update

Merge branch 'main' into ruff-update
AutoModel (#11115 )
2025-04-09 15:34:42 +05:30 · 2025-04-09 15:22:55 +05:30 · 2025-04-09 15:20:07 +05:30 · 2025-04-09 14:12:00 +05:30 · 2025-04-09 12:12:53 +05:30 · 2025-04-09 09:17:05 +05:30
375 changed files with 21649 additions and 6334 deletions
@@ -38,6 +38,7 @@ jobs:
          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
          python -m uv pip install -e [quality,test]
          python -m uv pip install pandas peft
+          python -m uv pip uninstall transformers && python -m uv pip install transformers==4.48.0
      - name: Environment
        run: |
          python utils/print_env.py
@@ -414,12 +414,16 @@ jobs:
        config:
          - backend: "bitsandbytes"
            test_location: "bnb"
+            additional_deps: ["peft"]
          - backend: "gguf"
            test_location: "gguf"
+            additional_deps: ["peft"]
          - backend: "torchao"
            test_location: "torchao"
+            additional_deps: []
          - backend: "optimum_quanto"
            test_location: "quanto"
+            additional_deps: []
    runs-on:
      group: aws-g6e-xlarge-plus
    container:
@@ -437,6 +441,9 @@ jobs:
          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
          python -m uv pip install -e [quality,test]
          python -m uv pip install -U ${{ matrix.config.backend }}
+          if [ "${{ join(matrix.config.additional_deps, ' ') }}" != "" ]; then
+              python -m uv pip install ${{ join(matrix.config.additional_deps, ' ') }}
+          fi
          python -m uv pip install pytest-reportlog
      - name: Environment
        run: |
@@ -13,39 +13,5 @@ jobs:
    uses: huggingface/huggingface_hub/.github/workflows/style-bot-action.yml@main
    with:
      python_quality_dependencies: "[quality]"
-      pre_commit_script_name: "Download and Compare files from the main branch"
-      pre_commit_script: |
-        echo "Downloading the files from the main branch"
-
-        curl -o main_Makefile https://raw.githubusercontent.com/huggingface/diffusers/main/Makefile
-        curl -o main_setup.py https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/setup.py
-        curl -o main_check_doc_toc.py https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/utils/check_doc_toc.py
-
-        echo "Compare the files and raise error if needed"
-
-        diff_failed=0
-        if ! diff -q main_Makefile Makefile; then
-          echo "Error: The Makefile has changed. Please ensure it matches the main branch."
-          diff_failed=1
-        fi
-
-        if ! diff -q main_setup.py setup.py; then
-          echo "Error: The setup.py has changed. Please ensure it matches the main branch."
-          diff_failed=1
-        fi
-
-        if ! diff -q main_check_doc_toc.py utils/check_doc_toc.py; then
-          echo "Error: The utils/check_doc_toc.py has changed. Please ensure it matches the main branch."
-          diff_failed=1
-        fi
-
-        if [ $diff_failed -eq 1 ]; then
-          echo "❌ Error happened as we detected changes in the files that should not be changed ❌"
-          exit 1
-        fi
-
-        echo "No changes in the files. Proceeding..."
-        rm -rf main_Makefile main_setup.py main_check_doc_toc.py
-      style_command: "make style && make quality"
    secrets:
      bot_token: ${{ secrets.GITHUB_TOKEN }}
@@ -28,7 +28,51 @@ env:
  PIPELINE_USAGE_CUTOFF: 1000000000 # set high cutoff so that only always-test pipelines run

 jobs:
+  check_code_quality:
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[quality]
+      - name: Check quality
+        run: make quality
+      - name: Check if failure
+        if: ${{ failure() }}
+        run: |
+          echo "Quality check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make style && make quality'" >> $GITHUB_STEP_SUMMARY
+
+  check_repository_consistency:
+    needs: check_code_quality
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[quality]
+      - name: Check repo consistency
+        run: |
+          python utils/check_copies.py
+          python utils/check_dummies.py
+          python utils/check_support_list.py
+          make deps_table_check_updated
+      - name: Check if failure
+        if: ${{ failure() }}
+        run: |
+          echo "Repo consistency check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make fix-copies'" >> $GITHUB_STEP_SUMMARY
+  
  setup_torch_cuda_pipeline_matrix:
+    needs: [check_code_quality, check_repository_consistency]
    name: Setup Torch Pipelines CUDA Slow Tests Matrix
    runs-on:
      group: aws-general-8-plus
@@ -133,6 +177,7 @@ jobs:

  torch_cuda_tests:
    name: Torch CUDA Tests
+    needs: [check_code_quality, check_repository_consistency]
    runs-on:
      group: aws-g4dn-2xlarge
    container:
@@ -201,7 +246,7 @@ jobs:

  run_examples_tests:
    name: Examples PyTorch CUDA tests on Ubuntu
-        pip uninstall transformers -y && python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
+    needs: [check_code_quality, check_repository_consistency]
    runs-on:
      group: aws-g4dn-2xlarge

@@ -220,6 +265,7 @@ jobs:
    - name: Install dependencies
      run: |
        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        pip uninstall transformers -y && python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
        python -m uv pip install -e [quality,test,training]

