update

2024-08-07 07:03:12 +00:00 · 2024-08-07 06:54:13 +00:00
140 changed files with 374 additions and 13639 deletions
@@ -202,7 +202,6 @@ 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
@@ -223,76 +223,64 @@
    sections:
    - local: api/models/overview
      title: Overview
-    - 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
+    - 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
    title: Models
  - isExpanded: false
    sections:
@@ -314,8 +302,6 @@
      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
@@ -22,7 +22,6 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:

 ## Supported pipelines

- [`CogVideoXPipeline`]
 - [`StableDiffusionPipeline`]
 - [`StableDiffusionImg2ImgPipeline`]
 - [`StableDiffusionInpaintPipeline`]
@@ -50,7 +49,6 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:
 - [`UNet2DConditionModel`]
 - [`StableCascadeUNet`]
 - [`AutoencoderKL`]
- [`AutoencoderKLCogVideoX`]
 - [`ControlNetModel`]
 - [`SD3Transformer2DModel`]
 - [`FluxTransformer2DModel`]
@@ -1,37 +0,0 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License. -->
-
-# 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
@@ -1,30 +0,0 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License. -->
-
-# 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
@@ -1,88 +0,0 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-->
-
-# CogVideoX
-
-[CogVideoX: 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
@@ -1,4 +1,4 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+<!--Copyright 2023 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
@@ -22,16 +22,7 @@ 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 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) |
-
+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.

 <Tip>

@@ -44,10 +35,5 @@ 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
@@ -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()
@@ -77,34 +77,6 @@ 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.
@@ -162,4 +134,4 @@ image.save("flux-fp8-dev.png")

 [[autodoc]] FluxPipeline
 	- all
-	- __call__
+	- __call__
@@ -14,7 +14,7 @@ specific language governing permissions and limitations under the License.

 ![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/kolors/kolors_header_collage.png)

-Kolors is a large-scale text-to-image generation model based on latent diffusion, developed by [the Kuaishou Kolors team](https://github.com/Kwai-Kolors/Kolors). Trained on billions of text-image pairs, Kolors exhibits significant advantages over both open-source and closed-source models in visual quality, complex semantic accuracy, and text rendering for both Chinese and English characters. Furthermore, Kolors supports both Chinese and English inputs, demonstrating strong performance in understanding and generating Chinese-specific content. For more details, please refer to this [technical report](https://github.com/Kwai-Kolors/Kolors/blob/master/imgs/Kolors_paper.pdf).
+Kolors is a large-scale text-to-image generation model based on latent diffusion, developed by [the Kuaishou Kolors team](kwai-kolors@kuaishou.com). Trained on billions of text-image pairs, Kolors exhibits significant advantages over both open-source and closed-source models in visual quality, complex semantic accuracy, and text rendering for both Chinese and English characters. Furthermore, Kolors supports both Chinese and English inputs, demonstrating strong performance in understanding and generating Chinese-specific content. For more details, please refer to this [technical report](https://github.com/Kwai-Kolors/Kolors/blob/master/imgs/Kolors_paper.pdf).

 The abstract from the technical report is:

@@ -74,7 +74,7 @@ image_encoder = CLIPVisionModelWithProjection.from_pretrained(

 pipe = KolorsPipeline.from_pretrained(
    "Kwai-Kolors/Kolors-diffusers", image_encoder=image_encoder, torch_dtype=torch.float16, variant="fp16"
-)
+).to("cuda")
 pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True)

 pipe.load_ip_adapter(
@@ -20,7 +20,7 @@ The abstract from the paper is:

 *Recent studies have demonstrated that diffusion models are capable of generating high-quality samples, but their quality heavily depends on sampling guidance techniques, such as classifier guidance (CG) and classifier-free guidance (CFG). These techniques are often not applicable in unconditional generation or in various downstream tasks such as image restoration. In this paper, we propose a novel sampling guidance, called Perturbed-Attention Guidance (PAG), which improves diffusion sample quality across both unconditional and conditional settings, achieving this without requiring additional training or the integration of external modules. PAG is designed to progressively enhance the structure of samples throughout the denoising process. It involves generating intermediate samples with degraded structure by substituting selected self-attention maps in diffusion U-Net with an identity matrix, by considering the self-attention mechanisms' ability to capture structural information, and guiding the denoising process away from these degraded samples. In both ADM and Stable Diffusion, PAG surprisingly improves sample quality in conditional and even unconditional scenarios. Moreover, PAG significantly improves the baseline performance in various downstream tasks where existing guidances such as CG or CFG cannot be fully utilized, including ControlNet with empty prompts and image restoration such as inpainting and deblurring.*

-PAG can be used by specifying the `pag_applied_layers` as a parameter when instantiating a PAG pipeline. It can be a single string or a list of strings. Each string can be a unique layer identifier or a regular expression to identify one or more layers.
+PAG can be used by specifying the `pag_applied_layers` as a parameter when instantiating a PAG pipeline. It can be a single string or a list of strings. Each string can be a unique layer identifier or a regular expression to identify one or more layers. 

