMatrix
 - https://github.com/invoke-ai/InvokeAI
+- https://github.com/InstantID/InstantID
 - https://github.com/apple/ml-stable-diffusion
 - https://github.com/Sanster/lama-cleaner
 - https://github.com/IDEA-Research/Grounded-Segment-Anything

docs/source/en/_toctree.yml

@@ -223,64 +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/stable_cascade_unet
-      title: StableCascadeUNet
-    - 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:
@@ -302,6 +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

@@ -22,6 +22,7 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:
 
 ## Supported pipelines
 
+- [`CogVideoXPipeline`]
 - [`StableDiffusionPipeline`]
 - [`StableDiffusionImg2ImgPipeline`]
 - [`StableDiffusionInpaintPipeline`]
@@ -49,6 +50,7 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:
 - [`UNet2DConditionModel`]
 - [`StableCascadeUNet`]
 - [`AutoencoderKL`]
+- [`AutoencoderKLCogVideoX`]
 - [`ControlNetModel`]
 - [`SD3Transformer2DModel`]
 - [`FluxTransformer2DModel`]

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

@@ -0,0 +1,37 @@
+<!--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. -->
+
+# AutoencoderKLCogVideoX
+
+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.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLCogVideoX
+
+vae = AutoencoderKLCogVideoX.from_pretrained("THUDM/CogVideoX-2b", subfolder="vae", torch_dtype=torch.float16).to("cuda")
+```
+
+## AutoencoderKLCogVideoX
+
+[[autodoc]] AutoencoderKLCogVideoX
+    - decode
+    - encode
+    - all
+
+## AutoencoderKLOutput
+
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

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

@@ -0,0 +1,30 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# CogVideoXTransformer3DModel
+
+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/pipelines/cogvideox.md

@@ -0,0 +1,88 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+-->
+
+# CogVideoX
+
+[CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://arxiv.org/abs/2408.06072) from Tsinghua University & ZhipuAI, by Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, Jie Tang.
+
+The abstract from the paper is:
+
+*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>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.md) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading.md#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+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.
+
+First, load the pipeline:
+
+```python
+import torch
+from diffusers import CogVideoXPipeline
+from diffusers.utils import export_to_video
+
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b").to("cuda")
+```
+
+Then change the memory layout of the pipelines `transformer` component to `torch.channels_last`:
+
+```python
+pipe.transformer.to(memory_format=torch.channels_last)
+```
+
+Finally, compile the components and run inference:
+
+```python
+pipe.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
+
+# CogVideoX works well with long and well-described prompts
+prompt = "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance."
+video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+```
+
+The [benchmark](https://gist.github.com/a-r-r-o-w/5183d75e452a368fd17448fcc810bd3f) results on an 80GB A100 machine are:
+
+```
+Without torch.compile(): Average inference time: 96.89 seconds.
+With torch.compile(): Average inference time: 76.27 seconds.
+```
+
+### Memory optimization
+
+CogVideoX requires about 19 GB of GPU memory to decode 49 frames (6 seconds of video at 8 FPS) with output resolution 720x480 (W x H), which makes it not possible to run on consumer GPUs or free-tier T4 Colab. The following memory optimizations could be used to reduce the memory footprint. For replication, you can refer to [this](https://gist.github.com/a-r-r-o-w/3959a03f15be5c9bd1fe545b09dfcc93) script.
+
+- `pipe.enable_model_cpu_offload()`:
+  - Without enabling cpu offloading, memory usage is `33 GB`
+  - With enabling cpu offloading, memory usage is `19 GB`
+- `pipe.vae.enable_tiling()`:
+  - With enabling cpu offloading and tiling, memory usage is `11 GB`
+- `pipe.vae.enable_slicing()`
+
+## CogVideoXPipeline
+
+[[autodoc]] CogVideoXPipeline
+  - all
+  - __call__
+
+## CogVideoXPipelineOutput
+
+[[autodoc]] pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput

docs/source/en/api/pipelines/controlnet_sd3.md

@@ -1,4 +1,4 @@
-<!--Copyright 2023 The HuggingFace Team and The InstantX Team. All rights reserved.
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
 the License. You may obtain a copy of the License at
@@ -22,7 +22,16 @@ The abstract from the paper is:
 
 *We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
 
-This code is implemented by [The InstantX Team](https://huggingface.co/InstantX). You can find pre-trained checkpoints for SD3-ControlNet on [The InstantX Team](https://huggingface.co/InstantX) Hub profile.
+This controlnet code is mainly implemented by [The InstantX Team](https://huggingface.co/InstantX). The inpainting-related code was developed by [The Alimama Creative Team](https://huggingface.co/alimama-creative). You can find pre-trained checkpoints for SD3-ControlNet in the table below: 
+
+
+| ControlNet type | Developer | Link |
+| -------- | ---------- | ---- |
+| Canny | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/InstantX/SD3-Controlnet-Canny) |
+| Pose | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/InstantX/SD3-Controlnet-Pose) |
+| Tile | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/InstantX/SD3-Controlnet-Tile) |
+| Inpainting | [The AlimamaCreative Team](https://huggingface.co/alimama-creative) | [link](https://huggingface.co/alimama-creative/SD3-Controlnet-Inpainting) |
+
 
 <Tip>
 
@@ -35,5 +44,10 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers)
 	- all
 	- __call__
 
+## StableDiffusion3ControlNetInpaintingPipeline
+[[autodoc]] pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet_inpainting.StableDiffusion3ControlNetInpaintingPipeline
+	- all
+	- __call__
+
 ## StableDiffusion3PipelineOutput
 [[autodoc]] pipelines.stable_diffusion_3.pipeline_output.StableDiffusion3PipelineOutput

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()
@@ -77,6 +77,34 @@ 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.
@@ -134,4 +162,4 @@ image.save("flux-fp8-dev.png")
 
 [[autodoc]] FluxPipeline
 	- all
-	- __call__
+	- __call__

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/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/distributed_inference.md

@@ -48,7 +48,7 @@ accelerate launch run_distributed.py --num_processes=2
 
 <Tip>
 
-To learn more, take a look at the [Distributed Inference with 🤗 Accelerate](https://huggingface.co/docs/accelerate/en/usage_guides/distributed_inference#distributed-inference-with-accelerate) guide.
+Refer to this minimal example [script](https://gist.github.com/sayakpaul/cfaebd221820d7b43fae638b4dfa01ba) for running inference across multiple GPUs. To learn more, take a look at the [Distributed Inference with 🤗 Accelerate](https://huggingface.co/docs/accelerate/en/usage_guides/distributed_inference#distributed-inference-with-accelerate) guide.
 
 </Tip>
 
@@ -108,4 +108,4 @@ torchrun run_distributed.py --nproc_per_node=2
 ```
 
 > [!TIP]
-> You can use `device_map` within a [`DiffusionPipeline`] to distribute its model-level components on multiple devices. Refer to the [Device placement](../tutorials/inference_with_big_models#device-placement) guide to learn more.
+> You can use `device_map` within a [`DiffusionPipeline`] to distribute its model-level components on multiple devices. Refer to the [Device placement](../tutorials/inference_with_big_models#device-placement) guide to learn more.

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

@@ -14,9 +14,9 @@ specific language governing permissions and limitations under the License.
 
 It can be fun and creative to use multiple [LoRAs]((https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora)) together to generate something entirely new and unique. This works by merging multiple LoRA weights together to produce images that are a blend of different styles. Diffusers provides a few methods to merge LoRAs depending on *how* you want to merge their weights, which can affect image quality.
 
-This guide will show you how to merge LoRAs using the [`~loaders.UNet2DConditionLoadersMixin.set_adapters`] and [`~peft.LoraModel.add_weighted_adapter`] methods. To improve inference speed and reduce memory-usage of merged LoRAs, you'll also see how to use the [`~loaders.StableDiffusionLoraLoaderMixin.fuse_lora`] method to fuse the LoRA weights with the original weights of the underlying model.
+This guide will show you how to merge LoRAs using the [`~loaders.PeftAdapterMixin.set_adapters`] and [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) methods. To improve inference speed and reduce memory-usage of merged LoRAs, you'll also see how to use the [`~loaders.StableDiffusionLoraLoaderMixin.fuse_lora`] method to fuse the LoRA weights with the original weights of the underlying model.
 
-For this guide, load a Stable Diffusion XL (SDXL) checkpoint and the [KappaNeuro/studio-ghibli-style]() and [Norod78/sdxl-chalkboarddrawing-lora]() LoRAs with the [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] method. You'll need to assign each LoRA an `adapter_name` to combine them later.
+For this guide, load a Stable Diffusion XL (SDXL) checkpoint and the [KappaNeuro/studio-ghibli-style](https://huggingface.co/KappaNeuro/studio-ghibli-style) and [Norod78/sdxl-chalkboarddrawing-lora](https://huggingface.co/Norod78/sdxl-chalkboarddrawing-lora) LoRAs with the [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] method. You'll need to assign each LoRA an `adapter_name` to combine them later.
 
 ```py
 from diffusers import DiffusionPipeline
@@ -29,7 +29,7 @@ pipeline.load_lora_weights("lordjia/by-feng-zikai", weight_name="fengzikai_v1.0_
 
 ## set_adapters
 
-The [`~loaders.UNet2DConditionLoadersMixin.set_adapters`] method merges LoRA adapters by concatenating their weighted matrices. Use the adapter name to specify which LoRAs to merge, and the `adapter_weights` parameter to control the scaling for each LoRA. For example, if `adapter_weights=[0.5, 0.5]`, then the merged LoRA output is an average of both LoRAs. Try adjusting the adapter weights to see how it affects the generated image!
+The [`~loaders.PeftAdapterMixin.set_adapters`] method merges LoRA adapters by concatenating their weighted matrices. Use the adapter name to specify which LoRAs to merge, and the `adapter_weights` parameter to control the scaling for each LoRA. For example, if `adapter_weights=[0.5, 0.5]`, then the merged LoRA output is an average of both LoRAs. Try adjusting the adapter weights to see how it affects the generated image!
 
 ```py
 pipeline.set_adapters(["ikea", "feng"], adapter_weights=[0.7, 0.8])
@@ -47,19 +47,19 @@ image
 ## add_weighted_adapter
 
 > [!WARNING]
-> This is an experimental method that adds PEFTs [`~peft.LoraModel.add_weighted_adapter`] method to Diffusers to enable more efficient merging methods. Check out this [issue](https://github.com/huggingface/diffusers/issues/6892) if you're interested in learning more about the motivation and design behind this integration.
+> This is an experimental method that adds PEFTs [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method to Diffusers to enable more efficient merging methods. Check out this [issue](https://github.com/huggingface/diffusers/issues/6892) if you're interested in learning more about the motivation and design behind this integration.
 
-The [`~peft.LoraModel.add_weighted_adapter`] method provides access to more efficient merging method such as [TIES and DARE](https://huggingface.co/docs/peft/developer_guides/model_merging). To use these merging methods, make sure you have the latest stable version of Diffusers and PEFT installed.
+The [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method provides access to more efficient merging method such as [TIES and DARE](https://huggingface.co/docs/peft/developer_guides/model_merging). To use these merging methods, make sure you have the latest stable version of Diffusers and PEFT installed.
 
 ```bash
 pip install -U diffusers peft
 ```
 
-There are three steps to merge LoRAs with the [`~peft.LoraModel.add_weighted_adapter`] method:
+There are three steps to merge LoRAs with the [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method:
 
-1. Create a [`~peft.PeftModel`] from the underlying model and LoRA checkpoint.
+1. Create a [PeftModel](https://huggingface.co/docs/peft/package_reference/peft_model#peft.PeftModel) from the underlying model and LoRA checkpoint.
 2. Load a base UNet model and the LoRA adapters.
-3. Merge the adapters using the [`~peft.LoraModel.add_weighted_adapter`] method and the merging method of your choice.
+3. Merge the adapters using the [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method and the merging method of your choice.
 
 Let's dive deeper into what these steps entail.
 
@@ -92,7 +92,7 @@ pipeline = DiffusionPipeline.from_pretrained(
 pipeline.load_lora_weights("ostris/ikea-instructions-lora-sdxl", weight_name="ikea_instructions_xl_v1_5.safetensors", adapter_name="ikea")
 ```
 
-Now you'll create a [`~peft.PeftModel`] from the loaded LoRA checkpoint by combining the SDXL UNet and the LoRA UNet from the pipeline.
+Now you'll create a [PeftModel](https://huggingface.co/docs/peft/package_reference/peft_model#peft.PeftModel) from the loaded LoRA checkpoint by combining the SDXL UNet and the LoRA UNet from the pipeline.
 
 ```python
 from peft import get_peft_model, LoraConfig
@@ -112,7 +112,7 @@ ikea_peft_model.load_state_dict(original_state_dict, strict=True)
 > [!TIP]
 > You can optionally push the ikea_peft_model to the Hub by calling `ikea_peft_model.push_to_hub("ikea_peft_model", token=TOKEN)`.
 
-Repeat this process to create a [`~peft.PeftModel`] from the [lordjia/by-feng-zikai](https://huggingface.co/lordjia/by-feng-zikai) LoRA.
+Repeat this process to create a [PeftModel](https://huggingface.co/docs/peft/package_reference/peft_model#peft.PeftModel) from the [lordjia/by-feng-zikai](https://huggingface.co/lordjia/by-feng-zikai) LoRA.
 
 ```python
 pipeline.delete_adapters("ikea")
@@ -148,7 +148,7 @@ model = PeftModel.from_pretrained(base_unet, "stevhliu/ikea_peft_model", use_saf
 model.load_adapter("stevhliu/feng_peft_model", use_safetensors=True, subfolder="feng", adapter_name="feng")
 ```
 
-3. Merge the adapters using the [`~peft.LoraModel.add_weighted_adapter`] method and the merging method of your choice (learn more about other merging methods in this [blog post](https://huggingface.co/blog/peft_merging)). For this example, let's use the `"dare_linear"` method to merge the LoRAs.
+3. Merge the adapters using the [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method and the merging method of your choice (learn more about other merging methods in this [blog post](https://huggingface.co/blog/peft_merging)). For this example, let's use the `"dare_linear"` method to merge the LoRAs.
 
 > [!WARNING]
 > Keep in mind the LoRAs need to have the same rank to be merged!
@@ -182,9 +182,9 @@ image
 
 ## fuse_lora
 
-Both the [`~loaders.UNet2DConditionLoadersMixin.set_adapters`] and [`~peft.LoraModel.add_weighted_adapter`] methods require loading the base model and the LoRA adapters separately which incurs some overhead. The [`~loaders.StableDiffusionLoraLoaderMixin.fuse_lora`] method allows you to fuse the LoRA weights directly with the original weights of the underlying model. This way, you're only loading the model once which can increase inference and lower memory-usage.
+Both the [`~loaders.PeftAdapterMixin.set_adapters`] and [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) methods require loading the base model and the LoRA adapters separately which incurs some overhead. The [`~loaders.lora_base.LoraBaseMixin.fuse_lora`] method allows you to fuse the LoRA weights directly with the original weights of the underlying model. This way, you're only loading the model once which can increase inference and lower memory-usage.
 
-You can use PEFT to easily fuse/unfuse multiple adapters directly into the model weights (both UNet and text encoder) using the [`~loaders.StableDiffusionLoraLoaderMixin.fuse_lora`] method, which can lead to a speed-up in inference and lower VRAM usage.
+You can use PEFT to easily fuse/unfuse multiple adapters directly into the model weights (both UNet and text encoder) using the [`~loaders.lora_base.LoraBaseMixin.fuse_lora`] method, which can lead to a speed-up in inference and lower VRAM usage.
 
 For example, if you have a base model and adapters loaded and set as active with the following adapter weights:
 
@@ -199,7 +199,7 @@ pipeline.load_lora_weights("lordjia/by-feng-zikai", weight_name="fengzikai_v1.0_
 pipeline.set_adapters(["ikea", "feng"], adapter_weights=[0.7, 0.8])
 ```
 
-Fuse these LoRAs into the UNet with the [`~loaders.StableDiffusionLoraLoaderMixin.fuse_lora`] method. The `lora_scale` parameter controls how much to scale the output by with the LoRA weights. It is important to make the `lora_scale` adjustments in the [`~loaders.StableDiffusionLoraLoaderMixin.fuse_lora`] method because it won’t work if you try to pass `scale` to the `cross_attention_kwargs` in the pipeline.
+Fuse these LoRAs into the UNet with the [`~loaders.lora_base.LoraBaseMixin.fuse_lora`] method. The `lora_scale` parameter controls how much to scale the output by with the LoRA weights. It is important to make the `lora_scale` adjustments in the [`~loaders.lora_base.LoraBaseMixin.fuse_lora`] method because it won’t work if you try to pass `scale` to the `cross_attention_kwargs` in the pipeline.
 
 ```py
 pipeline.fuse_lora(adapter_names=["ikea", "feng"], lora_scale=1.0)
@@ -226,7 +226,7 @@ image = pipeline("A bowl of ramen shaped like a cute kawaii bear, by Feng Zikai"
 image
 ```
 