    - name: Environment
@@ -81,6 +81,8 @@
    title: Overview
  - local: hybrid_inference/vae_decode
    title: VAE Decode
+  - local: hybrid_inference/vae_encode
+    title: VAE Encode
  - local: hybrid_inference/api_reference
    title: API Reference
  title: Hybrid Inference
@@ -494,6 +496,8 @@
      title: PixArt-Σ
    - local: api/pipelines/sana
      title: Sana
+    - local: api/pipelines/sana_sprint
+      title: Sana Sprint
    - local: api/pipelines/self_attention_guidance
      title: Self-Attention Guidance
    - local: api/pipelines/semantic_stable_diffusion
@@ -38,6 +38,33 @@ config = PyramidAttentionBroadcastConfig(
 pipe.transformer.enable_cache(config)
 ```

+## Faster Cache
+
+[FasterCache](https://huggingface.co/papers/2410.19355) from Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, Kwan-Yee K. Wong.
+
+FasterCache is a method that speeds up inference in diffusion transformers by:
+- Reusing attention states between successive inference steps, due to high similarity between them
+- Skipping unconditional branch prediction used in classifier-free guidance by revealing redundancies between unconditional and conditional branch outputs for the same timestep, and therefore approximating the unconditional branch output using the conditional branch output
+
+```python
+import torch
+from diffusers import CogVideoXPipeline, FasterCacheConfig
+
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+config = FasterCacheConfig(
+    spatial_attention_block_skip_range=2,
+    spatial_attention_timestep_skip_range=(-1, 681),
+    current_timestep_callback=lambda: pipe.current_timestep,
+    attention_weight_callback=lambda _: 0.3,
+    unconditional_batch_skip_range=5,
+    unconditional_batch_timestep_skip_range=(-1, 781),
+    tensor_format="BFCHW",
+)
+pipe.transformer.enable_cache(config)
+```
+
 ### CacheMixin

 [[autodoc]] CacheMixin
@@ -47,3 +74,9 @@ pipe.transformer.enable_cache(config)
 [[autodoc]] PyramidAttentionBroadcastConfig

 [[autodoc]] apply_pyramid_attention_broadcast
+
+### FasterCacheConfig
+
+[[autodoc]] FasterCacheConfig
+
+[[autodoc]] apply_faster_cache
@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.

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

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

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

 Flux is a series of text-to-image generation models based on diffusion transformers. To know more about Flux, check out the original [blog post](https://blackforestlabs.ai/announcing-black-forest-labs/) by the creators of Flux, Black Forest Labs.
@@ -50,7 +50,8 @@ The following models are available for the image-to-video pipeline:
 | Model name | Description |
 |:---|:---|
 | [`Skywork/SkyReels-V1-Hunyuan-I2V`](https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-I2V) | Skywork's custom finetune of HunyuanVideo (de-distilled). Performs best with `97x544x960` resolution. Performs best at `97x544x960` resolution, `guidance_scale=1.0`, `true_cfg_scale=6.0` and a negative prompt. |
-| [`hunyuanvideo-community/HunyuanVideo-I2V`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20) |
+| [`hunyuanvideo-community/HunyuanVideo-I2V-33ch`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo 33-channel I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20). |
+| [`hunyuanvideo-community/HunyuanVideo-I2V`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo 16-channel I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20) |

 ## Quantization

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

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

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

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

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

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

@@ -196,6 +198,12 @@ export_to_video(video, "ship.mp4", fps=24)
  - all
  - __call__

+## LTXConditionPipeline
+
+[[autodoc]] LTXConditionPipeline
+  - all
+  - __call__
+
 ## LTXPipelineOutput

 [[autodoc]] pipelines.ltx.pipeline_output.LTXPipelineOutput
@@ -58,10 +58,10 @@ Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fa
 First, load the pipeline:

 ```python
-from diffusers import LuminaText2ImgPipeline
+from diffusers import LuminaPipeline
 import torch

-pipeline = LuminaText2ImgPipeline.from_pretrained(
+pipeline = LuminaPipeline.from_pretrained(
 	"Alpha-VLLM/Lumina-Next-SFT-diffusers", torch_dtype=torch.bfloat16
 ).to("cuda")
 ```
@@ -86,11 +86,11 @@ image = pipeline(prompt="Upper body of a young woman in a Victorian-era outfit w

 Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.

-Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`LuminaText2ImgPipeline`] for inference with bitsandbytes.
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`LuminaPipeline`] for inference with bitsandbytes.

 ```py
 import torch
-from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, Transformer2DModel, LuminaText2ImgPipeline
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, Transformer2DModel, LuminaPipeline
 from transformers import BitsAndBytesConfig as BitsAndBytesConfig, T5EncoderModel

 quant_config = BitsAndBytesConfig(load_in_8bit=True)
@@ -109,7 +109,7 @@ transformer_8bit = Transformer2DModel.from_pretrained(
    torch_dtype=torch.float16,
 )

-pipeline = LuminaText2ImgPipeline.from_pretrained(
+pipeline = LuminaPipeline.from_pretrained(
    "Alpha-VLLM/Lumina-Next-SFT-diffusers",
    text_encoder=text_encoder_8bit,
    transformer=transformer_8bit,
@@ -122,9 +122,9 @@ image = pipeline(prompt).images[0]
 image.save("lumina.png")
 ```

-## LuminaText2ImgPipeline
+## LuminaPipeline

-[[autodoc]] LuminaText2ImgPipeline
+[[autodoc]] LuminaPipeline
 	- all
 	- __call__

@@ -36,14 +36,14 @@ Single file loading for Lumina Image 2.0 is available for the `Lumina2Transforme

 ```python
 import torch
-from diffusers import Lumina2Transformer2DModel, Lumina2Text2ImgPipeline
+from diffusers import Lumina2Transformer2DModel, Lumina2Pipeline

 ckpt_path = "https://huggingface.co/Alpha-VLLM/Lumina-Image-2.0/blob/main/consolidated.00-of-01.pth"
 transformer = Lumina2Transformer2DModel.from_single_file(
    ckpt_path, torch_dtype=torch.bfloat16
 )