 - Full identifier as a normal string: `down_blocks.2.attentions.0.transformer_blocks.0.attn1.processor`
 - Full identifier as a RegEx: `down_blocks.2.(attentions|motion_modules).0.transformer_blocks.0.attn1.processor`
@@ -46,7 +46,7 @@ Since RegEx is supported as a way for matching layer identifiers, it is crucial
 ## KolorsPAGPipeline
 [[autodoc]] KolorsPAGPipeline
  - all
-  - __call__
+  - __call__  

 ## StableDiffusionPAGPipeline
 [[autodoc]] StableDiffusionPAGPipeline
@@ -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-tools](https://github.com/Stability-AI/stable-audio-tools).
+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).

 ## Tips

@@ -125,5 +125,3 @@ 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).
@@ -48,7 +48,7 @@ accelerate launch run_distributed.py --num_processes=2

 <Tip>

-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.
+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.
@@ -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.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.
+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.

-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.
+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.

 ```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.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!
+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!

 ```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 [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.
+> 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.

-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.
+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.

 ```bash
 pip install -U diffusers peft
 ```

-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:
+There are three steps to merge LoRAs with the [`~peft.LoraModel.add_weighted_adapter`] method:

-1. Create a [PeftModel](https://huggingface.co/docs/peft/package_reference/peft_model#peft.PeftModel) from the underlying model and LoRA checkpoint.
+1. Create a [`~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 [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method and the merging method of your choice.
+3. Merge the adapters using the [`~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 [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.
+Now you'll create a [`~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 [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.
+Repeat this process to create a [`~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 [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.
+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.

 > [!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.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.
+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.

-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.
+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.

 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.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.
+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.

 ```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.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.
+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.

 ```py
 pipeline.unfuse_lora()
@@ -71,7 +71,7 @@ from diffusers.utils.import_utils import is_xformers_available


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

@@ -71,7 +71,6 @@ 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.

@@ -1647,6 +1646,7 @@ 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',
@@ -1661,6 +1661,7 @@ 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')
@@ -4208,52 +4209,6 @@ 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

@@ -4330,4 +4285,4 @@ 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'])
@@ -43,7 +43,7 @@ from diffusers.utils import BaseOutput, check_min_version


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")


 class MarigoldDepthOutput(BaseOutput):
@@ -73,7 +73,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

@@ -66,7 +66,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

@@ -78,7 +78,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = logging.getLogger(__name__)

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)
 if is_torch_npu_available():
@@ -63,7 +63,7 @@ from diffusers.utils.import_utils import is_xformers_available


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

@@ -1,195 +0,0 @@
-# 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 ♥️
@@ -1,8 +0,0 @@
-accelerate>=0.31.0
-torchvision
-transformers>=4.41.2
-ftfy
-tensorboard
-Jinja2
-peft>=0.11.1
-sentencepiece
@@ -1,203 +0,0 @@
-# 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"})
@@ -1,165 +0,0 @@
-# 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"})
@@ -63,7 +63,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

@@ -35,7 +35,7 @@ from diffusers.utils import check_min_version


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 # Cache compiled models across invocations of this script.
 cc.initialize_cache(os.path.expanduser("~/.cache/jax/compilation_cache"))
@@ -70,7 +70,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

@@ -1454,7 +1454,7 @@ def main(args):
            )

    # Clear the memory here
-    if not args.train_text_encoder and not train_dataset.custom_instance_prompts:
+    if not args.train_text_encoder and train_dataset.custom_instance_prompts:
        del tokenizers, text_encoders
        # Explicitly delete the objects as well, otherwise only the lists are deleted and the original references remain, preventing garbage collection
        del text_encoder_one, text_encoder_two, text_encoder_three
@@ -78,7 +78,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

@@ -64,7 +64,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -52,7 +52,7 @@ if is_wandb_available():


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -46,7 +46,7 @@ from diffusers.utils import check_min_version, is_wandb_available


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -46,7 +46,7 @@ from diffusers.utils import check_min_version, is_wandb_available


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -51,7 +51,7 @@ if is_wandb_available():