-You can call [`~loaders.StableDiffusionLoraLoaderMixin.unfuse_lora`] to restore the original model's weights (for example, if you want to use a different `lora_scale` value). However, this only works if you've only fused one LoRA adapter to the original model. If you've fused multiple LoRAs, you'll need to reload the model.
+You can call [`~~loaders.lora_base.LoraBaseMixin.unfuse_lora`] to restore the original model's weights (for example, if you want to use a different `lora_scale` value). However, this only works if you've only fused one LoRA adapter to the original model. If you've fused multiple LoRAs, you'll need to reload the model.
 
 ```py
 pipeline.unfuse_lora()

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.
 
@@ -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')
@@ -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/dreambooth/README_flux.md

@@ -0,0 +1,195 @@
+# DreamBooth training example for FLUX.1 [dev]
+
+[DreamBooth](https://arxiv.org/abs/2208.12242) is a method to personalize text2image models like stable diffusion given just a few (3~5) images of a subject.
+
+The `train_dreambooth_flux.py` script shows how to implement the training procedure and adapt it for [FLUX.1 [dev]](https://blackforestlabs.ai/announcing-black-forest-labs/). We also provide a LoRA implementation in the `train_dreambooth_lora_flux.py` script.
+> [!NOTE] 
+> **Memory consumption**
+> 
+> Flux can be quite expensive to run on consumer hardware devices and as a result finetuning it comes with high memory requirements - 
+> a LoRA with a rank of 16 (w/ all components trained) can exceed 40GB of VRAM for training.
+> For more tips & guidance on training on a resource-constrained device please visit [`@bghira`'s guide](https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md) 
+
+
+> [!NOTE]
+> **Gated model**
+> 
+> As the model is gated, before using it with diffusers you first need to go to the [FLUX.1 [dev] Hugging Face page](https://huggingface.co/black-forest-labs/FLUX.1-dev), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
+
+```bash
+huggingface-cli login
+```
+
+This will also allow us to push the trained model parameters to the Hugging Face Hub platform.
+
+## Running locally with PyTorch
+
+### Installing the dependencies
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+**Important**
+
+To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
+
+```bash
+git clone https://github.com/huggingface/diffusers
+cd diffusers
+pip install -e .
+```
+
+Then cd in the `examples/dreambooth` folder and run
+```bash
+pip install -r requirements_flux.txt
+```
+
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+Or for a default accelerate configuration without answering questions about your environment
+
+```bash
+accelerate config default
+```
+
+Or if your environment doesn't support an interactive shell (e.g., a notebook)
+
+```python
+from accelerate.utils import write_basic_config
+write_basic_config()
+```
+
+When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
+
+
+### Dog toy example
+
+Now let's get our dataset. For this example we will use some dog images: https://huggingface.co/datasets/diffusers/dog-example.
+
+Let's first download it locally:
+
+```python
+from huggingface_hub import snapshot_download
+
+local_dir = "./dog"
+snapshot_download(
+    "diffusers/dog-example",
+    local_dir=local_dir, repo_type="dataset",
+    ignore_patterns=".gitattributes",
+)
+```
+
+This will also allow us to push the trained LoRA parameters to the Hugging Face Hub platform.
+
+Now, we can launch training using:
+
+```bash
+export MODEL_NAME="black-forest-labs/FLUX.1-dev"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="trained-flux"
+
+accelerate launch train_dreambooth_flux.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --mixed_precision="bf16" \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=500 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --validation_epochs=25 \
+  --seed="0" \
+  --push_to_hub
+```
+
+To better track our training experiments, we're using the following flags in the command above:
+
+* `report_to="wandb` will ensure the training runs are tracked on Weights and Biases. To use it, be sure to install `wandb` with `pip install wandb`.
+* `validation_prompt` and `validation_epochs` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
+
+> [!NOTE]
+> If you want to train using long prompts with the T5 text encoder, you can use `--max_sequence_length` to set the token limit. The default is 77, but it can be increased to as high as 512. Note that this will use more resources and may slow down the training in some cases.
+
+> [!TIP]
+> You can pass `--use_8bit_adam` to reduce the memory requirements of training. Make sure to install `bitsandbytes` if you want to do so.
+
+## LoRA + DreamBooth
+
+[LoRA](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) is a popular parameter-efficient fine-tuning technique that allows you to achieve full-finetuning like performance but with a fraction of learnable parameters.
+
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
+
+To perform DreamBooth with LoRA, run:
+
+```bash
+export MODEL_NAME="black-forest-labs/FLUX.1-dev"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="trained-flux-lora"
+
+accelerate launch train_dreambooth_lora_flux.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --mixed_precision="bf16" \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --learning_rate=1e-5 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=500 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --validation_epochs=25 \
+  --seed="0" \
+  --push_to_hub
+```
+
+### Text Encoder Training
+
+Alongside the transformer, fine-tuning of the CLIP text encoder is also supported.
+To do so, just specify `--train_text_encoder` while launching training. Please keep the following points in mind:
+
+> [!NOTE]
+> FLUX.1 has 2 text encoders (CLIP L/14 and T5-v1.1-XXL).
+By enabling `--train_text_encoder`, fine-tuning of the **CLIP encoder** is performed. 
+> At the moment, T5 fine-tuning is not supported and weights remain frozen when text encoder training is enabled.
+
+To perform DreamBooth LoRA with text-encoder training, run:
+```bash
+export MODEL_NAME="black-forest-labs/FLUX.1-dev"
+export OUTPUT_DIR="trained-flux-dev-dreambooth-lora"
+
+accelerate launch train_dreambooth_lora_flux.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --mixed_precision="bf16" \
+  --train_text_encoder\
+  --instance_prompt="a photo of sks dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --learning_rate=1e-5 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=500 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --seed="0" \
+  --push_to_hub
+```
+
+## Other notes
+Thanks to `bghira` for their help with reviewing & insight sharing ♥️

examples/dreambooth/requirements_flux.txt

@@ -0,0 +1,8 @@
+accelerate>=0.31.0
+torchvision
+transformers>=4.41.2
+ftfy
+tensorboard
+Jinja2
+peft>=0.11.1
+sentencepiece

examples/dreambooth/test_dreambooth_flux.py

@@ -0,0 +1,203 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import os
+import shutil
+import sys
+import tempfile
+
+from diffusers import DiffusionPipeline, FluxTransformer2DModel
+
+
+sys.path.append("..")
+from test_examples_utils import ExamplesTestsAccelerate, run_command  # noqa: E402
+
+
+logging.basicConfig(level=logging.DEBUG)
+
+logger = logging.getLogger()
+stream_handler = logging.StreamHandler(sys.stdout)
+logger.addHandler(stream_handler)
+
+
+class DreamBoothFlux(ExamplesTestsAccelerate):
+    instance_data_dir = "docs/source/en/imgs"
+    instance_prompt = "photo"
+    pretrained_model_name_or_path = "hf-internal-testing/tiny-flux-pipe"
+    script_path = "examples/dreambooth/train_dreambooth_flux.py"
+
+    def test_dreambooth(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "transformer", "diffusion_pytorch_model.safetensors")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "scheduler", "scheduler_config.json")))
+
+    def test_dreambooth_checkpointing(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            # Run training script with checkpointing
+            # max_train_steps == 4, checkpointing_steps == 2
+            # Should create checkpoints at steps 2, 4
+
+            initial_run_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 4
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                --checkpointing_steps=2
+                --seed=0
+                """.split()
+
+            run_command(self._launch_args + initial_run_args)
+
+            # check can run the original fully trained output pipeline
+            pipe = DiffusionPipeline.from_pretrained(tmpdir)
+            pipe(self.instance_prompt, num_inference_steps=1)
+
+            # check checkpoint directories exist
+            self.assertTrue(os.path.isdir(os.path.join(tmpdir, "checkpoint-2")))
+            self.assertTrue(os.path.isdir(os.path.join(tmpdir, "checkpoint-4")))
+
+            # check can run an intermediate checkpoint
+            transformer = FluxTransformer2DModel.from_pretrained(tmpdir, subfolder="checkpoint-2/transformer")
+            pipe = DiffusionPipeline.from_pretrained(self.pretrained_model_name_or_path, transformer=transformer)
+            pipe(self.instance_prompt, num_inference_steps=1)
+
+            # Remove checkpoint 2 so that we can check only later checkpoints exist after resuming
+            shutil.rmtree(os.path.join(tmpdir, "checkpoint-2"))
+
+            # Run training script for 7 total steps resuming from checkpoint 4
+
+            resume_run_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 6
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                --checkpointing_steps=2
+                --resume_from_checkpoint=checkpoint-4
+                --seed=0
+                """.split()
+
+            run_command(self._launch_args + resume_run_args)
+
+            # check can run new fully trained pipeline
+            pipe = DiffusionPipeline.from_pretrained(tmpdir)
+            pipe(self.instance_prompt, num_inference_steps=1)
+
+            # check old checkpoints do not exist
+            self.assertFalse(os.path.isdir(os.path.join(tmpdir, "checkpoint-2")))
+
+            # check new checkpoints exist
+            self.assertTrue(os.path.isdir(os.path.join(tmpdir, "checkpoint-4")))
+            self.assertTrue(os.path.isdir(os.path.join(tmpdir, "checkpoint-6")))
+
+    def test_dreambooth_checkpointing_checkpoints_total_limit(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=6
+            --checkpoints_total_limit=2
+            --checkpointing_steps=2
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual(
+                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
+                {"checkpoint-4", "checkpoint-6"},
+            )
+
+    def test_dreambooth_checkpointing_checkpoints_total_limit_removes_multiple_checkpoints(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=4
+            --checkpointing_steps=2
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual(
+                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
+                {"checkpoint-2", "checkpoint-4"},
+            )
+
+            resume_run_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=8
+            --checkpointing_steps=2
+            --resume_from_checkpoint=checkpoint-4
+            --checkpoints_total_limit=2
+            """.split()
+
+            run_command(self._launch_args + resume_run_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-6", "checkpoint-8"})

examples/dreambooth/test_dreambooth_lora_flux.py

@@ -0,0 +1,165 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import os
+import sys
+import tempfile
+
+import safetensors
+
+
+sys.path.append("..")
+from test_examples_utils import ExamplesTestsAccelerate, run_command  # noqa: E402
+
+
+logging.basicConfig(level=logging.DEBUG)
+
+logger = logging.getLogger()
+stream_handler = logging.StreamHandler(sys.stdout)
+logger.addHandler(stream_handler)
+
+
+class DreamBoothLoRAFlux(ExamplesTestsAccelerate):
+    instance_data_dir = "docs/source/en/imgs"
+    instance_prompt = "photo"
+    pretrained_model_name_or_path = "hf-internal-testing/tiny-flux-pipe"
+    script_path = "examples/dreambooth/train_dreambooth_lora_flux.py"
+
+    def test_dreambooth_lora_flux(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            # when not training the text encoder, all the parameters in the state dict should start
+            # with `"transformer"` in their names.
+            starts_with_transformer = all(key.startswith("transformer") for key in lora_state_dict.keys())
+            self.assertTrue(starts_with_transformer)
+
+    def test_dreambooth_lora_text_encoder_flux(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --train_text_encoder
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            starts_with_expected_prefix = all(
+                (key.startswith("transformer") or key.startswith("text_encoder")) for key in lora_state_dict.keys()
+            )
+            self.assertTrue(starts_with_expected_prefix)
+
+    def test_dreambooth_lora_flux_checkpointing_checkpoints_total_limit(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=6
+            --checkpoints_total_limit=2
+            --checkpointing_steps=2
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual(
+                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
+                {"checkpoint-4", "checkpoint-6"},
+            )
+
+    def test_dreambooth_lora_flux_checkpointing_checkpoints_total_limit_removes_multiple_checkpoints(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=4
+            --checkpointing_steps=2
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-2", "checkpoint-4"})
+
+            resume_run_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=8
+            --checkpointing_steps=2
+            --resume_from_checkpoint=checkpoint-4
+            --checkpoints_total_limit=2
+            """.split()
+
+            run_command(self._launch_args + resume_run_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-6", "checkpoint-8"})

examples/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)
 

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -478,7 +478,7 @@ def parse_args(input_args=None):
     parser.add_argument(
         "--debug_loss",
         action="store_true",
-        help="debug loss for each image, if filenames are awailable in the dataset",
+        help="debug loss for each image, if filenames are available in the dataset",
     )
 
     if input_args is not None:

examples/textual_inversion/README_sdxl.md

@@ -23,4 +23,25 @@ accelerate launch textual_inversion_sdxl.py \
   --output_dir="./textual_inversion_cat_sdxl"
 ```
 
-For now, only training of the first text encoder is supported.
+Training of both text encoders is supported.
+
+### Inference Example
+
+Once you have trained a model using above command, the inference can be done simply using the `StableDiffusionXLPipeline`.
+Make sure to include the `placeholder_token` in your prompt.
+
+```python
+from diffusers import StableDiffusionXLPipeline
+import torch
+
+model_id = "./textual_inversion_cat_sdxl"
+pipe = StableDiffusionXLPipeline.from_pretrained(model_id,torch_dtype=torch.float16).to("cuda")
+
+prompt = "A <cat-toy> backpack"
+
+image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
+image.save("cat-backpack.png")
+
+image = pipe(prompt="", prompt_2=prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
+image.save("cat-backpack-prompt_2.png")
+```

examples/textual_inversion/textual_inversion_sdxl.py

@@ -135,7 +135,7 @@ def log_validation(
     pipeline = DiffusionPipeline.from_pretrained(
         args.pretrained_model_name_or_path,
         text_encoder=accelerator.unwrap_model(text_encoder_1),
-        text_encoder_2=text_encoder_2,
+        text_encoder_2=accelerator.unwrap_model(text_encoder_2),
         tokenizer=tokenizer_1,
         tokenizer_2=tokenizer_2,
         unet=unet,
@@ -678,36 +678,54 @@ def main():
             f"The tokenizer already contains the token {args.placeholder_token}. Please pass a different"
             " `placeholder_token` that is not already in the tokenizer."
         )
+    num_added_tokens = tokenizer_2.add_tokens(placeholder_tokens)
+    if num_added_tokens != args.num_vectors:
+        raise ValueError(
+            f"The 2nd tokenizer already contains the token {args.placeholder_token}. Please pass a different"
+            " `placeholder_token` that is not already in the tokenizer."
+        )
 
     # Convert the initializer_token, placeholder_token to ids
     token_ids = tokenizer_1.encode(args.initializer_token, add_special_tokens=False)
+    token_ids_2 = tokenizer_2.encode(args.initializer_token, add_special_tokens=False)
+
     # Check if initializer_token is a single token or a sequence of tokens
-    if len(token_ids) > 1:
+    if len(token_ids) > 1 or len(token_ids_2) > 1:
         raise ValueError("The initializer token must be a single token.")
 
     initializer_token_id = token_ids[0]
     placeholder_token_ids = tokenizer_1.convert_tokens_to_ids(placeholder_tokens)
+    initializer_token_id_2 = token_ids_2[0]
+    placeholder_token_ids_2 = tokenizer_2.convert_tokens_to_ids(placeholder_tokens)
 
     # Resize the token embeddings as we are adding new special tokens to the tokenizer
     text_encoder_1.resize_token_embeddings(len(tokenizer_1))
+    text_encoder_2.resize_token_embeddings(len(tokenizer_2))
 
     # Initialise the newly added placeholder token with the embeddings of the initializer token
     token_embeds = text_encoder_1.get_input_embeddings().weight.data
+    token_embeds_2 = text_encoder_2.get_input_embeddings().weight.data
     with torch.no_grad():
         for token_id in placeholder_token_ids:
             token_embeds[token_id] = token_embeds[initializer_token_id].clone()
+        for token_id in placeholder_token_ids_2:
+            token_embeds_2[token_id] = token_embeds_2[initializer_token_id_2].clone()
 
     # Freeze vae and unet
     vae.requires_grad_(False)
     unet.requires_grad_(False)
-    text_encoder_2.requires_grad_(False)
+
     # Freeze all parameters except for the token embeddings in text encoder
     text_encoder_1.text_model.encoder.requires_grad_(False)
     text_encoder_1.text_model.final_layer_norm.requires_grad_(False)
     text_encoder_1.text_model.embeddings.position_embedding.requires_grad_(False)
+    text_encoder_2.text_model.encoder.requires_grad_(False)
+    text_encoder_2.text_model.final_layer_norm.requires_grad_(False)
+    text_encoder_2.text_model.embeddings.position_embedding.requires_grad_(False)
 
     if args.gradient_checkpointing:
         text_encoder_1.gradient_checkpointing_enable()
+        text_encoder_2.gradient_checkpointing_enable()
 
     if args.enable_xformers_memory_efficient_attention:
         if is_xformers_available():
@@ -746,7 +764,11 @@ def main():
         optimizer_class = torch.optim.AdamW
 
     optimizer = optimizer_class(
-        text_encoder_1.get_input_embeddings().parameters(),  # only optimize the embeddings
+        # only optimize the embeddings
+        [
+            text_encoder_1.text_model.embeddings.token_embedding.weight,
+            text_encoder_2.text_model.embeddings.token_embedding.weight,
+        ],
         lr=args.learning_rate,
         betas=(args.adam_beta1, args.adam_beta2),
         weight_decay=args.adam_weight_decay,
@@ -786,9 +808,10 @@ def main():
     )
 
     text_encoder_1.train()
+    text_encoder_2.train()
     # Prepare everything with our `accelerator`.
-    text_encoder_1, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        text_encoder_1, optimizer, train_dataloader, lr_scheduler
+    text_encoder_1, text_encoder_2, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        text_encoder_1, text_encoder_2, optimizer, train_dataloader, lr_scheduler
     )
 
     # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision
@@ -866,11 +889,13 @@ def main():
 
     # keep original embeddings as reference
     orig_embeds_params = accelerator.unwrap_model(text_encoder_1).get_input_embeddings().weight.data.clone()
+    orig_embeds_params_2 = accelerator.unwrap_model(text_encoder_2).get_input_embeddings().weight.data.clone()
 
     for epoch in range(first_epoch, args.num_train_epochs):
         text_encoder_1.train()
+        text_encoder_2.train()
         for step, batch in enumerate(train_dataloader):
-            with accelerator.accumulate(text_encoder_1):
+            with accelerator.accumulate([text_encoder_1, text_encoder_2]):
                 # Convert images to latent space
                 latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample().detach()
                 latents = latents * vae.config.scaling_factor
@@ -892,9 +917,7 @@ def main():
                     .hidden_states[-2]
                     .to(dtype=weight_dtype)
                 )
-                encoder_output_2 = text_encoder_2(
-                    batch["input_ids_2"].reshape(batch["input_ids_1"].shape[0], -1), output_hidden_states=True
-                )
+                encoder_output_2 = text_encoder_2(batch["input_ids_2"], output_hidden_states=True)
                 encoder_hidden_states_2 = encoder_output_2.hidden_states[-2].to(dtype=weight_dtype)
                 original_size = [
                     (batch["original_size"][0][i].item(), batch["original_size"][1][i].item())
@@ -938,11 +961,16 @@ def main():
                 # Let's make sure we don't update any embedding weights besides the newly added token
                 index_no_updates = torch.ones((len(tokenizer_1),), dtype=torch.bool)
                 index_no_updates[min(placeholder_token_ids) : max(placeholder_token_ids) + 1] = False
+                index_no_updates_2 = torch.ones((len(tokenizer_2),), dtype=torch.bool)
+                index_no_updates_2[min(placeholder_token_ids_2) : max(placeholder_token_ids_2) + 1] = False
 
                 with torch.no_grad():
                     accelerator.unwrap_model(text_encoder_1).get_input_embeddings().weight[
                         index_no_updates
                     ] = orig_embeds_params[index_no_updates]
+                    accelerator.unwrap_model(text_encoder_2).get_input_embeddings().weight[
+                        index_no_updates_2
+                    ] = orig_embeds_params_2[index_no_updates_2]
 
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
@@ -960,6 +988,16 @@ def main():
                         save_path,
                         safe_serialization=True,
                     )
+                    weight_name = f"learned_embeds_2-steps-{global_step}.safetensors"
+                    save_path = os.path.join(args.output_dir, weight_name)
+                    save_progress(
+                        text_encoder_2,
+                        placeholder_token_ids_2,
+                        accelerator,
+                        args,
+                        save_path,
+                        safe_serialization=True,
+                    )
 
                 if accelerator.is_main_process:
                     if global_step % args.checkpointing_steps == 0:
@@ -1034,7 +1072,7 @@ def main():
             pipeline = DiffusionPipeline.from_pretrained(
                 args.pretrained_model_name_or_path,
                 text_encoder=accelerator.unwrap_model(text_encoder_1),
-                text_encoder_2=text_encoder_2,
+                text_encoder_2=accelerator.unwrap_model(text_encoder_2),
                 vae=vae,
                 unet=unet,
                 tokenizer=tokenizer_1,
@@ -1052,6 +1090,16 @@ def main():
             save_path,
             safe_serialization=True,
         )
+        weight_name = "learned_embeds_2.safetensors"
+        save_path = os.path.join(args.output_dir, weight_name)
+        save_progress(
+            text_encoder_2,
+            placeholder_token_ids_2,
+            accelerator,
+            args,
+            save_path,
+            safe_serialization=True,
+        )
 
         if args.push_to_hub:
             save_model_card(

scripts/convert_cogvideox_to_diffusers.py

@@ -0,0 +1,222 @@
+import argparse
+from typing import Any, Dict
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers import AutoencoderKLCogVideoX, CogVideoXDDIMScheduler, CogVideoXPipeline, CogVideoXTransformer3DModel
+
+
+def reassign_query_key_value_inplace(key: str, state_dict: Dict[str, Any]):
+    to_q_key = key.replace("query_key_value", "to_q")
+    to_k_key = key.replace("query_key_value", "to_k")
+    to_v_key = key.replace("query_key_value", "to_v")
+    to_q, to_k, to_v = torch.chunk(state_dict[key], chunks=3, dim=0)
+    state_dict[to_q_key] = to_q
+    state_dict[to_k_key] = to_k
+    state_dict[to_v_key] = to_v
+    state_dict.pop(key)
+
+
+def reassign_query_key_layernorm_inplace(key: str, state_dict: Dict[str, Any]):
+    layer_id, weight_or_bias = key.split(".")[-2:]
+
+    if "query" in key:
+        new_key = f"transformer_blocks.{layer_id}.attn1.norm_q.{weight_or_bias}"
+    elif "key" in key:
+        new_key = f"transformer_blocks.{layer_id}.attn1.norm_k.{weight_or_bias}"
+
+    state_dict[new_key] = state_dict.pop(key)
+
+
+def reassign_adaln_norm_inplace(key: str, state_dict: Dict[str, Any]):
+    layer_id, _, weight_or_bias = key.split(".")[-3:]
+
+    weights_or_biases = state_dict[key].chunk(12, dim=0)
+    norm1_weights_or_biases = torch.cat(weights_or_biases[0:3] + weights_or_biases[6:9])
+    norm2_weights_or_biases = torch.cat(weights_or_biases[3:6] + weights_or_biases[9:12])
+
+    norm1_key = f"transformer_blocks.{layer_id}.norm1.linear.{weight_or_bias}"
+    state_dict[norm1_key] = norm1_weights_or_biases
+
+    norm2_key = f"transformer_blocks.{layer_id}.norm2.linear.{weight_or_bias}"
+    state_dict[norm2_key] = norm2_weights_or_biases
+
+    state_dict.pop(key)
+
+
+def remove_keys_inplace(key: str, state_dict: Dict[str, Any]):
+    state_dict.pop(key)
+
+
+def replace_up_keys_inplace(key: str, state_dict: Dict[str, Any]):
+    key_split = key.split(".")
+    layer_index = int(key_split[2])
+    replace_layer_index = 4 - 1 - layer_index
+
+    key_split[1] = "up_blocks"
+    key_split[2] = str(replace_layer_index)
+    new_key = ".".join(key_split)
+
+    state_dict[new_key] = state_dict.pop(key)
+
+
+TRANSFORMER_KEYS_RENAME_DICT = {
+    "transformer.final_layernorm": "norm_final",
+    "transformer": "transformer_blocks",
+    "attention": "attn1",
+    "mlp": "ff.net",
+    "dense_h_to_4h": "0.proj",
+    "dense_4h_to_h": "2",
+    ".layers": "",
+    "dense": "to_out.0",
+    "input_layernorm": "norm1.norm",
+    "post_attn1_layernorm": "norm2.norm",
+    "time_embed.0": "time_embedding.linear_1",
+    "time_embed.2": "time_embedding.linear_2",
+    "mixins.patch_embed": "patch_embed",
+    "mixins.final_layer.norm_final": "norm_out.norm",
+    "mixins.final_layer.linear": "proj_out",
+    "mixins.final_layer.adaLN_modulation.1": "norm_out.linear",
+}
+
+TRANSFORMER_SPECIAL_KEYS_REMAP = {
+    "query_key_value": reassign_query_key_value_inplace,
+    "query_layernorm_list": reassign_query_key_layernorm_inplace,
+    "key_layernorm_list": reassign_query_key_layernorm_inplace,
+    "adaln_layer.adaLN_modulations": reassign_adaln_norm_inplace,
+    "embed_tokens": remove_keys_inplace,
+}
+
+VAE_KEYS_RENAME_DICT = {
+    "block.": "resnets.",
+    "down.": "down_blocks.",
+    "downsample": "downsamplers.0",
+    "upsample": "upsamplers.0",
+    "nin_shortcut": "conv_shortcut",
+    "encoder.mid.block_1": "encoder.mid_block.resnets.0",
+    "encoder.mid.block_2": "encoder.mid_block.resnets.1",
+    "decoder.mid.block_1": "decoder.mid_block.resnets.0",
+    "decoder.mid.block_2": "decoder.mid_block.resnets.1",
+}
+
+VAE_SPECIAL_KEYS_REMAP = {
+    "loss": remove_keys_inplace,
+    "up.": replace_up_keys_inplace,
+}
+
+TOKENIZER_MAX_LENGTH = 226
+
+
+def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
+    state_dict = saved_dict
+    if "model" in saved_dict.keys():
+        state_dict = state_dict["model"]
+    if "module" in saved_dict.keys():
+        state_dict = state_dict["module"]
+    if "state_dict" in saved_dict.keys():
+        state_dict = state_dict["state_dict"]
+    return state_dict
+
+
+def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
+    state_dict[new_key] = state_dict.pop(old_key)
+
+
+def convert_transformer(ckpt_path: str):
+    PREFIX_KEY = "model.diffusion_model."
+
+    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
+    transformer = CogVideoXTransformer3DModel()
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[len(PREFIX_KEY) :]
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_inplace(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    transformer.load_state_dict(original_state_dict, strict=True)
+    return transformer
+
+
+def convert_vae(ckpt_path: str):
+    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
+    vae = AutoencoderKLCogVideoX()
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in VAE_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_inplace(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in VAE_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    vae.load_state_dict(original_state_dict, strict=True)
+    return vae
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transformer_ckpt_path", type=str, default=None, help="Path to original transformer checkpoint"
+    )
+    parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original vae checkpoint")
+    parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
+    parser.add_argument("--fp16", action="store_true", default=True, help="Whether to save the model weights in fp16")
+    parser.add_argument(
+        "--push_to_hub", action="store_true", default=False, help="Whether to push to HF Hub after saving"
+    )
+    parser.add_argument(
+        "--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    transformer = None
+    vae = None
+
+    if args.transformer_ckpt_path is not None:
+        transformer = convert_transformer(args.transformer_ckpt_path)
+    if args.vae_ckpt_path is not None:
+        vae = convert_vae(args.vae_ckpt_path)
+
+    text_encoder_id = "google/t5-v1_1-xxl"
+    tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
+    text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
+
+    scheduler = CogVideoXDDIMScheduler.from_config(
+        {
+            "snr_shift_scale": 3.0,
+            "beta_end": 0.012,
+            "beta_schedule": "scaled_linear",
+            "beta_start": 0.00085,
+            "clip_sample": False,
+            "num_train_timesteps": 1000,
+            "prediction_type": "v_prediction",
+            "rescale_betas_zero_snr": True,
+            "set_alpha_to_one": True,
+            "timestep_spacing": "linspace",
+        }
+    )
+
+    pipe = CogVideoXPipeline(
+        tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+    )
+
+    if args.fp16:
+        pipe = pipe.to(dtype=torch.float16)
+
+    pipe.save_pretrained(args.output_path, safe_serialization=True, push_to_hub=args.push_to_hub)

src/diffusers/__init__.py

@@ -79,13 +79,16 @@ else:
             "AsymmetricAutoencoderKL",
             "AuraFlowTransformer2DModel",
             "AutoencoderKL",
+            "AutoencoderKLCogVideoX",
             "AutoencoderKLTemporalDecoder",
             "AutoencoderOobleck",
             "AutoencoderTiny",
+            "CogVideoXTransformer3DModel",
             "ConsistencyDecoderVAE",
             "ControlNetModel",
             "ControlNetXSAdapter",
             "DiTTransformer2DModel",
+            "FluxControlNetModel",
             "FluxTransformer2DModel",
             "HunyuanDiT2DControlNetModel",
             "HunyuanDiT2DModel",
@@ -155,6 +158,8 @@ else:
         [
             "AmusedScheduler",
             "CMStochasticIterativeScheduler",
+            "CogVideoXDDIMScheduler",
+            "CogVideoXDPMScheduler",
             "DDIMInverseScheduler",
             "DDIMParallelScheduler",
             "DDIMScheduler",
@@ -248,7 +253,9 @@ else:
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
             "CLIPImageProjection",
+            "CogVideoXPipeline",
             "CycleDiffusionPipeline",
+            "FluxControlNetPipeline",
             "FluxPipeline",
             "HunyuanDiTControlNetPipeline",
             "HunyuanDiTPAGPipeline",
@@ -303,6 +310,7 @@ else:
             "StableCascadeCombinedPipeline",
             "StableCascadeDecoderPipeline",
             "StableCascadePriorPipeline",
+            "StableDiffusion3ControlNetInpaintingPipeline",
             "StableDiffusion3ControlNetPipeline",
             "StableDiffusion3Img2ImgPipeline",
             "StableDiffusion3InpaintPipeline",
@@ -535,13 +543,16 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AsymmetricAutoencoderKL,
             AuraFlowTransformer2DModel,
             AutoencoderKL,
+            AutoencoderKLCogVideoX,
             AutoencoderKLTemporalDecoder,
             AutoencoderOobleck,
             AutoencoderTiny,
+            CogVideoXTransformer3DModel,
             ConsistencyDecoderVAE,
             ControlNetModel,
             ControlNetXSAdapter,
             DiTTransformer2DModel,
+            FluxControlNetModel,
             FluxTransformer2DModel,
             HunyuanDiT2DControlNetModel,
             HunyuanDiT2DModel,
@@ -608,6 +619,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .schedulers import (
             AmusedScheduler,
             CMStochasticIterativeScheduler,
+            CogVideoXDDIMScheduler,
+            CogVideoXDPMScheduler,
             DDIMInverseScheduler,
             DDIMParallelScheduler,
             DDIMScheduler,
@@ -682,7 +695,9 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDMPipeline,
             AuraFlowPipeline,
             CLIPImageProjection,
+            CogVideoXPipeline,
             CycleDiffusionPipeline,
+            FluxControlNetPipeline,
             FluxPipeline,
             HunyuanDiTControlNetPipeline,
             HunyuanDiTPAGPipeline,

src/diffusers/loaders/single_file_model.py

@@ -75,6 +75,9 @@ 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 +74,11 @@ 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",
 }
 
@@ -111,6 +113,8 @@ 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"},
 }
@@ -494,7 +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:

src/diffusers/models/__init__.py

@@ -28,12 +28,14 @@ if is_torch_available():
     _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
     _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
     _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl_cogvideox"] = ["AutoencoderKLCogVideoX"]
     _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
     _import_structure["autoencoders.autoencoder_oobleck"] = ["AutoencoderOobleck"]
     _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
     _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["autoencoders.vq_model"] = ["VQModel"]
     _import_structure["controlnet"] = ["ControlNetModel"]
+    _import_structure["controlnet_flux"] = ["FluxControlNetModel"]
     _import_structure["controlnet_hunyuan"] = ["HunyuanDiT2DControlNetModel", "HunyuanDiT2DMultiControlNetModel"]
     _import_structure["controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
     _import_structure["controlnet_sparsectrl"] = ["SparseControlNetModel"]
@@ -41,6 +43,7 @@ if is_torch_available():
     _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
     _import_structure["transformers.auraflow_transformer_2d"] = ["AuraFlowTransformer2DModel"]
+    _import_structure["transformers.cogvideox_transformer_3d"] = ["CogVideoXTransformer3DModel"]
     _import_structure["transformers.dit_transformer_2d"] = ["DiTTransformer2DModel"]
     _import_structure["transformers.dual_transformer_2d"] = ["DualTransformer2DModel"]
     _import_structure["transformers.hunyuan_transformer_2d"] = ["HunyuanDiT2DModel"]
@@ -77,6 +80,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .autoencoders import (
             AsymmetricAutoencoderKL,
             AutoencoderKL,
+            AutoencoderKLCogVideoX,
             AutoencoderKLTemporalDecoder,
             AutoencoderOobleck,
             AutoencoderTiny,
@@ -84,6 +88,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VQModel,
         )
         from .controlnet import ControlNetModel
+        from .controlnet_flux import FluxControlNetModel
         from .controlnet_hunyuan import HunyuanDiT2DControlNetModel, HunyuanDiT2DMultiControlNetModel
         from .controlnet_sd3 import SD3ControlNetModel, SD3MultiControlNetModel
         from .controlnet_sparsectrl import SparseControlNetModel
@@ -92,6 +97,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .modeling_utils import ModelMixin
         from .transformers import (
             AuraFlowTransformer2DModel,
+            CogVideoXTransformer3DModel,
             DiTTransformer2DModel,
             DualTransformer2DModel,
             FluxTransformer2DModel,

src/diffusers/models/autoencoders/__init__.py

@@ -1,5 +1,6 @@
 from .autoencoder_asym_kl import AsymmetricAutoencoderKL
 from .autoencoder_kl import AutoencoderKL
+from .autoencoder_kl_cogvideox import AutoencoderKLCogVideoX
 from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
 from .autoencoder_oobleck import AutoencoderOobleck
 from .autoencoder_tiny import AutoencoderTiny