-pipe = Lumina2Text2ImgPipeline.from_pretrained(
+pipe = Lumina2Pipeline.from_pretrained(
    "Alpha-VLLM/Lumina-Image-2.0", transformer=transformer, torch_dtype=torch.bfloat16
 )
 pipe.enable_model_cpu_offload()
@@ -60,7 +60,7 @@ image.save("lumina-single-file.png")
 GGUF Quantized checkpoints for the `Lumina2Transformer2DModel` can be loaded via `from_single_file` with the `GGUFQuantizationConfig` 

 ```python
-from diffusers import Lumina2Transformer2DModel, Lumina2Text2ImgPipeline, GGUFQuantizationConfig 
+from diffusers import Lumina2Transformer2DModel, Lumina2Pipeline, GGUFQuantizationConfig 

 ckpt_path = "https://huggingface.co/calcuis/lumina-gguf/blob/main/lumina2-q4_0.gguf"
 transformer = Lumina2Transformer2DModel.from_single_file(
@@ -69,7 +69,7 @@ transformer = Lumina2Transformer2DModel.from_single_file(
    torch_dtype=torch.bfloat16,
 )

-pipe = Lumina2Text2ImgPipeline.from_pretrained(
+pipe = Lumina2Pipeline.from_pretrained(
    "Alpha-VLLM/Lumina-Image-2.0", transformer=transformer, torch_dtype=torch.bfloat16
 )
 pipe.enable_model_cpu_offload()
@@ -80,8 +80,8 @@ image = pipe(
 image.save("lumina-gguf.png")
 ```

-## Lumina2Text2ImgPipeline
+## Lumina2Pipeline

-[[autodoc]] Lumina2Text2ImgPipeline
+[[autodoc]] Lumina2Pipeline
  - all
  - __call__
@@ -16,6 +16,7 @@

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

 [SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.
@@ -0,0 +1,100 @@
+<!-- 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. -->
+
+# SANA-Sprint
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
+[SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation](https://huggingface.co/papers/2503.09641) from NVIDIA, MIT HAN Lab, and Hugging Face by Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Enze Xie, Song Han
+
+The abstract from the paper is:
+
+*This paper presents SANA-Sprint, an efficient diffusion model for ultra-fast text-to-image (T2I) generation. SANA-Sprint is built on a pre-trained foundation model and augmented with hybrid distillation, dramatically reducing inference steps from 20 to 1-4. We introduce three key innovations: (1) We propose a training-free approach that transforms a pre-trained flow-matching model for continuous-time consistency distillation (sCM), eliminating costly training from scratch and achieving high training efficiency. Our hybrid distillation strategy combines sCM with latent adversarial distillation (LADD): sCM ensures alignment with the teacher model, while LADD enhances single-step generation fidelity. (2) SANA-Sprint is a unified step-adaptive model that achieves high-quality generation in 1-4 steps, eliminating step-specific training and improving efficiency. (3) We integrate ControlNet with SANA-Sprint for real-time interactive image generation, enabling instant visual feedback for user interaction. SANA-Sprint establishes a new Pareto frontier in speed-quality tradeoffs, achieving state-of-the-art performance with 7.59 FID and 0.74 GenEval in only 1 step — outperforming FLUX-schnell (7.94 FID / 0.71 GenEval) while being 10× faster (0.1s vs 1.1s on H100). It also achieves 0.1s (T2I) and 0.25s (ControlNet) latency for 1024×1024 images on H100, and 0.31s (T2I) on an RTX 4090, showcasing its exceptional efficiency and potential for AI-powered consumer applications (AIPC). Code and pre-trained models will be open-sourced.*
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+This pipeline was contributed by [lawrence-cj](https://github.com/lawrence-cj), [shuchen Xue](https://github.com/scxue) and [Enze Xie](https://github.com/xieenze). The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://huggingface.co/Efficient-Large-Model/).
+
+Available models:
+
+|                                                                    Model                                                                    | Recommended dtype |
+|:-------------------------------------------------------------------------------------------------------------------------------------------:|:-----------------:|
+| [`Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers) | `torch.bfloat16`  |
+| [`Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers) | `torch.bfloat16`  |
+
+Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-sprint-67d6810d65235085b3b17c76) collection for more information.
+
+Note: The recommended dtype mentioned is for the transformer weights. The text encoder must stay in `torch.bfloat16` and VAE weights must stay in `torch.bfloat16` or `torch.float32` for the model to work correctly. Please refer to the inference example below to see how to load the model with the recommended dtype. 
+
+
+## Quantization
+
+Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
+
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`SanaSprintPipeline`] for inference with bitsandbytes.
+
+```py
+import torch
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, SanaTransformer2DModel, SanaSprintPipeline
+from transformers import BitsAndBytesConfig as BitsAndBytesConfig, AutoModel
+
+quant_config = BitsAndBytesConfig(load_in_8bit=True)
+text_encoder_8bit = AutoModel.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
+    subfolder="text_encoder",
+    quantization_config=quant_config,
+    torch_dtype=torch.bfloat16,
+)
+
+quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
+transformer_8bit = SanaTransformer2DModel.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
+    subfolder="transformer",
+    quantization_config=quant_config,
+    torch_dtype=torch.bfloat16,
+)
+
+pipeline = SanaSprintPipeline.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
+    text_encoder=text_encoder_8bit,
+    transformer=transformer_8bit,
+    torch_dtype=torch.bfloat16,
+    device_map="balanced",
+)
+
+prompt = "a tiny astronaut hatching from an egg on the moon"
+image = pipeline(prompt).images[0]
+image.save("sana.png")
+```
+
+## Setting `max_timesteps`
+
+Users can tweak the `max_timesteps` value for experimenting with the visual quality of the generated outputs. The default `max_timesteps` value was obtained with an inference-time search process. For more details about it, check out the paper.
+
+## SanaSprintPipeline
+
+[[autodoc]] SanaSprintPipeline
+  - all
+  - __call__
+
+
+## SanaPipelineOutput
+
+[[autodoc]] pipelines.sana.pipeline_output.SanaPipelineOutput
@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.