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -2,8 +2,8 @@ diffusers==0.20.1
 accelerate==0.23.0
 transformers==4.38.0
 peft==0.5.0
-torch==2.2.0
+torch==2.0.1
 torchvision>=0.16
 ftfy==6.1.1
 tensorboard==2.14.0
-Jinja2==3.1.4
+Jinja2==3.1.3
@@ -60,7 +60,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

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


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -826,22 +826,17 @@ def main():
    )

    # Scheduler and math around the number of training 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
+    overrode_max_train_steps = False
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if args.max_train_steps is None:
-        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
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        overrode_max_train_steps = True

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
-        num_warmup_steps=num_warmup_steps_for_scheduler,
-        num_training_steps=num_training_steps_for_scheduler,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
    )

    # Prepare everything with our `accelerator`.
@@ -871,14 +866,8 @@ 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 args.max_train_steps is None:
+    if overrode_max_train_steps:
        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)

@@ -49,7 +49,7 @@ from diffusers.utils import check_min_version


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = logging.getLogger(__name__)

@@ -56,7 +56,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -68,7 +68,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)
 if is_torch_npu_available():
@@ -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 available in the dataset",
+        help="debug loss for each image, if filenames are awailable in the dataset",
    )

    if input_args is not None:
@@ -55,7 +55,7 @@ from diffusers.utils.torch_utils import is_compiled_module


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)
 if is_torch_npu_available():
@@ -109,9 +109,6 @@ import torch
 model_id = "path-to-your-trained-model"
 pipe = StableDiffusionPipeline.from_pretrained(model_id,torch_dtype=torch.float16).to("cuda")

-repo_id_embeds = "path-to-your-learned-embeds"
-pipe.load_textual_inversion(repo_id_embeds)
-
 prompt = "A <cat-toy> backpack"