src/diffusers/models/controlnet_flux.py

@@ -0,0 +1,374 @@
+# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX 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 dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..loaders import PeftAdapterMixin
+from ..models.attention_processor import AttentionProcessor
+from ..models.modeling_utils import ModelMixin
+from ..utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
+from .controlnet import BaseOutput, zero_module
+from .embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings
+from .modeling_outputs import Transformer2DModelOutput
+from .transformers.transformer_flux import EmbedND, FluxSingleTransformerBlock, FluxTransformerBlock
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class FluxControlNetOutput(BaseOutput):
+    controlnet_block_samples: Tuple[torch.Tensor]
+    controlnet_single_block_samples: Tuple[torch.Tensor]
+
+
+class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 1,
+        in_channels: int = 64,
+        num_layers: int = 19,
+        num_single_layers: int = 38,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 24,
+        joint_attention_dim: int = 4096,
+        pooled_projection_dim: int = 768,
+        guidance_embeds: bool = False,
+        axes_dims_rope: List[int] = [16, 56, 56],
+    ):
+        super().__init__()
+        self.out_channels = in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        self.pos_embed = EmbedND(dim=self.inner_dim, theta=10000, axes_dim=axes_dims_rope)
+        text_time_guidance_cls = (
+            CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
+        )
+        self.time_text_embed = text_time_guidance_cls(
+            embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim
+        )
+
+        self.context_embedder = nn.Linear(joint_attention_dim, self.inner_dim)
+        self.x_embedder = torch.nn.Linear(in_channels, self.inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                FluxTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        self.single_transformer_blocks = nn.ModuleList(
+            [
+                FluxSingleTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                )
+                for i in range(num_single_layers)
+            ]
+        )
+
+        # controlnet_blocks
+        self.controlnet_blocks = nn.ModuleList([])
+        for _ in range(len(self.transformer_blocks)):
+            self.controlnet_blocks.append(zero_module(nn.Linear(self.inner_dim, self.inner_dim)))
+
+        self.controlnet_single_blocks = nn.ModuleList([])
+        for _ in range(len(self.single_transformer_blocks)):
+            self.controlnet_single_blocks.append(zero_module(nn.Linear(self.inner_dim, self.inner_dim)))
+
+        self.controlnet_x_embedder = zero_module(torch.nn.Linear(in_channels, self.inner_dim))
+
+        self.gradient_checkpointing = False
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self):
+        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):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    @classmethod
+    def from_transformer(
+        cls,
+        transformer,
+        num_layers=4,
+        num_single_layers=10,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 24,
+        load_weights_from_transformer=True,
+    ):
+        config = transformer.config
+        config["num_layers"] = num_layers
+        config["num_single_layers"] = num_single_layers
+        config["attention_head_dim"] = attention_head_dim
+        config["num_attention_heads"] = num_attention_heads
+
+        controlnet = cls(**config)
+
+        if load_weights_from_transformer:
+            controlnet.pos_embed.load_state_dict(transformer.pos_embed.state_dict())
+            controlnet.time_text_embed.load_state_dict(transformer.time_text_embed.state_dict())
+            controlnet.context_embedder.load_state_dict(transformer.context_embedder.state_dict())
+            controlnet.x_embedder.load_state_dict(transformer.x_embedder.state_dict())
+            controlnet.transformer_blocks.load_state_dict(transformer.transformer_blocks.state_dict(), strict=False)
+            controlnet.single_transformer_blocks.load_state_dict(
+                transformer.single_transformer_blocks.state_dict(), strict=False
+            )
+
+            controlnet.controlnet_x_embedder = zero_module(controlnet.controlnet_x_embedder)
+
+        return controlnet
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        controlnet_cond: torch.Tensor,
+        conditioning_scale: float = 1.0,
+        encoder_hidden_states: torch.Tensor = None,
+        pooled_projections: torch.Tensor = None,
+        timestep: torch.LongTensor = None,
+        img_ids: torch.Tensor = None,
+        txt_ids: torch.Tensor = None,
+        guidance: torch.Tensor = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
+        """
+        The [`FluxTransformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
+                Input `hidden_states`.
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
+            pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
+                from the embeddings of input conditions.
+            timestep ( `torch.LongTensor`):
+                Used to indicate denoising step.
+            block_controlnet_hidden_states: (`list` of `torch.Tensor`):
+                A list of tensors that if specified are added to the residuals of transformer blocks.
+            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).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        if joint_attention_kwargs is not None:
+            joint_attention_kwargs = joint_attention_kwargs.copy()
+            lora_scale = joint_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+        hidden_states = self.x_embedder(hidden_states)
+
+        # add
+        hidden_states = hidden_states + self.controlnet_x_embedder(controlnet_cond)
+
+        timestep = timestep.to(hidden_states.dtype) * 1000
+        if guidance is not None:
+            guidance = guidance.to(hidden_states.dtype) * 1000
+        else:
+            guidance = None
+        temb = (
+            self.time_text_embed(timestep, pooled_projections)
+            if guidance is None
+            else self.time_text_embed(timestep, guidance, pooled_projections)
+        )
+        encoder_hidden_states = self.context_embedder(encoder_hidden_states)
+
+        txt_ids = txt_ids.expand(img_ids.size(0), -1, -1)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = self.pos_embed(ids)
+
+        block_samples = ()
+        for index_block, block in enumerate(self.transformer_blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    image_rotary_emb,
+                    **ckpt_kwargs,
+                )
+
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                )
+            block_samples = block_samples + (hidden_states,)
+
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        single_block_samples = ()
+        for index_block, block in enumerate(self.single_transformer_blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    temb,
+                    image_rotary_emb,
+                    **ckpt_kwargs,
+                )
+
+            else:
+                hidden_states = block(
+                    hidden_states=hidden_states,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                )
+            single_block_samples = single_block_samples + (hidden_states[:, encoder_hidden_states.shape[1] :],)
+
+        # controlnet block
+        controlnet_block_samples = ()
+        for block_sample, controlnet_block in zip(block_samples, self.controlnet_blocks):
+            block_sample = controlnet_block(block_sample)
+            controlnet_block_samples = controlnet_block_samples + (block_sample,)
+
+        controlnet_single_block_samples = ()
+        for single_block_sample, controlnet_block in zip(single_block_samples, self.controlnet_single_blocks):
+            single_block_sample = controlnet_block(single_block_sample)
+            controlnet_single_block_samples = controlnet_single_block_samples + (single_block_sample,)
+
+        # scaling
+        controlnet_block_samples = [sample * conditioning_scale for sample in controlnet_block_samples]
+        controlnet_single_block_samples = [sample * conditioning_scale for sample in controlnet_single_block_samples]
+
+        #
+        controlnet_block_samples = None if len(controlnet_block_samples) == 0 else controlnet_block_samples
+        controlnet_single_block_samples = (
+            None if len(controlnet_single_block_samples) == 0 else controlnet_single_block_samples
+        )
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (controlnet_block_samples, controlnet_single_block_samples)
+
+        return FluxControlNetOutput(
+            controlnet_block_samples=controlnet_block_samples,
+            controlnet_single_block_samples=controlnet_single_block_samples,
+        )

src/diffusers/models/controlnet_sd3.py

@@ -55,6 +55,7 @@ class SD3ControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginal
         pooled_projection_dim: int = 2048,
         out_channels: int = 16,
         pos_embed_max_size: int = 96,
+        extra_conditioning_channels: int = 0,
     ):
         super().__init__()
         default_out_channels = in_channels
@@ -98,7 +99,7 @@ class SD3ControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginal
             height=sample_size,
             width=sample_size,
             patch_size=patch_size,
-            in_channels=in_channels,
+            in_channels=in_channels + extra_conditioning_channels,
             embed_dim=self.inner_dim,
             pos_embed_type=None,
         )

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/downsampling.py

@@ -285,6 +285,74 @@ class KDownsample2D(nn.Module):
         return F.conv2d(inputs, weight, stride=2)
 
 
+class CogVideoXDownsample3D(nn.Module):
+    # Todo: Wait for paper relase.
+    r"""
+    A 3D Downsampling layer using in [CogVideoX]() by Tsinghua University & ZhipuAI
+
+    Args:
+        in_channels (`int`):
+            Number of channels in the input image.
+        out_channels (`int`):
+            Number of channels produced by the convolution.
+        kernel_size (`int`, defaults to `3`):
+            Size of the convolving kernel.
+        stride (`int`, defaults to `2`):
+            Stride of the convolution.
+        padding (`int`, defaults to `0`):
+            Padding added to all four sides of the input.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to compress the time dimension.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int = 3,
+        stride: int = 2,
+        padding: int = 0,
+        compress_time: bool = False,
+    ):
+        super().__init__()
+
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.compress_time:
+            batch_size, channels, frames, height, width = x.shape
+
+            # (batch_size, channels, frames, height, width) -> (batch_size, height, width, channels, frames) -> (batch_size * height * width, channels, frames)
+            x = x.permute(0, 3, 4, 1, 2).reshape(batch_size * height * width, channels, frames)
+
+            if x.shape[-1] % 2 == 1:
+                x_first, x_rest = x[..., 0], x[..., 1:]
+                if x_rest.shape[-1] > 0:
+                    # (batch_size * height * width, channels, frames - 1) -> (batch_size * height * width, channels, (frames - 1) // 2)
+                    x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
+
+                x = torch.cat([x_first[..., None], x_rest], dim=-1)
+                # (batch_size * height * width, channels, (frames // 2) + 1) -> (batch_size, height, width, channels, (frames // 2) + 1) -> (batch_size, channels, (frames // 2) + 1, height, width)
+                x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
+            else:
+                # (batch_size * height * width, channels, frames) -> (batch_size * height * width, channels, frames // 2)
+                x = F.avg_pool1d(x, kernel_size=2, stride=2)
+                # (batch_size * height * width, channels, frames // 2) -> (batch_size, height, width, channels, frames // 2) -> (batch_size, channels, frames // 2, height, width)
+                x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
+
+        # Pad the tensor
+        pad = (0, 1, 0, 1)
+        x = F.pad(x, pad, mode="constant", value=0)
+        batch_size, channels, frames, height, width = x.shape
+        # (batch_size, channels, frames, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size * frames, channels, height, width)
+        x = x.permute(0, 2, 1, 3, 4).reshape(batch_size * frames, channels, height, width)
+        x = self.conv(x)
+        # (batch_size * frames, channels, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size, channels, frames, height, width)
+        x = x.reshape(batch_size, frames, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
+        return x
+
+
 def downsample_2d(
     hidden_states: torch.Tensor,
     kernel: Optional[torch.Tensor] = None,

src/diffusers/models/embeddings.py

@@ -78,6 +78,53 @@ def get_timestep_embedding(
     return emb
 
 
+def get_3d_sincos_pos_embed(
+    embed_dim: int,
+    spatial_size: Union[int, Tuple[int, int]],
+    temporal_size: int,
+    spatial_interpolation_scale: float = 1.0,
+    temporal_interpolation_scale: float = 1.0,
+) -> np.ndarray:
+    r"""
+    Args:
+        embed_dim (`int`):
+        spatial_size (`int` or `Tuple[int, int]`):
+        temporal_size (`int`):
+        spatial_interpolation_scale (`float`, defaults to 1.0):
+        temporal_interpolation_scale (`float`, defaults to 1.0):
+    """
+    if embed_dim % 4 != 0:
+        raise ValueError("`embed_dim` must be divisible by 4")
+    if isinstance(spatial_size, int):
+        spatial_size = (spatial_size, spatial_size)
+
+    embed_dim_spatial = 3 * embed_dim // 4
+    embed_dim_temporal = embed_dim // 4
+
+    # 1. Spatial
+    grid_h = np.arange(spatial_size[1], dtype=np.float32) / spatial_interpolation_scale
+    grid_w = np.arange(spatial_size[0], dtype=np.float32) / spatial_interpolation_scale
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, spatial_size[1], spatial_size[0]])
+    pos_embed_spatial = get_2d_sincos_pos_embed_from_grid(embed_dim_spatial, grid)
+
+    # 2. Temporal
+    grid_t = np.arange(temporal_size, dtype=np.float32) / temporal_interpolation_scale
+    pos_embed_temporal = get_1d_sincos_pos_embed_from_grid(embed_dim_temporal, grid_t)
+
+    # 3. Concat
+    pos_embed_spatial = pos_embed_spatial[np.newaxis, :, :]
+    pos_embed_spatial = np.repeat(pos_embed_spatial, temporal_size, axis=0)  # [T, H*W, D // 4 * 3]
+
+    pos_embed_temporal = pos_embed_temporal[:, np.newaxis, :]
+    pos_embed_temporal = np.repeat(pos_embed_temporal, spatial_size[0] * spatial_size[1], axis=1)  # [T, H*W, D // 4]
+
+    pos_embed = np.concatenate([pos_embed_temporal, pos_embed_spatial], axis=-1)  # [T, H*W, D]
+    return pos_embed
+
+
 def get_2d_sincos_pos_embed(
     embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16
 ):
@@ -287,6 +334,46 @@ class LuminaPatchEmbed(nn.Module):
         )
 
 
+class CogVideoXPatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        embed_dim: int = 1920,
+        text_embed_dim: int = 4096,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+        self.patch_size = patch_size
+
+        self.proj = nn.Conv2d(
+            in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
+        )
+        self.text_proj = nn.Linear(text_embed_dim, embed_dim)
+
+    def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
+        r"""
+        Args:
+            text_embeds (`torch.Tensor`):
+                Input text embeddings. Expected shape: (batch_size, seq_length, embedding_dim).
+            image_embeds (`torch.Tensor`):
+                Input image embeddings. Expected shape: (batch_size, num_frames, channels, height, width).
+        """
+        text_embeds = self.text_proj(text_embeds)
+
+        batch, num_frames, channels, height, width = image_embeds.shape
+        image_embeds = image_embeds.reshape(-1, channels, height, width)
+        image_embeds = self.proj(image_embeds)
+        image_embeds = image_embeds.view(batch, num_frames, *image_embeds.shape[1:])
+        image_embeds = image_embeds.flatten(3).transpose(2, 3)  # [batch, num_frames, height x width, channels]
+        image_embeds = image_embeds.flatten(1, 2)  # [batch, num_frames x height x width, channels]
+
+        embeds = torch.cat(
+            [text_embeds, image_embeds], dim=1
+        ).contiguous()  # [batch, seq_length + num_frames x height x width, channels]
+        return embeds
+
+
 def get_2d_rotary_pos_embed(embed_dim, crops_coords, grid_size, use_real=True):
     """
     RoPE for image tokens with 2d structure.

src/diffusers/models/normalization.py

@@ -34,19 +34,53 @@ 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__(self, embedding_dim: int, num_embeddings: int):
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_embeddings: Optional[int] = None,
+        output_dim: Optional[int] = None,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-5,
+        chunk_dim: int = 0,
+    ):
         super().__init__()
-        self.emb = nn.Embedding(num_embeddings, embedding_dim)
-        self.silu = nn.SiLU()
-        self.linear = nn.Linear(embedding_dim, embedding_dim * 2)
-        self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False)
 
-    def forward(self, x: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor:
-        emb = self.linear(self.silu(self.emb(timestep)))
-        scale, shift = torch.chunk(emb, 2)
+        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
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, output_dim)
+        self.norm = nn.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine)
+
+    def forward(
+        self, x: torch.Tensor, timestep: Optional[torch.Tensor] = None, temb: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        if self.emb is not None:
+            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, :]
+        else:
+            scale, shift = temb.chunk(2, dim=0)
+
         x = self.norm(x) * (1 + scale) + shift
         return x
 
@@ -321,6 +355,30 @@ class LuminaLayerNormContinuous(nn.Module):
         return x
 