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

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

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

 Stable Diffusion XL (SDXL) was proposed in [SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis](https://huggingface.co/papers/2307.01952) by Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach.
@@ -14,22 +14,405 @@

 # Wan

+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.

 <!-- TODO(aryan): update abstract once paper is out -->

+## Generating Videos with Wan 2.1
+
+We will first need to install some addtional dependencies.
+
+```shell
+pip install -u ftfy imageio-ffmpeg imageio
+```
+
+### Text to Video Generation
+
+The following example requires 11GB VRAM to run and uses the smaller `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` model. You can switch it out
+for the larger `Wan2.1-I2V-14B-720P-Diffusers` or `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` if you have at least 35GB VRAM available.
+
+```python
+from diffusers import WanPipeline
+from diffusers.utils import export_to_video
+
+# Available models: Wan-AI/Wan2.1-I2V-14B-720P-Diffusers or Wan-AI/Wan2.1-I2V-14B-480P-Diffusers
+model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+
+pipe = WanPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
+pipe.enable_model_cpu_offload()
+
+prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+num_frames = 33
+
+frames = pipe(prompt=prompt, negative_prompt=negative_prompt, num_frames=num_frames).frames[0]
+export_to_video(frames, "wan-t2v.mp4", fps=16)
+```
+
 <Tip>
-
-Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
-
+You can improve the quality of the generated video by running the decoding step in full precision.
 </Tip>

-Recommendations for inference:
- VAE in `torch.float32` for better decoding quality.
- `num_frames` should be of the form `4 * k + 1`, for example `49` or `81`.
- For smaller resolution videos, try lower values of `shift` (between `2.0` to `5.0`) in the [Scheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete#diffusers.FlowMatchEulerDiscreteScheduler.shift). For larger resolution videos, try higher values (between `7.0` and `12.0`). The default value is `3.0` for Wan.
+```python
+from diffusers import WanPipeline, AutoencoderKLWan
+from diffusers.utils import export_to_video