 image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
@@ -23,25 +23,4 @@ accelerate launch textual_inversion_sdxl.py \
  --output_dir="./textual_inversion_cat_sdxl"
 ```

-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")
-```
+For now, only training of the first text encoder is supported.
@@ -81,7 +81,7 @@ else:


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

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

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = logging.getLogger(__name__)

@@ -76,7 +76,7 @@ else:


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

 logger = get_logger(__name__)

@@ -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=accelerator.unwrap_model(text_encoder_2),
+        text_encoder_2=text_encoder_2,
        tokenizer=tokenizer_1,
        tokenizer_2=tokenizer_2,
        unet=unet,
@@ -678,54 +678,36 @@ 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 or len(token_ids_2) > 1:
+    if len(token_ids) > 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():
@@ -764,11 +746,7 @@ def main():
        optimizer_class = torch.optim.AdamW

    optimizer = optimizer_class(
-        # only optimize the embeddings
-        [
-            text_encoder_1.text_model.embeddings.token_embedding.weight,
-            text_encoder_2.text_model.embeddings.token_embedding.weight,
-        ],
+        text_encoder_1.get_input_embeddings().parameters(),  # only optimize the embeddings
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
@@ -808,10 +786,9 @@ def main():
    )

    text_encoder_1.train()
-    text_encoder_2.train()
    # Prepare everything with our `accelerator`.
-    text_encoder_1, text_encoder_2, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        text_encoder_1, text_encoder_2, optimizer, train_dataloader, lr_scheduler
+    text_encoder_1, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        text_encoder_1, 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
@@ -889,13 +866,11 @@ 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, text_encoder_2]):
+            with accelerator.accumulate(text_encoder_1):
                # 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
@@ -917,7 +892,9 @@ def main():
                    .hidden_states[-2]
                    .to(dtype=weight_dtype)
                )
-                encoder_output_2 = text_encoder_2(batch["input_ids_2"], output_hidden_states=True)
+                encoder_output_2 = text_encoder_2(
+                    batch["input_ids_2"].reshape(batch["input_ids_1"].shape[0], -1), 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())
@@ -961,16 +938,11 @@ 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:
@@ -988,16 +960,6 @@ 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:
@@ -1072,7 +1034,7 @@ def main():
            pipeline = DiffusionPipeline.from_pretrained(
                args.pretrained_model_name_or_path,
                text_encoder=accelerator.unwrap_model(text_encoder_1),
-                text_encoder_2=accelerator.unwrap_model(text_encoder_2),
+                text_encoder_2=text_encoder_2,
                vae=vae,
                unet=unet,
                tokenizer=tokenizer_1,
@@ -1090,16 +1052,6 @@ 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(
@@ -29,7 +29,7 @@ from diffusers.utils.import_utils import is_xformers_available


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -50,7 +50,7 @@ if is_wandb_available():
    import wandb

 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -50,7 +50,7 @@ if is_wandb_available():


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -51,7 +51,7 @@ if is_wandb_available():


 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.31.0.dev0")
+check_min_version("0.30.0.dev0")

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

@@ -1,222 +0,0 @@
-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)
@@ -254,7 +254,7 @@ version_range_max = max(sys.version_info[1], 10) + 1

 setup(
    name="diffusers",
-    version="0.31.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="0.30.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
    description="State-of-the-art diffusion in PyTorch and JAX.",
    long_description=open("README.md", "r", encoding="utf-8").read(),
    long_description_content_type="text/markdown",
@@ -1,4 +1,4 @@
-__version__ = "0.31.0.dev0"
+__version__ = "0.30.0.dev0"

 from typing import TYPE_CHECKING

@@ -79,16 +79,13 @@ else:
            "AsymmetricAutoencoderKL",
            "AuraFlowTransformer2DModel",
            "AutoencoderKL",
-            "AutoencoderKLCogVideoX",
            "AutoencoderKLTemporalDecoder",
            "AutoencoderOobleck",
            "AutoencoderTiny",
-            "CogVideoXTransformer3DModel",
            "ConsistencyDecoderVAE",
            "ControlNetModel",
            "ControlNetXSAdapter",
            "DiTTransformer2DModel",
-            "FluxControlNetModel",
            "FluxTransformer2DModel",
            "HunyuanDiT2DControlNetModel",
            "HunyuanDiT2DModel",
@@ -158,8 +155,6 @@ else:
        [
            "AmusedScheduler",
            "CMStochasticIterativeScheduler",
-            "CogVideoXDDIMScheduler",
-            "CogVideoXDPMScheduler",
            "DDIMInverseScheduler",
            "DDIMParallelScheduler",
            "DDIMScheduler",
@@ -253,9 +248,7 @@ else:
            "BlipDiffusionControlNetPipeline",
            "BlipDiffusionPipeline",
            "CLIPImageProjection",
-            "CogVideoXPipeline",
            "CycleDiffusionPipeline",
-            "FluxControlNetPipeline",
            "FluxPipeline",
            "HunyuanDiTControlNetPipeline",
            "HunyuanDiTPAGPipeline",
@@ -310,7 +303,6 @@ else:
            "StableCascadeCombinedPipeline",
            "StableCascadeDecoderPipeline",
            "StableCascadePriorPipeline",
-            "StableDiffusion3ControlNetInpaintingPipeline",
            "StableDiffusion3ControlNetPipeline",
            "StableDiffusion3Img2ImgPipeline",
            "StableDiffusion3InpaintPipeline",
@@ -543,16 +535,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            AsymmetricAutoencoderKL,
            AuraFlowTransformer2DModel,
            AutoencoderKL,
-            AutoencoderKLCogVideoX,
            AutoencoderKLTemporalDecoder,
            AutoencoderOobleck,
            AutoencoderTiny,
-            CogVideoXTransformer3DModel,
            ConsistencyDecoderVAE,
            ControlNetModel,
            ControlNetXSAdapter,
            DiTTransformer2DModel,
-            FluxControlNetModel,
            FluxTransformer2DModel,
            HunyuanDiT2DControlNetModel,
            HunyuanDiT2DModel,
@@ -619,8 +608,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
        from .schedulers import (
            AmusedScheduler,
            CMStochasticIterativeScheduler,
-            CogVideoXDDIMScheduler,
-            CogVideoXDPMScheduler,
            DDIMInverseScheduler,
            DDIMParallelScheduler,
            DDIMScheduler,
@@ -695,9 +682,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            AudioLDMPipeline,
            AuraFlowPipeline,
            CLIPImageProjection,
-            CogVideoXPipeline,
            CycleDiffusionPipeline,
-            FluxControlNetPipeline,
            FluxPipeline,
            HunyuanDiTControlNetPipeline,
            HunyuanDiTPAGPipeline,
@@ -222,11 +222,7 @@ class IPAdapterMixin:

            # create feature extractor if it has not been registered to the pipeline yet
            if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is None:
-                # FaceID IP adapters don't need the image encoder so it's not present, in this case we default to 224
-                default_clip_size = 224
-                clip_image_size = (
-                    self.image_encoder.config.image_size if self.image_encoder is not None else default_clip_size
-                )
+                clip_image_size = self.image_encoder.config.image_size
                feature_extractor = CLIPImageProcessor(size=clip_image_size, crop_size=clip_image_size)
                self.register_modules(feature_extractor=feature_extractor)

@@ -75,9 +75,6 @@ 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",
@@ -74,11 +74,9 @@ 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.0.pos_encoder.pe",
+    "animatediff": "down_blocks.0.motion_modules.0.temporal_transformer.transformer_blocks.0.attention_blocks.1.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",
 }

@@ -113,8 +111,6 @@ 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"},
 }
@@ -498,13 +494,7 @@ def infer_diffusers_model_type(checkpoint):
        model_type = "sd3"

    elif CHECKPOINT_KEY_NAMES["animatediff"] 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:
+        if CHECKPOINT_KEY_NAMES["animatediff_v2"] in checkpoint:
            model_type = "animatediff_v2"

        elif checkpoint[CHECKPOINT_KEY_NAMES["animatediff_sdxl_beta"]].shape[-1] == 320:
@@ -28,14 +28,12 @@ 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"]
@@ -43,7 +41,6 @@ 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"]
@@ -80,7 +77,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
        from .autoencoders import (
            AsymmetricAutoencoderKL,
            AutoencoderKL,
-            AutoencoderKLCogVideoX,
            AutoencoderKLTemporalDecoder,
            AutoencoderOobleck,
            AutoencoderTiny,
@@ -88,7 +84,6 @@ 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
@@ -97,7 +92,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
        from .modeling_utils import ModelMixin
        from .transformers import (
            AuraFlowTransformer2DModel,
-            CogVideoXTransformer3DModel,
            DiTTransformer2DModel,
            DualTransformer2DModel,
            FluxTransformer2DModel,
@@ -449,7 +449,7 @@ class BasicTransformerBlock(nn.Module):
            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
        elif self.norm_type == "ada_norm_single":
            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
-                self.scale_shift_table[None].to(timestep.dtype) + timestep.reshape(batch_size, 6, -1)
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
            ).chunk(6, dim=1)
            norm_hidden_states = self.norm1(hidden_states)
            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
@@ -1,6 +1,5 @@
 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
@@ -60,8 +60,6 @@ class AsymmetricAutoencoderKL(ModelMixin, ConfigMixin):
            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
    """

-    _always_upcast_modules = ["MaskConditionDecoder"]
-
    @register_to_config
    def __init__(
        self,
@@ -70,7 +70,6 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalModelMixin):

    _supports_gradient_checkpointing = True
    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D"]
-    _always_upcast_modules = ["Decoder"]

    @register_to_config
    def __init__(
@@ -192,7 +192,6 @@ class AutoencoderKLTemporalDecoder(ModelMixin, ConfigMixin):
    """

    _supports_gradient_checkpointing = True
-    _always_upcast_modules = ["TemporalDecoder"]

    @register_to_config
    def __init__(
@@ -317,7 +317,6 @@ class AutoencoderOobleck(ModelMixin, ConfigMixin):
    """

    _supports_gradient_checkpointing = False
-    _always_upcast_modules = ["OobleckEncoder", "OobleckDecoder"]