 
+class CogVideoXLayerNormZero(nn.Module):
+    def __init__(
+        self,
+        conditioning_dim: int,
+        embedding_dim: int,
+        elementwise_affine: bool = True,
+        eps: float = 1e-5,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(conditioning_dim, 6 * embedding_dim, bias=bias)
+        self.norm = nn.LayerNorm(embedding_dim, eps=eps, elementwise_affine=elementwise_affine)
+
+    def forward(
+        self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, temb: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        shift, scale, gate, enc_shift, enc_scale, enc_gate = self.linear(self.silu(temb)).chunk(6, dim=1)
+        hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
+        encoder_hidden_states = self.norm(encoder_hidden_states) * (1 + enc_scale)[:, None, :] + enc_shift[:, None, :]
+        return hidden_states, encoder_hidden_states, gate[:, None, :], enc_gate[:, None, :]
+
+
 if is_torch_version(">=", "2.1.0"):
     LayerNorm = nn.LayerNorm
 else:

src/diffusers/models/transformers/__init__.py

@@ -3,6 +3,7 @@ from ...utils import is_torch_available
 
 if is_torch_available():
     from .auraflow_transformer_2d import AuraFlowTransformer2DModel
+    from .cogvideox_transformer_3d import CogVideoXTransformer3DModel
     from .dit_transformer_2d import DiTTransformer2DModel
     from .dual_transformer_2d import DualTransformer2DModel
     from .hunyuan_transformer_2d import HunyuanDiT2DModel

src/diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -0,0 +1,369 @@
+# 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 Any, Dict, Optional, Union
+
+import torch
+from torch import nn
+
+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 import Attention, FeedForward
+from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps, get_3d_sincos_pos_embed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNorm, CogVideoXLayerNormZero
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@maybe_allow_in_graph
+class CogVideoXBlock(nn.Module):
+    r"""
+    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.
+        time_embed_dim (`int`):
+            The number of channels in timestep embedding.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability to use.
+        activation_fn (`str`, defaults to `"gelu-approximate"`):
+            Activation function to be used in feed-forward.
+        attention_bias (`bool`, defaults to `False`):
+            Whether or not to use bias in attention projection layers.
+        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_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.
+        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__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        time_embed_dim: int,
+        dropout: float = 0.0,
+        activation_fn: str = "gelu-approximate",
+        attention_bias: bool = False,
+        qk_norm: bool = True,
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        final_dropout: bool = True,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+
+        # 1. Self Attention
+        self.norm1 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            qk_norm="layer_norm" if qk_norm else None,
+            eps=1e-6,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+        )
+
+        # 2. Feed Forward
+        self.norm2 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+    ) -> torch.Tensor:
+        norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
+            hidden_states, encoder_hidden_states, temb
+        )
+
+        # attention
+        text_length = norm_encoder_hidden_states.size(1)
+
+        # CogVideoX uses concatenated text + video embeddings with self-attention instead of using
+        # them in cross-attention individually
+        norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
+        attn_output = self.attn1(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=None,
+        )
+
+        hidden_states = hidden_states + gate_msa * attn_output[:, text_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_output[:, :text_length]
+
+        # norm & modulate
+        norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
+            hidden_states, encoder_hidden_states, temb
+        )
+
+        # feed-forward
+        norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
+        ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = hidden_states + gate_ff * ff_output[:, text_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_length]
+        return hidden_states, encoder_hidden_states
+
+
+class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
+    """
+    A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).
+
+    Parameters:
+        num_attention_heads (`int`, defaults to `30`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, defaults to `64`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, *optional*, defaults to `16`):
+            The number of channels in the output.
+        flip_sin_to_cos (`bool`, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        time_embed_dim (`int`, defaults to `512`):
+            Output dimension of timestep embeddings.
+        text_embed_dim (`int`, defaults to `4096`):
+            Input dimension of text embeddings from the text encoder.
+        num_layers (`int`, defaults to `30`):
+            The number of layers of Transformer blocks to use.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability to use.
+        attention_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in the attention projection layers.
+        sample_width (`int`, defaults to `90`):
+            The width of the input latents.
+        sample_height (`int`, defaults to `60`):
+            The height of the input latents.
+        sample_frames (`int`, defaults to `49`):
+            The number of frames in the input latents. Note that this parameter was incorrectly initialized to 49
+            instead of 13 because CogVideoX processed 13 latent frames at once in its default and recommended settings,
+            but cannot be changed to the correct value to ensure backwards compatibility. To create a transformer with
+            K latent frames, the correct value to pass here would be: ((K - 1) * temporal_compression_ratio + 1).
+        patch_size (`int`, defaults to `2`):
+            The size of the patches to use in the patch embedding layer.
+        temporal_compression_ratio (`int`, defaults to `4`):
+            The compression ratio across the temporal dimension. See documentation for `sample_frames`.
+        max_text_seq_length (`int`, defaults to `226`):
+            The maximum sequence length of the input text embeddings.
+        activation_fn (`str`, defaults to `"gelu-approximate"`):
+            Activation function to use in feed-forward.
+        timestep_activation_fn (`str`, defaults to `"silu"`):
+            Activation function to use when generating the timestep embeddings.
+        norm_elementwise_affine (`bool`, defaults to `True`):
+            Whether or not to use elementwise affine in normalization layers.
+        norm_eps (`float`, defaults to `1e-5`):
+            The epsilon value to use in normalization layers.
+        spatial_interpolation_scale (`float`, defaults to `1.875`):
+            Scaling factor to apply in 3D positional embeddings across spatial dimensions.
+        temporal_interpolation_scale (`float`, defaults to `1.0`):
+            Scaling factor to apply in 3D positional embeddings across temporal dimensions.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 30,
+        attention_head_dim: int = 64,
+        in_channels: int = 16,
+        out_channels: Optional[int] = 16,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        time_embed_dim: int = 512,
+        text_embed_dim: int = 4096,
+        num_layers: int = 30,
+        dropout: float = 0.0,
+        attention_bias: bool = True,
+        sample_width: int = 90,
+        sample_height: int = 60,
+        sample_frames: int = 49,
+        patch_size: int = 2,
+        temporal_compression_ratio: int = 4,
+        max_text_seq_length: int = 226,
+        activation_fn: str = "gelu-approximate",
+        timestep_activation_fn: str = "silu",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        spatial_interpolation_scale: float = 1.875,
+        temporal_interpolation_scale: float = 1.0,
+    ):
+        super().__init__()
+        inner_dim = num_attention_heads * attention_head_dim
+
+        post_patch_height = sample_height // patch_size
+        post_patch_width = sample_width // patch_size
+        post_time_compression_frames = (sample_frames - 1) // temporal_compression_ratio + 1
+        self.num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+
+        # 1. Patch embedding
+        self.patch_embed = CogVideoXPatchEmbed(patch_size, in_channels, inner_dim, text_embed_dim, bias=True)
+        self.embedding_dropout = nn.Dropout(dropout)
+
+        # 2. 3D positional embeddings
+        spatial_pos_embedding = get_3d_sincos_pos_embed(
+            inner_dim,
+            (post_patch_width, post_patch_height),
+            post_time_compression_frames,
+            spatial_interpolation_scale,
+            temporal_interpolation_scale,
+        )
+        spatial_pos_embedding = torch.from_numpy(spatial_pos_embedding).flatten(0, 1)
+        pos_embedding = torch.zeros(1, max_text_seq_length + self.num_patches, inner_dim, requires_grad=False)
+        pos_embedding.data[:, max_text_seq_length:].copy_(spatial_pos_embedding)
+        self.register_buffer("pos_embedding", pos_embedding, persistent=False)
+
+        # 3. Time embeddings
+        self.time_proj = Timesteps(inner_dim, flip_sin_to_cos, freq_shift)
+        self.time_embedding = TimestepEmbedding(inner_dim, time_embed_dim, timestep_activation_fn)
+
+        # 4. Define spatio-temporal transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                CogVideoXBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    time_embed_dim=time_embed_dim,
+                    dropout=dropout,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+        self.norm_final = nn.LayerNorm(inner_dim, norm_eps, norm_elementwise_affine)
+
+        # 5. Output blocks
+        self.norm_out = AdaLayerNorm(
+            embedding_dim=time_embed_dim,
+            output_dim=2 * inner_dim,
+            norm_elementwise_affine=norm_elementwise_affine,
+            norm_eps=norm_eps,
+            chunk_dim=1,
+        )
+        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        self.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: Union[int, float, torch.LongTensor],
+        timestep_cond: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        batch_size, num_frames, channels, height, width = hidden_states.shape
+
+        # 1. Time embedding
+        timesteps = timestep
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=hidden_states.dtype)
+        emb = self.time_embedding(t_emb, timestep_cond)
+
+        # 2. Patch embedding
+        hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
+
+        # 3. Position embedding
+        seq_length = height * width * num_frames // (self.config.patch_size**2)
+
+        pos_embeds = self.pos_embedding[:, : self.config.max_text_seq_length + seq_length]
+        hidden_states = hidden_states + pos_embeds
+        hidden_states = self.embedding_dropout(hidden_states)
+
+        encoder_hidden_states = hidden_states[:, : self.config.max_text_seq_length]
+        hidden_states = hidden_states[:, self.config.max_text_seq_length :]
+
+        # 4. Transformer blocks
+        for i, block in enumerate(self.transformer_blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    emb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb=emb,
+                )
+
+        hidden_states = self.norm_final(hidden_states)
+
+        # 5. Final block
+        hidden_states = self.norm_out(hidden_states, temb=emb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # 6. Unpatchify
+        p = self.config.patch_size
+        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, channels, p, p)
+        output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)

src/diffusers/models/transformers/transformer_flux.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Black Forest Labs, The HuggingFace Team. All rights reserved.
+# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX 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.
@@ -15,6 +15,7 @@
 
 from typing import Any, Dict, List, Optional, Union
 
+import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -125,6 +126,8 @@ 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
 
@@ -223,6 +226,8 @@ 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
 
@@ -317,6 +322,8 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         txt_ids: torch.Tensor = None,
         guidance: torch.Tensor = None,
         joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_block_samples=None,
+        controlnet_single_block_samples=None,
         return_dict: bool = True,
     ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
         """
@@ -373,6 +380,7 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         )
         encoder_hidden_states = self.context_embedder(encoder_hidden_states)
 
+        txt_ids = txt_ids.expand(img_ids.size(0), -1, -1)
         ids = torch.cat((txt_ids, img_ids), dim=1)
         image_rotary_emb = self.pos_embed(ids)
 
@@ -406,6 +414,12 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                     image_rotary_emb=image_rotary_emb,
                 )
 
+            # controlnet residual
+            if controlnet_block_samples is not None:
+                interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
+                interval_control = int(np.ceil(interval_control))
+                hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]
+
         hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 
         for index_block, block in enumerate(self.single_transformer_blocks):
@@ -436,6 +450,15 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                     image_rotary_emb=image_rotary_emb,
                 )
 
+            # controlnet residual
+            if controlnet_single_block_samples is not None:
+                interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
+                interval_control = int(np.ceil(interval_control))
+                hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
+                    hidden_states[:, encoder_hidden_states.shape[1] :, ...]
+                    + controlnet_single_block_samples[index_block // interval_control]
+                )
+
         hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
 
         hidden_states = self.norm_out(hidden_states, temb)

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,
                 )
             )
 
@@ -385,6 +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 = []
@@ -466,6 +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,
                 )
             )
 

src/diffusers/models/upsampling.py

@@ -348,6 +348,70 @@ class KUpsample2D(nn.Module):
         return F.conv_transpose2d(inputs, weight, stride=2, padding=self.pad * 2 + 1)
 
 
+class CogVideoXUpsample3D(nn.Module):
+    r"""
+    A 3D Upsample layer using in CogVideoX by Tsinghua University & ZhipuAI # Todo: Wait for paper relase.
+
+    Args:
+        in_channels (`int`):
+            Number of channels in the input image.
+        out_channels (`int`):
+            Number of channels produced by the convolution.
+        kernel_size (`int`, defaults to `3`):
+            Size of the convolving kernel.
+        stride (`int`, defaults to `1`):
+            Stride of the convolution.
+        padding (`int`, defaults to `1`):
+            Padding added to all four sides of the input.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to compress the time dimension.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int = 3,
+        stride: int = 1,
+        padding: int = 1,
+        compress_time: bool = False,
+    ) -> None:
+        super().__init__()
+
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        if self.compress_time:
+            if inputs.shape[2] > 1 and inputs.shape[2] % 2 == 1:
+                # split first frame
+                x_first, x_rest = inputs[:, :, 0], inputs[:, :, 1:]
+
+                x_first = F.interpolate(x_first, scale_factor=2.0)
+                x_rest = F.interpolate(x_rest, scale_factor=2.0)
+                x_first = x_first[:, :, None, :, :]
+                inputs = torch.cat([x_first, x_rest], dim=2)
+            elif inputs.shape[2] > 1:
+                inputs = F.interpolate(inputs, scale_factor=2.0)
+            else:
+                inputs = inputs.squeeze(2)
+                inputs = F.interpolate(inputs, scale_factor=2.0)
+                inputs = inputs[:, :, None, :, :]
+        else:
+            # only interpolate 2D
+            b, c, t, h, w = inputs.shape
+            inputs = inputs.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+            inputs = F.interpolate(inputs, scale_factor=2.0)
+            inputs = inputs.reshape(b, t, c, *inputs.shape[2:]).permute(0, 2, 1, 3, 4)
+
+        b, c, t, h, w = inputs.shape
+        inputs = inputs.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        inputs = self.conv(inputs)
+        inputs = inputs.reshape(b, t, *inputs.shape[1:]).permute(0, 2, 1, 3, 4)
+
+        return inputs
+
+
 def upfirdn2d_native(
     tensor: torch.Tensor,
     kernel: torch.Tensor,

src/diffusers/pipelines/__init__.py

@@ -124,7 +124,7 @@ else:
         "AnimateDiffSparseControlNetPipeline",
         "AnimateDiffVideoToVideoPipeline",
     ]
-    _import_structure["flux"] = ["FluxPipeline"]
+    _import_structure["flux"] = ["FluxPipeline", "FluxControlNetPipeline"]
     _import_structure["audioldm"] = ["AudioLDMPipeline"]
     _import_structure["audioldm2"] = [
         "AudioLDM2Pipeline",
@@ -132,6 +132,7 @@ else:
         "AudioLDM2UNet2DConditionModel",
     ]
     _import_structure["blip_diffusion"] = ["BlipDiffusionPipeline"]
+    _import_structure["cogvideo"] = ["CogVideoXPipeline"]
     _import_structure["controlnet"].extend(
         [
             "BlipDiffusionControlNetPipeline",
@@ -172,6 +173,7 @@ else:
     _import_structure["controlnet_sd3"].extend(
         [
             "StableDiffusion3ControlNetPipeline",
+            "StableDiffusion3ControlNetInpaintingPipeline",
         ]
     )
     _import_structure["deepfloyd_if"] = [
@@ -451,6 +453,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         )
         from .aura_flow import AuraFlowPipeline
         from .blip_diffusion import BlipDiffusionPipeline
+        from .cogvideo import CogVideoXPipeline
         from .controlnet import (
             BlipDiffusionControlNetPipeline,
             StableDiffusionControlNetImg2ImgPipeline,
@@ -463,9 +466,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .controlnet_hunyuandit import (
             HunyuanDiTControlNetPipeline,
         )
-        from .controlnet_sd3 import (
-            StableDiffusion3ControlNetPipeline,
-        )
+        from .controlnet_sd3 import StableDiffusion3ControlNetInpaintingPipeline, StableDiffusion3ControlNetPipeline
         from .controlnet_xs import (
             StableDiffusionControlNetXSPipeline,
             StableDiffusionXLControlNetXSPipeline,
@@ -492,7 +493,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VersatileDiffusionTextToImagePipeline,
             VQDiffusionPipeline,
         )
-        from .flux import FluxPipeline
+        from .flux import FluxControlNetPipeline, FluxPipeline
         from .hunyuandit import HunyuanDiTPipeline
         from .i2vgen_xl import I2VGenXLPipeline
         from .kandinsky import (

src/diffusers/pipelines/cogvideo/__init__.py

@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_cogvideox"] = ["CogVideoXPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_cogvideox import CogVideoXPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/cogvideo/pipeline_cogvideox.py