-### Using a custom scheduler
+model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+
+# replace this with pipe.to("cuda") if you have sufficient VRAM
+pipe.enable_model_cpu_offload()
+
+prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+num_frames = 33
+
+frames = pipe(prompt=prompt, num_frames=num_frames).frames[0]
+export_to_video(frames, "wan-t2v.mp4", fps=16)
+```
+
+### Image to Video Generation
+
+The Image to Video pipeline requires loading the `AutoencoderKLWan` and the `CLIPVisionModel` components in full precision. The following example will need at least
+35GB of VRAM to run.
+
+```python
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, WanImageToVideoPipeline
+from diffusers.utils import export_to_video, load_image
+from transformers import CLIPVisionModel
+
+# Available models: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers, Wan-AI/Wan2.1-I2V-14B-720P-Diffusers
+model_id = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(
+    model_id, subfolder="image_encoder", torch_dtype=torch.float32
+)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id, vae=vae, image_encoder=image_encoder, torch_dtype=torch.bfloat16
+)
+
+# replace this with pipe.to("cuda") if you have sufficient VRAM
+pipe.enable_model_cpu_offload()
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+)
+
+max_area = 480 * 832
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+image = image.resize((width, height))
+
+prompt = (
+    "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in "
+    "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+)
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+num_frames = 33
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    num_frames=num_frames,
+    guidance_scale=5.0,
+).frames[0]
+export_to_video(output, "wan-i2v.mp4", fps=16)
+```
+
+### Video to Video Generation
+
+```python
+import torch
+from diffusers.utils import load_video, export_to_video
+from diffusers import AutoencoderKLWan, WanVideoToVideoPipeline, UniPCMultistepScheduler
+
+# Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(
+    model_id, subfolder="vae", torch_dtype=torch.float32
+)
+pipe = WanVideoToVideoPipeline.from_pretrained(
+    model_id, vae=vae, torch_dtype=torch.bfloat16
+)
+flow_shift = 3.0  # 5.0 for 720P, 3.0 for 480P
+pipe.scheduler = UniPCMultistepScheduler.from_config(
+    pipe.scheduler.config, flow_shift=flow_shift
+)
+# change to pipe.to("cuda") if you have sufficient VRAM
+pipe.enable_model_cpu_offload()
+
+prompt = "A robot standing on a mountain top. The sun is setting in the background"
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+video = load_video(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/hiker.mp4"
+)
+output = pipe(
+    video=video,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=480,
+    width=512,
+    guidance_scale=7.0,
+    strength=0.7,
+).frames[0]
+
+export_to_video(output, "wan-v2v.mp4", fps=16)
+```
+
+## Memory Optimizations for Wan 2.1
+
+Base inference with the large 14B Wan 2.1 models can take up to 35GB of VRAM when generating videos at 720p resolution. We'll outline a few memory optimizations we can apply to reduce the VRAM required to run the model.
+
+We'll use `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers` model in these examples to demonstrate the memory savings, but the techniques are applicable to all model checkpoints.
+
+### Group Offloading the Transformer and UMT5 Text Encoder
+
+Find more information about group offloading [here](../optimization/memory.md)
+
+#### Block Level Group Offloading
+
+We can reduce our VRAM requirements by applying group offloading to the larger model components of the pipeline; the `WanTransformer3DModel` and `UMT5EncoderModel`. Group offloading will break up the individual modules of a model and offload/onload them onto your GPU as needed during inference. In this example, we'll apply `block_level` offloading, which will group the modules in a model into blocks of size `num_blocks_per_group` and offload/onload them to GPU. Moving to between CPU and GPU does add latency to the inference process. You can trade off between latency and memory savings by increasing or decreasing the `num_blocks_per_group`.
+
+The following example will now only require 14GB of VRAM to run, but will take approximately 30 minutes to generate a video.
+
+```python
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, WanTransformer3DModel, WanImageToVideoPipeline
+from diffusers.hooks.group_offloading import apply_group_offloading
+from diffusers.utils import export_to_video, load_image
+from transformers import UMT5EncoderModel, CLIPVisionModel
+
+# Available models: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers, Wan-AI/Wan2.1-I2V-14B-720P-Diffusers
+model_id = "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(
+    model_id, subfolder="image_encoder", torch_dtype=torch.float32
+)
+
+text_encoder = UMT5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+transformer = WanTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+
+onload_device = torch.device("cuda")
+offload_device = torch.device("cpu")
+
+apply_group_offloading(text_encoder,
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="block_level",
+    num_blocks_per_group=4
+)
+
+transformer.enable_group_offload(
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="block_level",
+    num_blocks_per_group=4,
+)
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id,
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    image_encoder=image_encoder,
+    torch_dtype=torch.bfloat16
+)
+# Since we've offloaded the larger models alrady, we can move the rest of the model components to GPU
+pipe.to("cuda")
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+)
+
+max_area = 720 * 832
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+image = image.resize((width, height))
+
+prompt = (
+    "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in "
+    "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+)
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+num_frames = 33
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    num_frames=num_frames,
+    guidance_scale=5.0,
+).frames[0]
+
+export_to_video(output, "wan-i2v.mp4", fps=16)
+```
+
+#### Block Level Group Offloading with CUDA Streams
+
+We can speed up group offloading inference, by enabling the use of [CUDA streams](https://pytorch.org/docs/stable/generated/torch.cuda.Stream.html). However, using CUDA streams requires moving the model parameters into pinned memory. This allocation is handled by Pytorch under the hood, and can result in a significant spike in CPU RAM usage. Please consider this option if your CPU RAM is atleast 2X the size of the model you are group offloading.
+
+In the following example we will use CUDA streams when group offloading the `WanTransformer3DModel`. When testing on an A100, this example will require 14GB of VRAM, 52GB of CPU RAM, but will generate a video in approximately 9 minutes.
+
+```python
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, WanTransformer3DModel, WanImageToVideoPipeline
+from diffusers.hooks.group_offloading import apply_group_offloading
+from diffusers.utils import export_to_video, load_image
+from transformers import UMT5EncoderModel, CLIPVisionModel
+
+# Available models: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers, Wan-AI/Wan2.1-I2V-14B-720P-Diffusers
+model_id = "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(
+    model_id, subfolder="image_encoder", torch_dtype=torch.float32
+)
+
+text_encoder = UMT5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+transformer = WanTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+
+onload_device = torch.device("cuda")
+offload_device = torch.device("cpu")
+
+apply_group_offloading(text_encoder,
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="block_level",
+    num_blocks_per_group=4
+)
+
+transformer.enable_group_offload(
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="leaf_level",
+    use_stream=True
+)
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id,
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    image_encoder=image_encoder,
+    torch_dtype=torch.bfloat16
+)
+# Since we've offloaded the larger models alrady, we can move the rest of the model components to GPU
+pipe.to("cuda")
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+)
+
+max_area = 720 * 832
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+image = image.resize((width, height))
+
+prompt = (
+    "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in "
+    "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+)
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+num_frames = 33
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    num_frames=num_frames,
+    guidance_scale=5.0,
+).frames[0]
+
+export_to_video(output, "wan-i2v.mp4", fps=16)
+```
+
+### Applying Layerwise Casting to the Transformer
+
+Find more information about layerwise casting [here](../optimization/memory.md)
+
+In this example, we will model offloading with layerwise casting. Layerwise casting will downcast each layer's weights to `torch.float8_e4m3fn`, temporarily upcast to `torch.bfloat16` during the forward pass of the layer, then revert to `torch.float8_e4m3fn` afterward. This approach reduces memory requirements by approximately 50% while introducing a minor quality reduction in the generated video due to the precision trade-off.
+
+This example will require 20GB of VRAM.
+
+```python
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, WanTransformer3DModel, WanImageToVideoPipeline
+from diffusers.hooks.group_offloading import apply_group_offloading
+from diffusers.utils import export_to_video, load_image
+from transformers import UMT5EncoderModel, CLIPVisionModel
+
+model_id = "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(
+    model_id, subfolder="image_encoder", torch_dtype=torch.float32
+)
+text_encoder = UMT5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.bfloat16)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+
+transformer = WanTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+transformer.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
+
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id,
+    vae=vae,
+    transformer=transformer,
+    text_encoder=text_encoder,
+    image_encoder=image_encoder,
+    torch_dtype=torch.bfloat16
+)
+pipe.enable_model_cpu_offload()
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg")
+
+max_area = 720 * 832
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+image = image.resize((width, height))
+prompt = (
+    "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in "
+    "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+)
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+num_frames = 33
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    num_frames=num_frames,
+    num_inference_steps=50,
+    guidance_scale=5.0,
+).frames[0]
+export_to_video(output, "wan-i2v.mp4", fps=16)
+```
+
+## Using a Custom Scheduler

 Wan can be used with many different schedulers, each with their own benefits regarding speed and generation quality. By default, Wan uses the `UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)` scheduler. You can use a different scheduler as follows:

@@ -45,11 +428,10 @@ pipe = WanPipeline.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", scheduler
 pipe.scheduler = <CUSTOM_SCHEDULER_HERE>
 ```

-### Using single file loading with Wan
-
-The `WanTransformer3DModel` and `AutoencoderKLWan` models support loading checkpoints in their original format via the `from_single_file` loading 
-method. 
+## Using Single File Loading with Wan 2.1

+The `WanTransformer3DModel` and `AutoencoderKLWan` models support loading checkpoints in their original format via the `from_single_file` loading
+method.