    @register_to_config
    def __init__(
@@ -330,7 +330,7 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
            Union[DecoderOutput, Tuple[torch.Tensor]]: The decoded output.

        """
-        z = (z * self.config.scaling_factor - self.means.to(z.dtype)) / self.stds.to(z.dtype)
+        z = (z * self.config.scaling_factor - self.means) / self.stds

        scale_factor = 2 ** (len(self.config.block_out_channels) - 1)
        z = F.interpolate(z, mode="nearest", scale_factor=scale_factor)
@@ -71,8 +71,6 @@ class VQModel(ModelMixin, ConfigMixin):
            Type of normalization layer to use. Can be one of `"group"` or `"spatial"`.
    """

-    _always_upcast_modules = ["Decoder", "VectorQuantizer"]
-
    @register_to_config
    def __init__(
        self,
@@ -1,374 +0,0 @@
-# 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,
-        )
@@ -55,7 +55,6 @@ 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
@@ -99,7 +98,7 @@ class SD3ControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginal
            height=sample_size,
            width=sample_size,
            patch_size=patch_size,
-            in_channels=in_channels + extra_conditioning_channels,
+            in_channels=in_channels,
            embed_dim=self.inner_dim,
            pos_embed_type=None,
        )
@@ -20,7 +20,6 @@ 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,
@@ -93,7 +92,7 @@ class SparseControlNetConditioningEmbedding(nn.Module):
        return embedding


-class SparseControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+class SparseControlNetModel(ModelMixin, ConfigMixin):
    """
    A SparseControlNet model as described in [SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion
    Models](https://arxiv.org/abs/2311.16933).
@@ -315,7 +314,6 @@ class SparseControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
                    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(
@@ -333,7 +331,6 @@ class SparseControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
                    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(
@@ -285,74 +285,6 @@ 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,
@@ -78,53 +78,6 @@ 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
 ):
@@ -334,46 +287,6 @@ 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.
@@ -263,80 +263,6 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
        """
        self.set_use_memory_efficient_attention_xformers(False)

-    def enable_layerwise_upcasting(self, upcast_dtype=None):
-        r"""
-        Enable layerwise dynamic upcasting. This allows models to be loaded into the GPU in a low memory dtype e.g.
-        torch.float8_e4m3fn, but perform inference using a dtype that is supported by the GPU, by upcasting the
-        individual modules in the model to the appropriate dtype right before the foward pass.
-
-        The module is then moved back to the low memory dtype after the foward pass.
-        """
-
-        upcast_dtype = upcast_dtype or torch.float32
-        original_dtype = self.dtype
-
-        def upcast_dtype_hook_fn(module, *args, **kwargs):
-            module = module.to(upcast_dtype)
-
-        def cast_to_original_dtype_hook_fn(module, *args, **kwargs):
-            module = module.to(original_dtype)
-
-        def fn_recursive_upcast(module):
-            """In certain cases modules will apply casting internally or reference the dtype of internal blocks.
-
-            e.g.
-
-            ```
-            class MyModel(nn.Module):
-                def forward(self, x):
-                    dtype = next(iter(self.blocks.parameters())).dtype
-                    x = self.blocks(x) + torch.ones(x.size()).to(dtype)
-            ```
-            Layerwise upcasting will not work here, since the internal blocks remain in the low memory dtype until
-            their `forward` method is called. We need to add the upcast hook on the entire module in order for the
-            operation to work.
-
-            The `_always_upcast_modules` class attribute is a list of modules within the model that we must upcast
-            entirely, rather than layerwise.
-
-            """
-            if hasattr(self, "_always_upcast_modules") and module.__class__.__name__ in self._always_upcast_modules:
-                # Upcast entire module and exist recursion
-                module.register_forward_pre_hook(upcast_dtype_hook_fn)
-                module.register_forward_hook(cast_to_original_dtype_hook_fn)
-
-                return
-
-            has_children = list(module.children())
-            if not has_children:
-                module.register_forward_pre_hook(upcast_dtype_hook_fn)
-                module.register_forward_hook(cast_to_original_dtype_hook_fn)
-
-            for child in module.children():
-                fn_recursive_upcast(child)
-
-        for module in self.children():
-            fn_recursive_upcast(module)
-
-    def disable_layerwise_upcasting(self):
-        def fn_recursive_upcast(module):
-            if hasattr(self, "_always_upcast_modules") and module.__class__.__name__ in self._always_upcast_modules:
-                module._forward_pre_hooks = OrderedDict()
-                module._forward_hooks = OrderedDict()
-
-                return
-
-            has_children = list(module.children())
-            if not has_children:
-                module._forward_pre_hooks = OrderedDict()
-                module._forward_hooks = OrderedDict()
-
-            for child in module.children():
-                fn_recursive_upcast(child)
-
-        for module in self.children():
-            fn_recursive_upcast(module)
-
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