@@ -0,0 +1,677 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
+from ...pipelines.pipeline_utils import DiffusionPipeline
+from ...schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
+from ...utils import BaseOutput, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+
+
+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.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 "
+        ...     "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)
+        ```
+"""
+
+
+# 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
+
+
+@dataclass
+class CogVideoXPipelineOutput(BaseOutput):
+    r"""
+    Output class for CogVideo pipelines.
+
+    Args:
+        frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
+            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+            `(batch_size, num_frames, channels, height, width)`.
+    """
+
+    frames: torch.Tensor
+
+
+class CogVideoXPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-video generation using CogVideoX.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. CogVideoX uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
+            [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`CogVideoXTransformer3DModel`]):
+            A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
+    """
+
+    _optional_components = []
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+    ]
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+        self.vae_scale_factor_spatial = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.vae_scale_factor_temporal = (
+            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+    ):
+        shape = (
+            batch_size,
+            (num_frames - 1) // self.vae_scale_factor_temporal + 1,
+            num_channels_latents,
+            height // self.vae_scale_factor_spatial,
+            width // self.vae_scale_factor_spatial,
+        )
+        if 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)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
+        latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
+        latents = 1 / self.vae.config.scaling_factor * latents
+
+        frames = self.vae.decode(latents).sample
+        return frames
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_on_step_end_tensor_inputs,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @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: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 480,
+        width: int = 720,
+        num_frames: int = 49,
+        num_inference_steps: int = 50,
+        timesteps: Optional[List[int]] = None,
+        guidance_scale: float = 6,
+        use_dynamic_cfg: bool = False,
+        num_videos_per_prompt: int = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: str = "pil",
+        return_dict: bool = True,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 226,
+    ) -> Union[CogVideoXPipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            height (`int`, *optional*, defaults to self.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_frames (`int`, defaults to `48`):
+                Number of frames to generate. Must be divisible by self.vae_scale_factor_temporal. Generated video will
+                contain 1 extra frame because CogVideoX is conditioned with (num_seconds * fps + 1) frames where
+                num_seconds is 6 and fps is 4. However, since videos can be saved at any fps, the only condition that
+                needs to be satisfied is that of divisibility mentioned above.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, defaults to `226`):
+                Maximum sequence length in encoded prompt. Must be consistent with
+                `self.transformer.config.max_text_seq_length` otherwise may lead to poor results.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] or `tuple`:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        if num_frames > 49:
+            raise ValueError(
+                "The number of frames must be less than 49 for now due to static positional embeddings. This will be updated in the future to remove this limitation."
+            )
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        height = height or self.transformer.config.sample_size * self.vae_scale_factor_spatial
+        width = width or self.transformer.config.sample_size * self.vae_scale_factor_spatial
+        num_videos_per_prompt = 1
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_on_step_end_tensor_inputs,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+        self._guidance_scale = guidance_scale
+        self._interrupt = False
+
+        # 2. Default call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            latent_channels,
+            num_frames,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            # for DPM-solver++
+            old_pred_original_sample = None
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                # perform guidance
+                if use_dynamic_cfg:
+                    self._guidance_scale = 1 + guidance_scale * (
+                        (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
+                    )
+                if do_classifier_free_guidance:
+                    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
+                if not isinstance(self.scheduler, CogVideoXDPMScheduler):
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                else:
+                    latents, old_pred_original_sample = self.scheduler.step(
+                        noise_pred,
+                        old_pred_original_sample,
+                        t,
+                        timesteps[i - 1] if i > 0 else None,
+                        latents,
+                        **extra_step_kwargs,
+                        return_dict=False,
+                    )
+                latents = latents.to(prompt_embeds.dtype)
+
+                # call the callback, if provided
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if not output_type == "latent":
+            video = self.decode_latents(latents)
+            video = self.video_processor.postprocess_video(video=video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return CogVideoXPipelineOutput(frames=video)

src/diffusers/pipelines/controlnet_sd3/__init__.py

@@ -23,6 +23,9 @@ except OptionalDependencyNotAvailable:
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_stable_diffusion_3_controlnet"] = ["StableDiffusion3ControlNetPipeline"]
+    _import_structure["pipeline_stable_diffusion_3_controlnet_inpainting"] = [
+        "StableDiffusion3ControlNetInpaintingPipeline"
+    ]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -33,6 +36,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
         from .pipeline_stable_diffusion_3_controlnet import StableDiffusion3ControlNetPipeline
+        from .pipeline_stable_diffusion_3_controlnet_inpainting import StableDiffusion3ControlNetInpaintingPipeline
 
     try:
         if not (is_transformers_available() and is_flax_available()):

src/diffusers/pipelines/flux/__init__.py

@@ -23,6 +23,7 @@ except OptionalDependencyNotAvailable:
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_flux"] = ["FluxPipeline"]
+    _import_structure["pipeline_flux_controlnet"] = ["FluxControlNetPipeline"]
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
         if not (is_transformers_available() and is_torch_available()):
@@ -31,6 +32,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
     else:
         from .pipeline_flux import FluxPipeline
+        from .pipeline_flux_controlnet import FluxControlNetPipeline
 else:
     import sys
 

src/diffusers/pipelines/flux/pipeline_flux_controlnet.py

@@ -0,0 +1,861 @@
+# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX 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 numpy as np
+import torch
+from transformers import (
+    CLIPTextModel,
+    CLIPTokenizer,
+    T5EncoderModel,
+    T5TokenizerFast,
+)
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FluxLoraLoaderMixin
+from ...models.autoencoders import AutoencoderKL
+from ...models.controlnet_flux import FluxControlNetModel
+from ...models.transformers import FluxTransformer2DModel
+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 .pipeline_output import FluxPipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers.utils import load_image
+        >>> from diffusers import FluxControlNetPipeline
+        >>> from diffusers import FluxControlNetModel
+
+        >>> controlnet_model = "InstantX/FLUX.1-dev-controlnet-canny-alpha"
+        >>> controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16)
+        >>> pipe = FluxControlNetPipeline.from_pretrained(
+        ...     base_model, controlnet=controlnet, torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe.to("cuda")
+        >>> control_image = load_image("https://huggingface.co/InstantX/SD3-Controlnet-Canny/resolve/main/canny.jpg")
+        >>> prompt = "A girl in city, 25 years old, cool, futuristic"
+        >>> image = pipe(
+        ...     prompt,
+        ...     control_image=control_image,
+        ...     controlnet_conditioning_scale=0.6,
+        ...     num_inference_steps=28,
+        ...     guidance_scale=3.5,
+        ... ).images[0]
+        >>> image.save("flux.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    """
+    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 FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
+    r"""
+    The Flux pipeline for text-to-image generation.
+
+    Reference: https://blackforestlabs.ai/announcing-black-forest-labs/
+
+    Args:
+        transformer ([`FluxTransformer2DModel`]):
+            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 ([`CLIPTextModel`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        text_encoder_2 ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        text_encoder_2: T5EncoderModel,
+        tokenizer_2: T5TokenizerFast,
+        transformer: FluxTransformer2DModel,
+        controlnet: FluxControlNetModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            transformer=transformer,
+            scheduler=scheduler,
+            controlnet=controlnet,
+        )
+        self.vae_scale_factor = (
+            2 ** (len(self.vae.config.block_out_channels)) if hasattr(self, "vae") and self.vae is not None else 16
+        )
+        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 = 64
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_images_per_prompt: int = 1,
+        max_sequence_length: int = 512,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer_2(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_length=False,
+            return_overflowing_tokens=False,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, 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_2(text_input_ids.to(device), output_hidden_states=False)[0]
+
+        dtype = self.text_encoder_2.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
+
+    def _get_clip_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+    ):
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer_max_length,
+            truncation=True,
+            return_overflowing_tokens=False,
+            return_length=False,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, 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 = self.text_encoder(text_input_ids.to(device), output_hidden_states=False)
+
+        # Use pooled output of CLIPTextModel
+        prompt_embeds = prompt_embeds.pooler_output
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_2: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        max_sequence_length: int = 512,
+        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
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            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.
+            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.
+            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, FluxLoraLoaderMixin):
+            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
+
+            # We only use the pooled prompt output from the CLIPTextModel
+            pooled_prompt_embeds = self._get_clip_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+            )
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt_2,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, FluxLoraLoaderMixin) 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, FluxLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype
+        text_ids = torch.zeros(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
+        text_ids = text_ids.repeat(num_images_per_prompt, 1, 1)
+
+        return prompt_embeds, pooled_prompt_embeds, text_ids
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        height,
+        width,
+        prompt_embeds=None,
+        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 is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+        if prompt_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 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}")
+
+    @staticmethod
+    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids
+    def _prepare_latent_image_ids(batch_size, height, width, device, dtype):
+        latent_image_ids = torch.zeros(height // 2, width // 2, 3)
+        latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height // 2)[:, None]
+        latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width // 2)[None, :]
+
+        latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape
+
+        latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1)
+        latent_image_ids = latent_image_ids.reshape(
+            batch_size, latent_image_id_height * latent_image_id_width, latent_image_id_channels
+        )
+
+        return latent_image_ids.to(device=device, dtype=dtype)
+
+    @staticmethod
+    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents
+    def _pack_latents(latents, batch_size, num_channels_latents, height, width):
+        latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
+        latents = latents.permute(0, 2, 4, 1, 3, 5)
+        latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4)
+
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents
+    def _unpack_latents(latents, height, width, vae_scale_factor):
+        batch_size, num_patches, channels = latents.shape
+
+        height = height // vae_scale_factor
+        width = width // vae_scale_factor
+
+        latents = latents.view(batch_size, height, width, channels // 4, 2, 2)
+        latents = latents.permute(0, 3, 1, 4, 2, 5)
+
+        latents = latents.reshape(batch_size, channels // (2 * 2), height * 2, width * 2)
+
+        return latents
+
+    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_latents
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        height = 2 * (int(height) // self.vae_scale_factor)
+        width = 2 * (int(width) // self.vae_scale_factor)
+
+        shape = (batch_size, num_channels_latents, height, width)
+
+        if latents is not None:
+            latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, device, dtype)
+            return latents.to(device=device, dtype=dtype), latent_image_ids
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
+
+        latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, device, dtype)
+
+        return latents, latent_image_ids
+
+    # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image
+    def prepare_image(
+        self,
+        image,
+        width,
+        height,
+        batch_size,
+        num_images_per_prompt,
+        device,
+        dtype,
+        do_classifier_free_guidance=False,
+        guess_mode=False,
+    ):
+        if isinstance(image, torch.Tensor):
+            pass
+        else:
+            image = self.image_processor.preprocess(image, height=height, width=width)
+
+        image_batch_size = image.shape[0]
+
+        if image_batch_size == 1:
+            repeat_by = batch_size
+        else:
+            # image batch size is the same as prompt batch size
+            repeat_by = num_images_per_prompt
+
+        image = image.repeat_interleave(repeat_by, dim=0)
+
+        image = image.to(device=device, dtype=dtype)
+
+        if do_classifier_free_guidance and not guess_mode:
+            image = torch.cat([image] * 2)
+
+        return image
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @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,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 28,
+        timesteps: List[int] = None,
+        guidance_scale: float = 7.0,
+        control_image: PipelineImageInput = None,
+        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+        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,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+        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
+            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.
+            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.
+            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.
+            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.flux.FluxPipelineOutput`] 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 512): Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] 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,
+            height,
+            width,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=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._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+
+        # 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
+        dtype = self.transformer.dtype
+
+        lora_scale = (
+            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            pooled_prompt_embeds,
+            text_ids,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+
+        # 3. Prepare control image
+        num_channels_latents = self.transformer.config.in_channels // 4
+        if isinstance(self.controlnet, FluxControlNetModel):
+            control_image = self.prepare_image(
+                image=control_image,
+                width=width,
+                height=height,
+                batch_size=batch_size * num_images_per_prompt,
+                num_images_per_prompt=num_images_per_prompt,
+                device=device,
+                dtype=dtype,
+            )
+            height, width = control_image.shape[-2:]
+
+            # vae encode
+            control_image = self.vae.encode(control_image).latent_dist.sample()
+            control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+
+            # pack
+            height_control_image, width_control_image = control_image.shape[2:]
+            control_image = self._pack_latents(
+                control_image,
+                batch_size * num_images_per_prompt,
+                num_channels_latents,
+                height_control_image,
+                width_control_image,
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels // 4
+        latents, latent_image_ids = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        image_seq_len = latents.shape[1]
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.base_image_seq_len,
+            self.scheduler.config.max_image_seq_len,
+            self.scheduler.config.base_shift,
+            self.scheduler.config.max_shift,
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas,
+            mu=mu,
+        )
+
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                # handle guidance
+                if self.transformer.config.guidance_embeds:
+                    guidance = torch.tensor([guidance_scale], device=device)
+                    guidance = guidance.expand(latents.shape[0])
+                else:
+                    guidance = None
+
+                # controlnet
+                controlnet_block_samples, controlnet_single_block_samples = self.controlnet(
+                    hidden_states=latents,
+                    controlnet_cond=control_image,
+                    conditioning_scale=controlnet_conditioning_scale,
+                    timestep=timestep / 1000,
+                    guidance=guidance,
+                    pooled_projections=pooled_prompt_embeds,
+                    encoder_hidden_states=prompt_embeds,
+                    txt_ids=text_ids,
+                    img_ids=latent_image_ids,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )
+
+                noise_pred = self.transformer(
+                    hidden_states=latents,
+                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
+                    timestep=timestep / 1000,
+                    guidance=guidance,
+                    pooled_projections=pooled_prompt_embeds,
+                    encoder_hidden_states=prompt_embeds,
+                    controlnet_block_samples=controlnet_block_samples,
+                    controlnet_single_block_samples=controlnet_single_block_samples,
+                    txt_ids=text_ids,
+                    img_ids=latent_image_ids,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # 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)
+
+                # 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 = self._unpack_latents(latents, height, width, self.vae_scale_factor)
+            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 not return_dict:
+            return (image,)
+
+        return FluxPipelineOutput(images=image)

src/diffusers/schedulers/__init__.py

@@ -43,12 +43,14 @@ else:
     _import_structure["scheduling_consistency_decoder"] = ["ConsistencyDecoderScheduler"]
     _import_structure["scheduling_consistency_models"] = ["CMStochasticIterativeScheduler"]
     _import_structure["scheduling_ddim"] = ["DDIMScheduler"]
+    _import_structure["scheduling_ddim_cogvideox"] = ["CogVideoXDDIMScheduler"]
     _import_structure["scheduling_ddim_inverse"] = ["DDIMInverseScheduler"]
     _import_structure["scheduling_ddim_parallel"] = ["DDIMParallelScheduler"]
     _import_structure["scheduling_ddpm"] = ["DDPMScheduler"]
     _import_structure["scheduling_ddpm_parallel"] = ["DDPMParallelScheduler"]
     _import_structure["scheduling_ddpm_wuerstchen"] = ["DDPMWuerstchenScheduler"]
     _import_structure["scheduling_deis_multistep"] = ["DEISMultistepScheduler"]
+    _import_structure["scheduling_dpm_cogvideox"] = ["CogVideoXDPMScheduler"]
     _import_structure["scheduling_dpmsolver_multistep"] = ["DPMSolverMultistepScheduler"]
     _import_structure["scheduling_dpmsolver_multistep_inverse"] = ["DPMSolverMultistepInverseScheduler"]
     _import_structure["scheduling_dpmsolver_singlestep"] = ["DPMSolverSinglestepScheduler"]
@@ -141,12 +143,14 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .scheduling_consistency_decoder import ConsistencyDecoderScheduler
         from .scheduling_consistency_models import CMStochasticIterativeScheduler
         from .scheduling_ddim import DDIMScheduler
+        from .scheduling_ddim_cogvideox import CogVideoXDDIMScheduler
         from .scheduling_ddim_inverse import DDIMInverseScheduler
         from .scheduling_ddim_parallel import DDIMParallelScheduler
         from .scheduling_ddpm import DDPMScheduler
         from .scheduling_ddpm_parallel import DDPMParallelScheduler
         from .scheduling_ddpm_wuerstchen import DDPMWuerstchenScheduler
         from .scheduling_deis_multistep import DEISMultistepScheduler
+        from .scheduling_dpm_cogvideox import CogVideoXDPMScheduler
         from .scheduling_dpmsolver_multistep import DPMSolverMultistepScheduler
         from .scheduling_dpmsolver_multistep_inverse import DPMSolverMultistepInverseScheduler
         from .scheduling_dpmsolver_singlestep import DPMSolverSinglestepScheduler

src/diffusers/schedulers/scheduling_ddim_cogvideox.py

@@ -0,0 +1,449 @@
+# 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.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM
+class DDIMSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+        pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
+            The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+            `pred_original_sample` can be used to preview progress or for guidance.
+    """
+
+    prev_sample: torch.Tensor
+    pred_original_sample: Optional[torch.Tensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+    (1-beta) over time from t = [0,1].
+
+    Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+    to that part of the diffusion process.
+
+
+    Args:
+        num_diffusion_timesteps (`int`): the number of betas to produce.
+        max_beta (`float`): the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
+
+    Returns:
+        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+    """
+    if alpha_transform_type == "cosine":
+
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
+
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+    return torch.tensor(betas, dtype=torch.float32)
+
+
+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)
+
+
+    Args:
+        betas (`torch.Tensor`):
+            the betas that the scheduler is being initialized with.
+
+    Returns:
+        `torch.Tensor`: rescaled betas with zero terminal SNR