 ```python
 import torch
@@ -61,6 +443,11 @@ transformer = WanTransformer3DModel.from_single_file(ckpt_path, torch_dtype=torc
 pipe = WanPipeline.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", transformer=transformer)
 ```

+## Recommendations for Inference
+- Keep `AutencoderKLWan` in `torch.float32` for better decoding quality.
+- `num_frames` should satisfy the following constraint: `(num_frames - 1) % 4 == 0`
+- For smaller resolution videos, try lower values of `shift` (between `2.0` to `5.0`) in the [Scheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete#diffusers.FlowMatchEulerDiscreteScheduler.shift). For larger resolution videos, try higher values (between `7.0` and `12.0`). The default value is `3.0` for Wan.
+
 ## WanPipeline

 [[autodoc]] WanPipeline
@@ -3,3 +3,7 @@
 ## Remote Decode

 [[autodoc]] utils.remote_utils.remote_decode
+
+## Remote Encode
+
+[[autodoc]] utils.remote_utils.remote_encode
@@ -36,7 +36,7 @@ Hybrid Inference offers a fast and simple way to offload local generation requir
 ## Available Models

 * **VAE Decode 🖼️:** Quickly decode latent representations into high-quality images without compromising performance or workflow speed.
-* **VAE Encode 🔢 (coming soon):** Efficiently encode images into latent representations for generation and training.
+* **VAE Encode 🔢:** Efficiently encode images into latent representations for generation and training.
 * **Text Encoders 📃 (coming soon):** Compute text embeddings for your prompts quickly and accurately, ensuring a smooth and high-quality workflow.

 ---
@@ -46,9 +46,15 @@ Hybrid Inference offers a fast and simple way to offload local generation requir
 * **[SD.Next](https://github.com/vladmandic/sdnext):** All-in-one UI with direct supports Hybrid Inference.
 * **[ComfyUI-HFRemoteVae](https://github.com/kijai/ComfyUI-HFRemoteVae):** ComfyUI node for Hybrid Inference.

+## Changelog
+
+- March 10 2025: Added VAE encode
+- March 2 2025: Initial release with VAE decoding
+
 ## Contents

-The documentation is organized into two sections:
+The documentation is organized into three sections:

 * **VAE Decode** Learn the basics of how to use VAE Decode with Hybrid Inference.
+* **VAE Encode** Learn the basics of how to use VAE Encode with Hybrid Inference.
 * **API Reference** Dive into task-specific settings and parameters.
@@ -0,0 +1,183 @@
+# Getting Started: VAE Encode with Hybrid Inference
+
+VAE encode is used for training, image-to-image and image-to-video - turning into images or videos into latent representations.
+
+## Memory
+
+These tables demonstrate the VRAM requirements for VAE encode with SD v1 and SD XL on different GPUs.
+
+For the majority of these GPUs the memory usage % dictates other models (text encoders, UNet/Transformer) must be offloaded, or tiled encoding has to be used which increases time taken and impacts quality.
+
+<details><summary>SD v1.5</summary>
+
+| GPU                           | Resolution   |   Time (seconds) |   Memory (%) |   Tiled Time (secs) |   Tiled Memory (%) |
+|:------------------------------|:-------------|-----------------:|-------------:|--------------------:|-------------------:|
+| NVIDIA GeForce RTX 4090       | 512x512      |            0.015 |      3.51901 |               0.015 |            3.51901 |
+| NVIDIA GeForce RTX 4090       | 256x256      |            0.004 |      1.3154  |               0.005 |            1.3154  |
+| NVIDIA GeForce RTX 4090       | 2048x2048    |            0.402 |     47.1852  |               0.496 |            3.51901 |
+| NVIDIA GeForce RTX 4090       | 1024x1024    |            0.078 |     12.2658  |               0.094 |            3.51901 |
+| NVIDIA GeForce RTX 4080 SUPER | 512x512      |            0.023 |      5.30105 |               0.023 |            5.30105 |
+| NVIDIA GeForce RTX 4080 SUPER | 256x256      |            0.006 |      1.98152 |               0.006 |            1.98152 |
+| NVIDIA GeForce RTX 4080 SUPER | 2048x2048    |            0.574 |     71.08    |               0.656 |            5.30105 |
+| NVIDIA GeForce RTX 4080 SUPER | 1024x1024    |            0.111 |     18.4772  |               0.14  |            5.30105 |
+| NVIDIA GeForce RTX 3090       | 512x512      |            0.032 |      3.52782 |               0.032 |            3.52782 |
+| NVIDIA GeForce RTX 3090       | 256x256      |            0.01  |      1.31869 |               0.009 |            1.31869 |
+| NVIDIA GeForce RTX 3090       | 2048x2048    |            0.742 |     47.3033  |               0.954 |            3.52782 |
+| NVIDIA GeForce RTX 3090       | 1024x1024    |            0.136 |     12.2965  |               0.207 |            3.52782 |
+| NVIDIA GeForce RTX 3080       | 512x512      |            0.036 |      8.51761 |               0.036 |            8.51761 |
+| NVIDIA GeForce RTX 3080       | 256x256      |            0.01  |      3.18387 |               0.01  |            3.18387 |
+| NVIDIA GeForce RTX 3080       | 2048x2048    |            0.863 |     86.7424  |               1.191 |            8.51761 |
+| NVIDIA GeForce RTX 3080       | 1024x1024    |            0.157 |     29.6888  |               0.227 |            8.51761 |
+| NVIDIA GeForce RTX 3070       | 512x512      |            0.051 |     10.6941  |               0.051 |           10.6941  |
+| NVIDIA GeForce RTX 3070       | 256x256      |            0.015 |      3.99743 |               0.015 |            3.99743 |
+| NVIDIA GeForce RTX 3070       | 2048x2048    |            1.217 |     96.054   |               1.482 |           10.6941  |
+| NVIDIA GeForce RTX 3070       | 1024x1024    |            0.223 |     37.2751  |               0.327 |           10.6941  |
+
+
+</details>
+
+<details><summary>SDXL</summary>
+
+| GPU                           | Resolution   |   Time (seconds) |   Memory Consumed (%) |   Tiled Time (seconds) |   Tiled Memory (%) |
+|:------------------------------|:-------------|-----------------:|----------------------:|-----------------------:|-------------------:|
+| NVIDIA GeForce RTX 4090       | 512x512      |            0.029 |               4.95707 |                  0.029 |            4.95707 |
+| NVIDIA GeForce RTX 4090       | 256x256      |            0.007 |               2.29666 |                  0.007 |            2.29666 |
+| NVIDIA GeForce RTX 4090       | 2048x2048    |            0.873 |              66.3452  |                  0.863 |           15.5649  |