@@ -34,53 +34,19 @@ class AdaLayerNorm(nn.Module):

    Parameters:
        embedding_dim (`int`): The size of each embedding vector.
-        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`):
+        num_embeddings (`int`): The size of the embeddings dictionary.
    """

-    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,
-    ):
+    def __init__(self, embedding_dim: int, num_embeddings: int):
        super().__init__()
-
-        self.chunk_dim = chunk_dim
-        output_dim = output_dim or embedding_dim * 2
-
-        if num_embeddings is not None:
-            self.emb = nn.Embedding(num_embeddings, embedding_dim)
-        else:
-            self.emb = None
-
+        self.emb = nn.Embedding(num_embeddings, embedding_dim)
        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)
+        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)
        x = self.norm(x) * (1 + scale) + shift
        return x

@@ -355,30 +321,6 @@ 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:
@@ -3,7 +3,6 @@ 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
@@ -68,21 +68,6 @@ class AuraFlowPatchEmbed(nn.Module):
        self.height, self.width = height // patch_size, width // patch_size
        self.base_size = height // patch_size

-    def pe_selection_index_based_on_dim(self, h, w):
-        # select subset of positional embedding based on H, W, where H, W is size of latent
-        # PE will be viewed as 2d-grid, and H/p x W/p of the PE will be selected
-        # because original input are in flattened format, we have to flatten this 2d grid as well.
-        h_p, w_p = h // self.patch_size, w // self.patch_size
-        original_pe_indexes = torch.arange(self.pos_embed.shape[1])
-        h_max, w_max = int(self.pos_embed_max_size**0.5), int(self.pos_embed_max_size**0.5)
-        original_pe_indexes = original_pe_indexes.view(h_max, w_max)
-        starth = h_max // 2 - h_p // 2
-        endh = starth + h_p
-        startw = w_max // 2 - w_p // 2
-        endw = startw + w_p
-        original_pe_indexes = original_pe_indexes[starth:endh, startw:endw]
-        return original_pe_indexes.flatten()
-
    def forward(self, latent):
        batch_size, num_channels, height, width = latent.size()
        latent = latent.view(
@@ -95,8 +80,7 @@ class AuraFlowPatchEmbed(nn.Module):
        )
        latent = latent.permute(0, 2, 4, 1, 3, 5).flatten(-3).flatten(1, 2)
        latent = self.proj(latent)
-        pe_index = self.pe_selection_index_based_on_dim(height, width)
-        return latent + self.pos_embed[:, pe_index]
+        return latent + self.pos_embed


 # Taken from the original Aura flow inference code.
@@ -274,9 +258,7 @@ class AuraFlowTransformer2DModel(ModelMixin, ConfigMixin):
        pos_embed_max_size (`int`, defaults to 4096): Maximum positions to embed from the image latents.
    """

-    _no_split_modules = ["AuraFlowJointTransformerBlock", "AuraFlowSingleTransformerBlock", "AuraFlowPatchEmbed"]
    _supports_gradient_checkpointing = True
-    _always_upcast_modules = ["AuraFlowPatchEmbed"]

    @register_to_config
    def __init__(
@@ -458,15 +440,11 @@ class AuraFlowTransformer2DModel(ModelMixin, ConfigMixin):

        # Apply patch embedding, timestep embedding, and project the caption embeddings.
        hidden_states = self.pos_embed(hidden_states)  # takes care of adding positional embeddings too.
-        temb = self.time_step_embed(timestep).to(dtype=hidden_states.dtype)
+        temb = self.time_step_embed(timestep).to(dtype=next(self.parameters()).dtype)
        temb = self.time_step_proj(temb)
        encoder_hidden_states = self.context_embedder(encoder_hidden_states)
        encoder_hidden_states = torch.cat(
-            [
-                self.register_tokens.to(encoder_hidden_states.dtype).repeat(encoder_hidden_states.size(0), 1, 1),
-                encoder_hidden_states,
-            ],
-            dim=1,
+            [self.register_tokens.repeat(encoder_hidden_states.size(0), 1, 1), encoder_hidden_states], dim=1
        )

        # MMDiT blocks.
@@ -1,369 +0,0 @@
-# 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)
@@ -65,7 +65,6 @@ class DiTTransformer2DModel(ModelMixin, ConfigMixin):
    """

    _supports_gradient_checkpointing = True
-    _always_upcast_modules = ["PatchEmbed"]

    @register_to_config
    def __init__(
@@ -244,8 +244,6 @@ class HunyuanDiT2DModel(ModelMixin, ConfigMixin):
            Whether or not to use style condition and image meta size. True for version <=1.1, False for version >= 1.2
    """

-    _always_upcast_modules = ["HunyuanDiTAttentionPool"]
-
    @register_to_config
    def __init__(
        self,
@@ -486,9 +484,7 @@ class HunyuanDiT2DModel(ModelMixin, ConfigMixin):
        text_embedding_mask = torch.cat([text_embedding_mask, text_embedding_mask_t5], dim=-1)
        text_embedding_mask = text_embedding_mask.unsqueeze(2).bool()

-        encoder_hidden_states = torch.where(
-            text_embedding_mask, encoder_hidden_states, self.text_embedding_padding.to(encoder_hidden_states.dtype)
-        )
+        encoder_hidden_states = torch.where(text_embedding_mask, encoder_hidden_states, self.text_embedding_padding)

        skips = []
        for layer, block in enumerate(self.blocks):
@@ -64,7 +64,6 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
        video_length (`int`, *optional*):
            The number of frames in the video-like data.
    """
-    _always_upcast_modules = ["PatchEmbed"]

    @register_to_config
    def __init__(
@@ -302,9 +301,7 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
                hidden_states = hidden_states.reshape(-1, hidden_states.shape[-2], hidden_states.shape[-1])