+    """
+
+    alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+    # Store old values.
+    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+    # Shift so the last timestep is zero.
+    alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+    # Scale so the first timestep is back to the old value.
+    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+    # Convert alphas_bar_sqrt to betas
+    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
+
+    return alphas_bar
+
+
+class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin):
+    """
+    `DDIMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with
+    non-Markovian guidance.
+
+    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+    methods the library implements for all schedulers such as loading and saving.
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            The number of diffusion steps to train the model.
+        beta_start (`float`, defaults to 0.0001):
+            The starting `beta` value of inference.
+        beta_end (`float`, defaults to 0.02):
+            The final `beta` value.
+        beta_schedule (`str`, defaults to `"linear"`):
+            The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+        trained_betas (`np.ndarray`, *optional*):
+            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+        clip_sample (`bool`, defaults to `True`):
+            Clip the predicted sample for numerical stability.
+        clip_sample_range (`float`, defaults to 1.0):
+            The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+        set_alpha_to_one (`bool`, defaults to `True`):
+            Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
+            there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+            otherwise it uses the alpha value at step 0.
+        steps_offset (`int`, defaults to 0):
+            An offset added to the inference steps, as required by some model families.
+        prediction_type (`str`, defaults to `epsilon`, *optional*):
+            Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+            `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+            Video](https://imagen.research.google/video/paper.pdf) paper).
+        thresholding (`bool`, defaults to `False`):
+            Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+            as Stable Diffusion.
+        dynamic_thresholding_ratio (`float`, defaults to 0.995):
+            The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+        sample_max_value (`float`, defaults to 1.0):
+            The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+        timestep_spacing (`str`, defaults to `"leading"`):
+            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+        rescale_betas_zero_snr (`bool`, defaults to `False`):
+            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+            dark samples instead of limiting it to samples with medium brightness. Loosely related to
+            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+    """
+
+    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        beta_start: float = 0.00085,
+        beta_end: float = 0.0120,
+        beta_schedule: str = "scaled_linear",
+        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+        clip_sample: bool = True,
+        set_alpha_to_one: bool = True,
+        steps_offset: int = 0,
+        prediction_type: str = "epsilon",
+        clip_sample_range: float = 1.0,
+        sample_max_value: float = 1.0,
+        timestep_spacing: str = "leading",
+        rescale_betas_zero_snr: bool = False,
+        snr_shift_scale: float = 3.0,
+    ):
+        if trained_betas is not None:
+            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+        elif beta_schedule == "linear":
+            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+        elif beta_schedule == "scaled_linear":
+            # this schedule is very specific to the latent diffusion model.
+            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float64) ** 2
+        elif beta_schedule == "squaredcos_cap_v2":
+            # Glide cosine schedule
+            self.betas = betas_for_alpha_bar(num_train_timesteps)
+        else:
+            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
+
+        self.alphas = 1.0 - self.betas
+        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+        # Modify: SNR shift following SD3
+        self.alphas_cumprod = self.alphas_cumprod / (snr_shift_scale + (1 - snr_shift_scale) * self.alphas_cumprod)
+
+        # Rescale for zero SNR
+        if rescale_betas_zero_snr:
+            self.alphas_cumprod = rescale_zero_terminal_snr(self.alphas_cumprod)
+
+        # At every step in ddim, we are looking into the previous alphas_cumprod
+        # For the final step, there is no previous alphas_cumprod because we are already at 0
+        # `set_alpha_to_one` decides whether we set this parameter simply to one or
+        # whether we use the final alpha of the "non-previous" one.
+        self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+        # standard deviation of the initial noise distribution
+        self.init_noise_sigma = 1.0
+
+        # setable values
+        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
+        current timestep.
+
+        Args:
+            sample (`torch.Tensor`):
+                The input sample.
+            timestep (`int`, *optional*):
+                The current timestep in the diffusion chain.
+
+        Returns:
+            `torch.Tensor`:
+                A scaled input sample.
+        """
+        return sample
+
+    def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+        """
+        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+        Args:
+            num_inference_steps (`int`):
+                The number of diffusion steps used when generating samples with a pre-trained model.
+        """
+
+        if num_inference_steps > self.config.num_train_timesteps:
+            raise ValueError(
+                f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+                f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+                f" maximal {self.config.num_train_timesteps} timesteps."
+            )
+
+        self.num_inference_steps = num_inference_steps
+
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        if self.config.timestep_spacing == "linspace":
+            timesteps = (
+                np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
+                .round()[::-1]
+                .copy()
+                .astype(np.int64)
+            )
+        elif self.config.timestep_spacing == "leading":
+            step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
+            timesteps += self.config.steps_offset
+        elif self.config.timestep_spacing == "trailing":
+            step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64)
+            timesteps -= 1
+        else:
+            raise ValueError(
+                f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'."
+            )
+
+        self.timesteps = torch.from_numpy(timesteps).to(device)
+
+    def step(
+        self,
+        model_output: torch.Tensor,
+        timestep: int,
+        sample: torch.Tensor,
+        eta: float = 0.0,
+        use_clipped_model_output: bool = False,
+        generator=None,
+        variance_noise: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> 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).
+
+        Args:
+            model_output (`torch.Tensor`):
+                The direct output from learned diffusion model.
+            timestep (`float`):
+                The current discrete timestep in the diffusion chain.
+            sample (`torch.Tensor`):
+                A current instance of a sample created by the diffusion process.
+            eta (`float`):
+                The weight of noise for added noise in diffusion step.
+            use_clipped_model_output (`bool`, defaults to `False`):
+                If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary
+                because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no
+                clipping has happened, "corrected" `model_output` would coincide with the one provided as input and
+                `use_clipped_model_output` has no effect.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            variance_noise (`torch.Tensor`):
+                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.DDIMSchedulerOutput`] or `tuple`.
+
+        Returns:
+            [`~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:
+            raise ValueError(
+                "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+            )
+
+        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # Ideally, read DDIM paper in-detail understanding
+
+        # Notation (<variable name> -> <name in paper>
+        # - pred_noise_t -> e_theta(x_t, t)
+        # - pred_original_sample -> f_theta(x_t, t) or x_0
+        # - std_dev_t -> sigma_t
+        # - eta -> η
+        # - pred_sample_direction -> "direction pointing to x_t"
+        # - pred_prev_sample -> "x_t-1"
+
+        # 1. get previous step value (=t-1)
+        prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
+
+        # 2. compute alphas, betas
+        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
+
+        # 3. compute predicted original sample from predicted noise also called
+        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # To make style tests pass, commented out `pred_epsilon` as it is an unused variable
+        if self.config.prediction_type == "epsilon":
+            pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+            # pred_epsilon = model_output
+        elif self.config.prediction_type == "sample":
+            pred_original_sample = model_output
+            # pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+        elif self.config.prediction_type == "v_prediction":
+            pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+            # pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+        else:
+            raise ValueError(
+                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+                " `v_prediction`"
+            )
+
+        a_t = ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) ** 0.5
+        b_t = alpha_prod_t_prev**0.5 - alpha_prod_t**0.5 * a_t
+
+        prev_sample = a_t * sample + b_t * pred_original_sample
+
+        if not return_dict:
+            return (prev_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(
+        self,
+        original_samples: torch.Tensor,
+        noise: torch.Tensor,
+        timesteps: torch.IntTensor,
+    ) -> torch.Tensor:
+        # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+        # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+        # for the subsequent add_noise calls
+        self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+        alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+        timesteps = timesteps.to(original_samples.device)
+
+        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+        while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+        while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+        noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+        return noisy_samples
+
+    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity
+    def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor:
+        # Make sure alphas_cumprod and timestep have same device and dtype as sample
+        self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device)
+        alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype)
+        timesteps = timesteps.to(sample.device)
+
+        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+        while len(sqrt_alpha_prod.shape) < len(sample.shape):
+            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+        while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+        velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+        return velocity
+
+    def __len__(self):
+        return self.config.num_train_timesteps

src/diffusers/schedulers/scheduling_dpm_cogvideox.py

@@ -0,0 +1,489 @@
+# 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.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM
+class DDIMSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+        pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
+            The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+            `pred_original_sample` can be used to preview progress or for guidance.
+    """
+
+    prev_sample: torch.Tensor
+    pred_original_sample: Optional[torch.Tensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+    (1-beta) over time from t = [0,1].
+
+    Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+    to that part of the diffusion process.
+
+
+    Args:
+        num_diffusion_timesteps (`int`): the number of betas to produce.
+        max_beta (`float`): the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
+
+    Returns:
+        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+    """
+    if alpha_transform_type == "cosine":
+
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
+
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+    return torch.tensor(betas, dtype=torch.float32)
+
+
+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)
+
+
+    Args:
+        betas (`torch.Tensor`):
+            the betas that the scheduler is being initialized with.
+
+    Returns:
+        `torch.Tensor`: rescaled betas with zero terminal SNR
+    """
+
+    alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+    # Store old values.
+    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+    # Shift so the last timestep is zero.
+    alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+    # Scale so the first timestep is back to the old value.
+    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+    # Convert alphas_bar_sqrt to betas
+    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
+
+    return alphas_bar
+
+
+class CogVideoXDPMScheduler(SchedulerMixin, ConfigMixin):
+    """
+    `DDIMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with
+    non-Markovian guidance.
+
+    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+    methods the library implements for all schedulers such as loading and saving.
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            The number of diffusion steps to train the model.
+        beta_start (`float`, defaults to 0.0001):
+            The starting `beta` value of inference.
+        beta_end (`float`, defaults to 0.02):
+            The final `beta` value.
+        beta_schedule (`str`, defaults to `"linear"`):
+            The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+        trained_betas (`np.ndarray`, *optional*):
+            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+        clip_sample (`bool`, defaults to `True`):
+            Clip the predicted sample for numerical stability.
+        clip_sample_range (`float`, defaults to 1.0):
+            The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+        set_alpha_to_one (`bool`, defaults to `True`):
+            Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
+            there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+            otherwise it uses the alpha value at step 0.
+        steps_offset (`int`, defaults to 0):
+            An offset added to the inference steps, as required by some model families.
+        prediction_type (`str`, defaults to `epsilon`, *optional*):
+            Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+            `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+            Video](https://imagen.research.google/video/paper.pdf) paper).
+        thresholding (`bool`, defaults to `False`):
+            Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+            as Stable Diffusion.
+        dynamic_thresholding_ratio (`float`, defaults to 0.995):
+            The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+        sample_max_value (`float`, defaults to 1.0):
+            The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+        timestep_spacing (`str`, defaults to `"leading"`):
+            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+        rescale_betas_zero_snr (`bool`, defaults to `False`):
+            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+            dark samples instead of limiting it to samples with medium brightness. Loosely related to
+            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+    """
+
+    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        beta_start: float = 0.00085,
+        beta_end: float = 0.0120,
+        beta_schedule: str = "scaled_linear",
+        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+        clip_sample: bool = True,
+        set_alpha_to_one: bool = True,
+        steps_offset: int = 0,
+        prediction_type: str = "epsilon",
+        clip_sample_range: float = 1.0,
+        sample_max_value: float = 1.0,
+        timestep_spacing: str = "leading",
+        rescale_betas_zero_snr: bool = False,
+        snr_shift_scale: float = 3.0,
+    ):
+        if trained_betas is not None:
+            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+        elif beta_schedule == "linear":
+            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+        elif beta_schedule == "scaled_linear":
+            # this schedule is very specific to the latent diffusion model.
+            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float64) ** 2
+        elif beta_schedule == "squaredcos_cap_v2":
+            # Glide cosine schedule
+            self.betas = betas_for_alpha_bar(num_train_timesteps)
+        else:
+            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
+
+        self.alphas = 1.0 - self.betas
+        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+        # Modify: SNR shift following SD3
+        self.alphas_cumprod = self.alphas_cumprod / (snr_shift_scale + (1 - snr_shift_scale) * self.alphas_cumprod)
+
+        # Rescale for zero SNR
+        if rescale_betas_zero_snr:
+            self.alphas_cumprod = rescale_zero_terminal_snr(self.alphas_cumprod)
+
+        # At every step in ddim, we are looking into the previous alphas_cumprod
+        # For the final step, there is no previous alphas_cumprod because we are already at 0
+        # `set_alpha_to_one` decides whether we set this parameter simply to one or
+        # whether we use the final alpha of the "non-previous" one.
+        self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+        # standard deviation of the initial noise distribution
+        self.init_noise_sigma = 1.0
+
+        # setable values
+        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
+        current timestep.
+
+        Args:
+            sample (`torch.Tensor`):
+                The input sample.
+            timestep (`int`, *optional*):
+                The current timestep in the diffusion chain.
+
+        Returns:
+            `torch.Tensor`:
+                A scaled input sample.
+        """
+        return sample
+
+    def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+        """
+        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+        Args:
+            num_inference_steps (`int`):
+                The number of diffusion steps used when generating samples with a pre-trained model.
+        """
+
+        if num_inference_steps > self.config.num_train_timesteps:
+            raise ValueError(
+                f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+                f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+                f" maximal {self.config.num_train_timesteps} timesteps."
+            )
+
+        self.num_inference_steps = num_inference_steps
+
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        if self.config.timestep_spacing == "linspace":
+            timesteps = (
+                np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
+                .round()[::-1]
+                .copy()
+                .astype(np.int64)
+            )
+        elif self.config.timestep_spacing == "leading":
+            step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
+            timesteps += self.config.steps_offset
+        elif self.config.timestep_spacing == "trailing":
+            step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64)
+            timesteps -= 1
+        else:
+            raise ValueError(
+                f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'."
+            )
+
+        self.timesteps = torch.from_numpy(timesteps).to(device)
+
+    def get_variables(self, alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back=None):
+        lamb = ((alpha_prod_t / (1 - alpha_prod_t)) ** 0.5).log()
+        lamb_next = ((alpha_prod_t_prev / (1 - alpha_prod_t_prev)) ** 0.5).log()
+        h = lamb_next - lamb
+
+        if alpha_prod_t_back is not None:
+            lamb_previous = ((alpha_prod_t_back / (1 - alpha_prod_t_back)) ** 0.5).log()
+            h_last = lamb - lamb_previous
+            r = h_last / h
+            return h, r, lamb, lamb_next
+        else:
+            return h, None, lamb, lamb_next
+
+    def get_mult(self, h, r, alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back):
+        mult1 = ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) ** 0.5 * (-h).exp()
+        mult2 = (-2 * h).expm1() * alpha_prod_t_prev**0.5
+
+        if alpha_prod_t_back is not None:
+            mult3 = 1 + 1 / (2 * r)
+            mult4 = 1 / (2 * r)
+            return mult1, mult2, mult3, mult4
+        else:
+            return mult1, mult2
+
+    def step(
+        self,
+        model_output: torch.Tensor,
+        old_pred_original_sample: torch.Tensor,
+        timestep: int,
+        timestep_back: int,
+        sample: torch.Tensor,
+        eta: float = 0.0,
+        use_clipped_model_output: bool = False,
+        generator=None,
+        variance_noise: Optional[torch.Tensor] = None,
+        return_dict: bool = False,
+    ) -> 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).
+
+        Args:
+            model_output (`torch.Tensor`):
+                The direct output from learned diffusion model.
+            timestep (`float`):
+                The current discrete timestep in the diffusion chain.
+            sample (`torch.Tensor`):
+                A current instance of a sample created by the diffusion process.
+            eta (`float`):
+                The weight of noise for added noise in diffusion step.
+            use_clipped_model_output (`bool`, defaults to `False`):
+                If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary
+                because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no
+                clipping has happened, "corrected" `model_output` would coincide with the one provided as input and
+                `use_clipped_model_output` has no effect.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            variance_noise (`torch.Tensor`):
+                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.DDIMSchedulerOutput`] or `tuple`.
+
+        Returns:
+            [`~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:
+            raise ValueError(
+                "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+            )
+
+        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # Ideally, read DDIM paper in-detail understanding
+
+        # Notation (<variable name> -> <name in paper>
+        # - pred_noise_t -> e_theta(x_t, t)
+        # - pred_original_sample -> f_theta(x_t, t) or x_0
+        # - std_dev_t -> sigma_t
+        # - eta -> η
+        # - pred_sample_direction -> "direction pointing to x_t"
+        # - pred_prev_sample -> "x_t-1"
+
+        # 1. get previous step value (=t-1)
+        prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
+
+        # 2. compute alphas, betas
+        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
+        alpha_prod_t_back = self.alphas_cumprod[timestep_back] if timestep_back is not None else None
+
+        beta_prod_t = 1 - alpha_prod_t
+
+        # 3. compute predicted original sample from predicted noise also called
+        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # To make style tests pass, commented out `pred_epsilon` as it is an unused variable
+        if self.config.prediction_type == "epsilon":
+            pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+            # pred_epsilon = model_output
+        elif self.config.prediction_type == "sample":
+            pred_original_sample = model_output
+            # pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+        elif self.config.prediction_type == "v_prediction":