+| NVIDIA GeForce RTX 4090       | 1024x1024    |            0.142 |              15.5479  |                  0.143 |           15.5479  |
+| NVIDIA GeForce RTX 4080 SUPER | 512x512      |            0.044 |               7.46735 |                  0.044 |            7.46735 |
+| NVIDIA GeForce RTX 4080 SUPER | 256x256      |            0.01  |               3.4597  |                  0.01  |            3.4597  |
+| NVIDIA GeForce RTX 4080 SUPER | 2048x2048    |            1.317 |              87.1615  |                  1.291 |           23.447   |
+| NVIDIA GeForce RTX 4080 SUPER | 1024x1024    |            0.213 |              23.4215  |                  0.214 |           23.4215  |
+| NVIDIA GeForce RTX 3090       | 512x512      |            0.058 |               5.65638 |                  0.058 |            5.65638 |
+| NVIDIA GeForce RTX 3090       | 256x256      |            0.016 |               2.45081 |                  0.016 |            2.45081 |
+| NVIDIA GeForce RTX 3090       | 2048x2048    |            1.755 |              77.8239  |                  1.614 |           18.4193  |
+| NVIDIA GeForce RTX 3090       | 1024x1024    |            0.265 |              18.4023  |                  0.265 |           18.4023  |
+| NVIDIA GeForce RTX 3080       | 512x512      |            0.064 |              13.6568  |                  0.064 |           13.6568  |
+| NVIDIA GeForce RTX 3080       | 256x256      |            0.018 |               5.91728 |                  0.018 |            5.91728 |
+| NVIDIA GeForce RTX 3080       | 2048x2048    |          OOM     |             OOM       |                  1.866 |           44.4717  |
+| NVIDIA GeForce RTX 3080       | 1024x1024    |            0.302 |              44.4308  |                  0.302 |           44.4308  |
+| NVIDIA GeForce RTX 3070       | 512x512      |            0.093 |              17.1465  |                  0.093 |           17.1465  |
+| NVIDIA GeForce RTX 3070       | 256x256      |            0.025 |               7.42931 |                  0.026 |            7.42931 |
+| NVIDIA GeForce RTX 3070       | 2048x2048    |          OOM     |             OOM       |                  2.674 |           55.8355  |
+| NVIDIA GeForce RTX 3070       | 1024x1024    |            0.443 |              55.7841  |                  0.443 |           55.7841  |
+
+</details>
+
+## Available VAEs
+
+|   | **Endpoint** | **Model** |
+|:-:|:-----------:|:--------:|
+| **Stable Diffusion v1** | [https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud](https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud) | [`stabilityai/sd-vae-ft-mse`](https://hf.co/stabilityai/sd-vae-ft-mse) |
+| **Stable Diffusion XL** | [https://xjqqhmyn62rog84g.us-east-1.aws.endpoints.huggingface.cloud](https://xjqqhmyn62rog84g.us-east-1.aws.endpoints.huggingface.cloud) | [`madebyollin/sdxl-vae-fp16-fix`](https://hf.co/madebyollin/sdxl-vae-fp16-fix) |
+| **Flux** | [https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud](https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud) | [`black-forest-labs/FLUX.1-schnell`](https://hf.co/black-forest-labs/FLUX.1-schnell) |
+
+
+> [!TIP]
+> Model support can be requested [here](https://github.com/huggingface/diffusers/issues/new?template=remote-vae-pilot-feedback.yml).
+
+
+## Code
+
+> [!TIP]
+> Install `diffusers` from `main` to run the code: `pip install git+https://github.com/huggingface/diffusers@main`
+
+
+A helper method simplifies interacting with Hybrid Inference.
+
+```python
+from diffusers.utils.remote_utils import remote_encode
+```
+
+### Basic example
+
+Let's encode an image, then decode it to demonstrate.
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"/>
+</figure>
+
+<details><summary>Code</summary>
+
+```python
+from diffusers.utils import load_image
+from diffusers.utils.remote_utils import remote_decode
+
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg?download=true")
+
+latent = remote_encode(
+    endpoint="https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud/",
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+
+decoded = remote_decode(
+    endpoint="https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/decoded.png"/>
+</figure>
+
+
+### Generation
+
+Now let's look at a generation example, we'll encode the image, generate then remotely decode too!
+
+<details><summary>Code</summary>
+
+```python
+import torch
+from diffusers import StableDiffusionImg2ImgPipeline
+from diffusers.utils import load_image
+from diffusers.utils.remote_utils import remote_decode, remote_encode
+
+pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    vae=None,
+).to("cuda")
+
+init_image = load_image(
+    "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+)
+init_image = init_image.resize((768, 512))
+
+init_latent = remote_encode(
+    endpoint="https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud/",
+    image=init_image,
+    scaling_factor=0.18215,
+)
+
+prompt = "A fantasy landscape, trending on artstation"
+latent = pipe(
+    prompt=prompt,
+    image=init_latent,
+    strength=0.75,
+    output_type="latent",
+).images
+
+image = remote_decode(
+    endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    scaling_factor=0.18215,
+)
+image.save("fantasy_landscape.jpg")
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/fantasy_landscape.png"/>
+</figure>
+
+## Integrations
+
+* **[SD.Next](https://github.com/vladmandic/sdnext):** All-in-one UI with direct supports Hybrid Inference.
+* **[ComfyUI-HFRemoteVae](https://github.com/kijai/ComfyUI-HFRemoteVae):** ComfyUI node for Hybrid Inference.
@@ -161,10 +161,10 @@ Your Python environment will find the `main` version of 🤗 Diffusers on the ne

 Model weights and files are downloaded from the Hub to a cache which is usually your home directory. You can change the cache location by specifying the `HF_HOME` or `HUGGINFACE_HUB_CACHE` environment variables or configuring the `cache_dir` parameter in methods like [`~DiffusionPipeline.from_pretrained`].

-Cached files allow you to run 🤗 Diffusers offline. To prevent 🤗 Diffusers from connecting to the internet, set the `HF_HUB_OFFLINE` environment variable to `True` and 🤗 Diffusers will only load previously downloaded files in the cache.
+Cached files allow you to run 🤗 Diffusers offline. To prevent 🤗 Diffusers from connecting to the internet, set the `HF_HUB_OFFLINE` environment variable to `1` and 🤗 Diffusers will only load previously downloaded files in the cache.