        embedded_timestep = embedded_timestep.repeat_interleave(num_frame, dim=0).view(-1, embedded_timestep.shape[-1])
-        shift, scale = (self.scale_shift_table[None].to(embedded_timestep.dtype) + embedded_timestep[:, None]).chunk(
-            2, dim=1
-        )
+        shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
        hidden_states = self.norm_out(hidden_states)
        # Modulation
        hidden_states = hidden_states * (1 + scale) + shift
@@ -19,7 +19,7 @@ from torch import nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import is_torch_version, logging
 from ..attention import BasicTransformerBlock
-from ..attention_processor import Attention, AttentionProcessor, AttnProcessor, FusedAttnProcessor2_0
+from ..attention_processor import Attention, AttentionProcessor, FusedAttnProcessor2_0
 from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -79,7 +79,6 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):

    _supports_gradient_checkpointing = True
    _no_split_modules = ["BasicTransformerBlock", "PatchEmbed"]
-    _always_upcast_modules = ["PatchEmbed"]

    @register_to_config
    def __init__(
@@ -248,14 +247,6 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
        for name, module in self.named_children():
            fn_recursive_attn_processor(name, module, processor)

-    def set_default_attn_processor(self):
-        """
-        Disables custom attention processors and sets the default attention implementation.
-
-        Safe to just use `AttnProcessor()` as PixArt doesn't have any exotic attention processors in default model.
-        """
-        self.set_attn_processor(AttnProcessor())
-
    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
    def fuse_qkv_projections(self):
        """
@@ -423,8 +414,7 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):

        # 3. Output
        shift, scale = (
-            self.scale_shift_table[None].to(embedded_timestep.dtype)
-            + embedded_timestep[:, None].to(self.scale_shift_table.device)
+            self.scale_shift_table[None] + embedded_timestep[:, None].to(self.scale_shift_table.device)
        ).chunk(2, dim=1)
        hidden_states = self.norm_out(hidden_states)
        # Modulation
@@ -289,7 +289,7 @@ class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Pef

        # timesteps does not contain any weights and will always return f32 tensors
        # but time_embedding might be fp16, so we need to cast here.
-        timesteps_projected = timesteps_projected.to(dtype=hidden_states.dtype)
+        timesteps_projected = timesteps_projected.to(dtype=self.dtype)
        time_embeddings = self.time_embedding(timesteps_projected)

        if self.embedding_proj_norm is not None:
@@ -1,4 +1,4 @@
-# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved.
+# Copyright 2024 Black Forest Labs, 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.
@@ -15,7 +15,6 @@

 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
@@ -126,8 +125,6 @@ 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

@@ -226,8 +223,6 @@ 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

@@ -251,7 +246,6 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
    """

    _supports_gradient_checkpointing = True
-    _no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]

    @register_to_config
    def __init__(
@@ -323,8 +317,6 @@ 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]:
        """
@@ -381,7 +373,6 @@ 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)

@@ -415,12 +406,6 @@ 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):
@@ -451,15 +436,6 @@ 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)
@@ -54,7 +54,6 @@ class SD3Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrigi
    """

    _supports_gradient_checkpointing = True
-    _always_upcast_modules = ["PatchEmbed"]

    @register_to_config
    def __init__(
@@ -283,7 +283,7 @@ class UNet2DModel(ModelMixin, ConfigMixin):
        # 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=sample.dtype)
+        t_emb = t_emb.to(dtype=self.dtype)
        emb = self.time_embedding(t_emb)

        if self.class_embedding is not None:
@@ -641,7 +641,7 @@ class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
        # 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=sample.dtype)
+        t_emb = t_emb.to(dtype=self.dtype)

        emb = self.time_embedding(t_emb, timestep_cond)
        emb = emb.repeat_interleave(repeats=num_frames, dim=0)
@@ -590,7 +590,7 @@ class I2VGenXLUNet(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
        # 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=sample.dtype)
+        t_emb = t_emb.to(dtype=self.dtype)
        t_emb = self.time_embedding(t_emb, timestep_cond)

        # 2. FPS
@@ -233,7 +233,6 @@ 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 = []
@@ -283,7 +282,6 @@ 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,
                )
            )

@@ -387,7 +385,6 @@ 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 = []
@@ -469,7 +466,6 @@ 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,
                )
            )

@@ -2152,7 +2148,7 @@ class UNetMotionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Peft
        # 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=sample.dtype)
+        t_emb = t_emb.to(dtype=self.dtype)

        emb = self.time_embedding(t_emb, timestep_cond)
        aug_emb = None
@@ -348,70 +348,6 @@ 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,
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Dhruv Nair	94bfe7da73	update	2024-08-07 07:03:12 +00:00
Dhruv Nair	cb508450de	update	2024-08-07 06:54:13 +00:00