+            pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+            # pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+        else:
+            raise ValueError(
+                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+                " `v_prediction`"
+            )
+
+        h, r, lamb, lamb_next = self.get_variables(alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back)
+        mult = list(self.get_mult(h, r, alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back))
+        mult_noise = (1 - alpha_prod_t_prev) ** 0.5 * (1 - (-2 * h).exp()) ** 0.5
+
+        noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype)
+        prev_sample = mult[0] * sample - mult[1] * pred_original_sample + mult_noise * noise
+
+        if old_pred_original_sample is None or prev_timestep < 0:
+            # Save a network evaluation if all noise levels are 0 or on the first step
+            return prev_sample, pred_original_sample
+        else:
+            denoised_d = mult[2] * pred_original_sample - mult[3] * old_pred_original_sample
+            noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype)
+            x_advanced = mult[0] * sample - mult[1] * denoised_d + mult_noise * noise
+
+            prev_sample = x_advanced
+
+        if not return_dict:
+            return (prev_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(
+        self,
+        original_samples: torch.Tensor,
+        noise: torch.Tensor,
+        timesteps: torch.IntTensor,
+    ) -> torch.Tensor:
+        # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+        # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+        # for the subsequent add_noise calls
+        self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+        alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+        timesteps = timesteps.to(original_samples.device)
+
+        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+        while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+        while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+        noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+        return noisy_samples
+
+    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity
+    def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor:
+        # Make sure alphas_cumprod and timestep have same device and dtype as sample
+        self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device)
+        alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype)
+        timesteps = timesteps.to(sample.device)
+
+        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+        while len(sqrt_alpha_prod.shape) < len(sample.shape):
+            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+        while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+        velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+        return velocity
+
+    def __len__(self):
+        return self.config.num_train_timesteps

src/diffusers/utils/dummy_pt_objects.py

@@ -47,6 +47,21 @@ class AutoencoderKL(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class AutoencoderKLCogVideoX(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class AutoencoderKLTemporalDecoder(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -92,6 +107,21 @@ class AutoencoderTiny(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class CogVideoXTransformer3DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class ConsistencyDecoderVAE(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -152,6 +182,21 @@ class DiTTransformer2DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class FluxControlNetModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class FluxTransformer2DModel(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -975,6 +1020,36 @@ class CMStochasticIterativeScheduler(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class CogVideoXDDIMScheduler(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
+class CogVideoXDPMScheduler(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class DDIMInverseScheduler(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -257,6 +257,21 @@ class CLIPImageProjection(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class CogVideoXPipeline(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 CycleDiffusionPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 
@@ -272,6 +287,21 @@ class CycleDiffusionPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class FluxControlNetPipeline(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 FluxPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

src/diffusers/utils/export_utils.py

@@ -9,7 +9,7 @@ import numpy as np
 import PIL.Image
 import PIL.ImageOps
 
-from .import_utils import BACKENDS_MAPPING, is_opencv_available
+from .import_utils import BACKENDS_MAPPING, is_imageio_available, is_opencv_available
 from .logging import get_logger
 
 
@@ -112,9 +112,9 @@ def export_to_obj(mesh, output_obj_path: str = None):
         f.writelines("\n".join(combined_data))
 
 
-def export_to_video(
+def _legacy_export_to_video(
     video_frames: Union[List[np.ndarray], List[PIL.Image.Image]], output_video_path: str = None, fps: int = 10
-) -> str:
+):
     if is_opencv_available():
         import cv2
     else:
@@ -134,4 +134,51 @@ def export_to_video(
     for i in range(len(video_frames)):
         img = cv2.cvtColor(video_frames[i], cv2.COLOR_RGB2BGR)
         video_writer.write(img)
+
+    return output_video_path
+
+
+def export_to_video(
+    video_frames: Union[List[np.ndarray], List[PIL.Image.Image]], output_video_path: str = None, fps: int = 10
+) -> str:
+    # TODO: Dhruv. Remove by Diffusers release 0.33.0
+    # Added to prevent breaking existing code
+    if not is_imageio_available():
+        logger.warning(
+            (
+                "It is recommended to use `export_to_video` with `imageio` and `imageio-ffmpeg` as a backend. \n"
+                "These libraries are not present in your environment. Attempting to use legacy OpenCV backend to export video. \n"
+                "Support for the OpenCV backend will be deprecated in a future Diffusers version"
+            )
+        )
+        return _legacy_export_to_video(video_frames, output_video_path, fps)
+
+    if is_imageio_available():
+        import imageio
+    else:
+        raise ImportError(BACKENDS_MAPPING["imageio"][1].format("export_to_video"))
+
+    try:
+        imageio.plugins.ffmpeg.get_exe()
+    except AttributeError:
+        raise AttributeError(
+            (
+                "Found an existing imageio backend in your environment. Attempting to export video with imageio. \n"
+                "Unable to find a compatible ffmpeg installation in your environment to use with imageio. Please install via `pip install imageio-ffmpeg"
+            )
+        )
+
+    if output_video_path is None:
+        output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name
+
+    if isinstance(video_frames[0], np.ndarray):
+        video_frames = [(frame * 255).astype(np.uint8) for frame in video_frames]
+
+    elif isinstance(video_frames[0], PIL.Image.Image):
+        video_frames = [np.array(frame) for frame in video_frames]
+
+    with imageio.get_writer(output_video_path, fps=fps) as writer:
+        for frame in video_frames:
+            writer.append_data(frame)
+
     return output_video_path

src/diffusers/utils/import_utils.py

@@ -330,6 +330,15 @@ except importlib_metadata.PackageNotFoundError:
 
 _is_google_colab = "google.colab" in sys.modules or any(k.startswith("COLAB_") for k in os.environ)
 
+_imageio_available = importlib.util.find_spec("imageio") is not None
+if _imageio_available:
+    try:
+        _imageio_version = importlib_metadata.version("imageio")
+        logger.debug(f"Successfully imported imageio version {_imageio_version}")
+
+    except importlib_metadata.PackageNotFoundError:
+        _imageio_available = False
+
 
 def is_torch_available():
     return _torch_available
@@ -447,6 +456,10 @@ def is_sentencepiece_available():
     return _sentencepiece_available
 
 
+def is_imageio_available():
+    return _imageio_available
+
+
 # docstyle-ignore
 FLAX_IMPORT_ERROR = """
 {0} requires the FLAX library but it was not found in your environment. Checkout the instructions on the
@@ -575,6 +588,11 @@ 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`
 """
 
+# docstyle-ignore
+IMAGEIO_IMPORT_ERROR = """
+{0} requires the imageio library and ffmpeg but it was not found in your environment. You can install it with pip: `pip install imageio imageio-ffmpeg`
+"""
+
 BACKENDS_MAPPING = OrderedDict(
     [
         ("bs4", (is_bs4_available, BS4_IMPORT_ERROR)),
@@ -599,6 +617,7 @@ BACKENDS_MAPPING = OrderedDict(
         ("safetensors", (is_safetensors_available, SAFETENSORS_IMPORT_ERROR)),
         ("bitsandbytes", (is_bitsandbytes_available, BITSANDBYTES_IMPORT_ERROR)),
         ("sentencepiece", (is_sentencepiece_available, SENTENCEPIECE_IMPORT_ERROR)),
+        ("imageio", (is_imageio_available, IMAGEIO_IMPORT_ERROR)),
     ]
 )
 

src/diffusers/utils/loading_utils.py

@@ -1,12 +1,13 @@
 import os
 import tempfile
 from typing import Callable, List, Optional, Union
+from urllib.parse import unquote, urlparse
 
 import PIL.Image
 import PIL.ImageOps
 import requests
 
-from .import_utils import BACKENDS_MAPPING, is_opencv_available
+from .import_utils import BACKENDS_MAPPING, is_imageio_available
 
 
 def load_image(
@@ -80,11 +81,22 @@ def load_video(
         )
 
     if is_url:
-        video_data = requests.get(video, stream=True).raw
-        video_path = tempfile.NamedTemporaryFile(suffix=os.path.splitext(video)[1], delete=False).name
+        response = requests.get(video, stream=True)
+        if response.status_code != 200:
+            raise ValueError(f"Failed to download video. Status code: {response.status_code}")
+
+        parsed_url = urlparse(video)
+        file_name = os.path.basename(unquote(parsed_url.path))
+
+        suffix = os.path.splitext(file_name)[1] or ".mp4"
+        video_path = tempfile.NamedTemporaryFile(suffix=suffix, delete=False).name
+
         was_tempfile_created = True
+
+        video_data = response.iter_content(chunk_size=8192)
         with open(video_path, "wb") as f:
-            f.write(video_data.read())
+            for chunk in video_data:
+                f.write(chunk)
 
         video = video_path
 
@@ -99,19 +111,22 @@ def load_video(
             pass
 
     else:
-        if is_opencv_available():
-            import cv2
+        if is_imageio_available():
+            import imageio
         else:
-            raise ImportError(BACKENDS_MAPPING["opencv"][1].format("load_video"))
+            raise ImportError(BACKENDS_MAPPING["imageio"][1].format("load_video"))
 
-        video_capture = cv2.VideoCapture(video)
-        success, frame = video_capture.read()
-        while success:
-            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
-            pil_images.append(PIL.Image.fromarray(frame))
-            success, frame = video_capture.read()
+        try:
+            imageio.plugins.ffmpeg.get_exe()
+        except AttributeError:
+            raise AttributeError(
+                "`Unable to find an ffmpeg installation on your machine. Please install via `pip install imageio-ffmpeg"
+            )
 
-        video_capture.release()
+        with imageio.get_reader(video) as reader:
+            # Read all frames
+            for frame in reader:
+                pil_images.append(PIL.Image.fromarray(frame))
 
     if was_tempfile_created:
         os.remove(video_path)

tests/models/transformers/test_models_transformer_cogvideox.py

@@ -0,0 +1,82 @@
+# 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 CogVideoXTransformer3DModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    torch_device,
+)
+
+from ..test_modeling_common import ModelTesterMixin
+
+
+enable_full_determinism()
+
+
+class CogVideoXTransformerTests(ModelTesterMixin, unittest.TestCase):
+    model_class = CogVideoXTransformer3DModel
+    main_input_name = "hidden_states"
+
+    @property
+    def dummy_input(self):
+        batch_size = 2
+        num_channels = 4
+        num_frames = 1
+        height = 8
+        width = 8
+        embedding_dim = 8
+        sequence_length = 8
+
+        hidden_states = torch.randn((batch_size, num_frames, num_channels, height, width)).to(torch_device)
+        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
+        timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "timestep": timestep,
+        }
+
+    @property
+    def input_shape(self):
+        return (1, 4, 8, 8)
+
+    @property
+    def output_shape(self):
+        return (1, 4, 8, 8)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
+            # Product of num_attention_heads * attention_head_dim must be divisible by 16 for 3D positional embeddings.
+            "num_attention_heads": 2,
+            "attention_head_dim": 8,
+            "in_channels": 4,
+            "out_channels": 4,
+            "time_embed_dim": 2,
+            "text_embed_dim": 8,
+            "num_layers": 1,
+            "sample_width": 8,
+            "sample_height": 8,
+            "sample_frames": 8,
+            "patch_size": 2,
+            "temporal_compression_ratio": 4,
+            "max_text_seq_length": 8,
+        }
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict

tests/pipelines/cogvideox/test_cogvideox.py

@@ -0,0 +1,315 @@
+# Copyright 2024 The HuggingFace Team.
+#
+# 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 gc
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, T5EncoderModel
+
+from diffusers import AutoencoderKLCogVideoX, CogVideoXPipeline, CogVideoXTransformer3DModel, DDIMScheduler
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+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
+
+
+enable_full_determinism()
+
+
+class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = CogVideoXPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    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 = 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(
+            # 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,
+            in_channels=4,
+            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 = 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,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        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": "dance monkey",
+            "negative_prompt": "",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            # Cannot reduce because convolution kernel becomes bigger than sample
+            "height": 16,
+            "width": 16,
+            "num_frames": 8,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+        }
+        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)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+
+        self.assertEqual(generated_video.shape, (8, 3, 16, 16))
+        expected_video = torch.randn(8, 3, 16, 16)
+        max_diff = np.abs(generated_video - expected_video).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    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, 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)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        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",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_overlap_factor_height=1 / 12,
+            tile_overlap_factor_width=1 / 12,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )
+
+
+@slow
+@require_torch_gpu
+class CogVideoXPipelineIntegrationTests(unittest.TestCase):
+    prompt = "A painting of a squirrel eating a burger."
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_cogvideox(self):
+        generator = torch.Generator("cpu").manual_seed(0)
+
+        pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16)
+        pipe.enable_model_cpu_offload()
+        prompt = self.prompt
+
+        videos = pipe(
+            prompt=prompt,
+            height=480,
+            width=720,
+            num_frames=16,
+            generator=generator,
+            num_inference_steps=2,
+            output_type="pt",
+        ).frames
+
+        video = videos[0]
+        expected_video = torch.randn(1, 16, 480, 720, 3).numpy()
+
+        max_diff = numpy_cosine_similarity_distance(video, expected_video)
+        assert max_diff < 1e-3, f"Max diff is too high. got {video}"

tests/pipelines/controlnet_flux/test_controlnet_flux.py

@@ -0,0 +1,233 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc and The InstantX Team.
+#
+# 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 gc
+import unittest
+
+import numpy as np
+import torch
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
+
+from diffusers import (
+    AutoencoderKL,
+    FlowMatchEulerDiscreteScheduler,
+    FluxControlNetPipeline,
+    FluxTransformer2DModel,
+)
+from diffusers.models import FluxControlNetModel
+from diffusers.utils import load_image
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+from diffusers.utils.torch_utils import randn_tensor
+
+from ..test_pipelines_common import PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+class FluxControlNetPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
+    pipeline_class = FluxControlNetPipeline
+
+    params = frozenset(["prompt", "height", "width", "guidance_scale", "prompt_embeds", "pooled_prompt_embeds"])
+    batch_params = frozenset(["prompt"])
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = FluxTransformer2DModel(
+            patch_size=1,
+            in_channels=16,
+            num_layers=1,
+            num_single_layers=1,
+            attention_head_dim=16,
+            num_attention_heads=2,
+            joint_attention_dim=32,
+            pooled_projection_dim=32,
+            axes_dims_rope=[4, 4, 8],
+        )
+
+        torch.manual_seed(0)
+        controlnet = FluxControlNetModel(
+            patch_size=1,
+            in_channels=16,
+            num_layers=1,
+            num_single_layers=1,
+            attention_head_dim=16,
+            num_attention_heads=2,
+            joint_attention_dim=32,
+            pooled_projection_dim=32,
+            axes_dims_rope=[4, 4, 8],
+        )
+
+        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 = CLIPTextModel(clip_text_encoder_config)
+
+        torch.manual_seed(0)
+        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+        tokenizer_2 = T5TokenizerFast.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,
+            "tokenizer": tokenizer,
+            "tokenizer_2": tokenizer_2,
+            "transformer": transformer,
+            "vae": vae,
+            "controlnet": controlnet,
+        }
+
+    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)
+
+        control_image = randn_tensor(
+            (1, 3, 32, 32),
+            generator=generator,
+            device=torch.device(device),
+            dtype=torch.float16,
+        )
+
+        controlnet_conditioning_scale = 0.5
+
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 3.5,
+            "output_type": "np",
+            "control_image": control_image,
+            "controlnet_conditioning_scale": controlnet_conditioning_scale,
+        }
+
+        return inputs
+
+    def test_controlnet_flux(self):
+        components = self.get_dummy_components()
+        flux_pipe = FluxControlNetPipeline(**components)
+        flux_pipe = flux_pipe.to(torch_device, dtype=torch.float16)
+        flux_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        output = flux_pipe(**inputs)
+        image = output.images
+
+        image_slice = image[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 32, 32, 3)
+
+        expected_slice = np.array(
+            [0.7348633, 0.41333008, 0.6621094, 0.5444336, 0.47607422, 0.5859375, 0.44677734, 0.4506836, 0.40454102]
+        )
+
+        assert (
+            np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+        ), f"Expected: {expected_slice}, got: {image_slice.flatten()}"
+
+    @unittest.skip("xFormersAttnProcessor does not work with SD3 Joint Attention")
+    def test_xformers_attention_forwardGenerator_pass(self):
+        pass
+
+
+@slow
+@require_torch_gpu
+class FluxControlNetPipelineSlowTests(unittest.TestCase):
+    pipeline_class = FluxControlNetPipeline
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_canny(self):
+        controlnet = FluxControlNetModel.from_pretrained(
+            "InstantX/FLUX.1-dev-Controlnet-Canny-alpha", torch_dtype=torch.bfloat16
+        )
+        pipe = FluxControlNetPipeline.from_pretrained(
+            "black-forest-labs/FLUX.1-dev", controlnet=controlnet, torch_dtype=torch.bfloat16
+        )
+        pipe.enable_model_cpu_offload()
+        pipe.set_progress_bar_config(disable=None)
+
+        generator = torch.Generator(device="cpu").manual_seed(0)
+        prompt = "A girl in city, 25 years old, cool, futuristic"
+        control_image = load_image(
+            "https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny-alpha/resolve/main/canny.jpg"
+        )
+
+        output = pipe(
+            prompt,
+            control_image=control_image,
+            controlnet_conditioning_scale=0.6,
+            num_inference_steps=2,
+            guidance_scale=3.5,
+            output_type="np",
+            generator=generator,
+        )
+
+        image = output.images[0]
+
+        assert image.shape == (1024, 1024, 3)
+
+        original_image = image[-3:, -3:, -1].flatten()
+
+        expected_image = np.array(
+            [0.33007812, 0.33984375, 0.33984375, 0.328125, 0.34179688, 0.33984375, 0.30859375, 0.3203125, 0.3203125]
+        )
+
+        assert np.abs(original_image.flatten() - expected_image).max() < 1e-2

tests/pipelines/controlnet_sd3/test_controlnet_inpaint_sd3.py

@@ -0,0 +1,201 @@
+# 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 numpy as np
+import torch
+from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
+
+from diffusers import (
+    AutoencoderKL,
+    FlowMatchEulerDiscreteScheduler,
+    SD3Transformer2DModel,
+    StableDiffusion3ControlNetInpaintingPipeline,
+)
+from diffusers.models import SD3ControlNetModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    torch_device,
+)
+from diffusers.utils.torch_utils import randn_tensor
+
+from ..test_pipelines_common import PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+class StableDiffusion3ControlInpaintNetPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
+    pipeline_class = StableDiffusion3ControlNetInpaintingPipeline
+    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=8,
+            num_layers=4,
+            attention_head_dim=8,
+            num_attention_heads=4,
+            joint_attention_dim=32,
+            caption_projection_dim=32,
+            pooled_projection_dim=64,
+            out_channels=8,
+        )
+
+        torch.manual_seed(0)
+        controlnet = SD3ControlNetModel(
+            sample_size=32,
+            patch_size=1,
+            in_channels=8,
+            num_layers=1,
+            attention_head_dim=8,
+            num_attention_heads=4,
+            joint_attention_dim=32,
+            caption_projection_dim=32,
+            pooled_projection_dim=64,
+            out_channels=8,
+            extra_conditioning_channels=1,
+        )
+        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)
+
+        torch.manual_seed(0)
+        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=8,
+            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,
+            "controlnet": controlnet,
+        }
+
+    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)
+
+        control_image = randn_tensor(
+            (1, 3, 32, 32),
+            generator=generator,
+            device=torch.device(device),
+            dtype=torch.float16,
+        )
+
+        control_mask = randn_tensor(
+            (1, 1, 32, 32),
+            generator=generator,
+            device=torch.device(device),
+            dtype=torch.float16,
+        )
+
+        controlnet_conditioning_scale = 0.95
+
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 7.0,
+            "output_type": "np",
+            "control_image": control_image,
+            "control_mask": control_mask,
+            "controlnet_conditioning_scale": controlnet_conditioning_scale,
+        }
+
+        return inputs
+
+    def test_controlnet_inpaint_sd3(self):
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusion3ControlNetInpaintingPipeline(**components)
+        sd_pipe = sd_pipe.to(torch_device, dtype=torch.float16)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        output = sd_pipe(**inputs)
+        image = output.images
+
+        image_slice = image[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 32, 32, 3)
+
+        expected_slice = np.array(
+            [0.51708984, 0.7421875, 0.4580078, 0.6435547, 0.65625, 0.43603516, 0.5151367, 0.65722656, 0.60839844]
+        )
+
+        assert (
+            np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+        ), f"Expected: {expected_slice}, got: {image_slice.flatten()}"
+
+    @unittest.skip("xFormersAttnProcessor does not work with SD3 Joint Attention")
+    def test_xformers_attention_forwardGenerator_pass(self):
+        pass