 ```shell
-export HF_HUB_OFFLINE=True
+export HF_HUB_OFFLINE=1
 ```

 For more details about managing and cleaning the cache, take a look at the [caching](https://huggingface.co/docs/huggingface_hub/guides/manage-cache) guide.
@@ -179,14 +179,16 @@ Telemetry is only sent when loading models and pipelines from the Hub,
 and it is not collected if you're loading local files.

 We understand that not everyone wants to share additional information,and we respect your privacy.
-You can disable telemetry collection by setting the `DISABLE_TELEMETRY` environment variable from your terminal:
+You can disable telemetry collection by setting the `HF_HUB_DISABLE_TELEMETRY` environment variable from your terminal:

 On Linux/MacOS:
+
 ```bash
-export DISABLE_TELEMETRY=YES
+export HF_HUB_DISABLE_TELEMETRY=1
 ```

 On Windows:
+
 ```bash
-set DISABLE_TELEMETRY=YES
+set HF_HUB_DISABLE_TELEMETRY=1
 ```
@@ -178,6 +178,9 @@ pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch
 # We can utilize the enable_group_offload method for Diffusers model implementations
 pipe.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True)

+# Uncomment the following to also allow recording the current streams.
+# pipe.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True, record_stream=True)
+
 # For any other model implementations, the apply_group_offloading function can be used
 apply_group_offloading(pipe.text_encoder, onload_device=onload_device, offload_type="block_level", num_blocks_per_group=2)
 apply_group_offloading(pipe.vae, onload_device=onload_device, offload_type="leaf_level")
@@ -198,6 +201,19 @@ export_to_video(video, "output.mp4", fps=8)

 Group offloading (for CUDA devices with support for asynchronous data transfer streams) overlaps data transfer and computation to reduce the overall execution time compared to sequential offloading. This is enabled using layer prefetching with CUDA streams. The next layer to be executed is loaded onto the accelerator device while the current layer is being executed - this increases the memory requirements slightly. Group offloading also supports leaf-level offloading (equivalent to sequential CPU offloading) but can be made much faster when using streams.

+<Tip>
+
+- Group offloading may not work with all models out-of-the-box. If the forward implementations of the model contain weight-dependent device-casting of inputs, it may clash with the offloading mechanism's handling of device-casting.
+- The `offload_type` parameter can be set to either `block_level` or `leaf_level`. `block_level` offloads groups of `torch::nn::ModuleList` or `torch::nn:Sequential` modules based on a configurable attribute `num_blocks_per_group`. For example, if you set `num_blocks_per_group=2` on a standard transformer model containing 40 layers, it will onload/offload 2 layers at a time for a total of 20 onload/offloads. This drastically reduces the VRAM requirements. `leaf_level` offloads individual layers at the lowest level, which is equivalent to sequential offloading. However, unlike sequential offloading, group offloading can be made much faster when using streams, with minimal compromise to end-to-end generation time.
+- The `use_stream` parameter can be used with CUDA devices to enable prefetching layers for onload. It defaults to `False`. Layer prefetching allows overlapping computation and data transfer of model weights, which drastically reduces the overall execution time compared to other offloading methods. However, it can increase the CPU RAM usage significantly. Ensure that available CPU RAM that is at least twice the size of the model when setting `use_stream=True`. You can find more information about CUDA streams [here](https://pytorch.org/docs/stable/generated/torch.cuda.Stream.html)
+- If specifying `use_stream=True` on VAEs with tiling enabled, make sure to do a dummy forward pass (possibly with dummy inputs) before the actual inference to avoid device-mismatch errors. This may not work on all implementations. Please open an issue if you encounter any problems.
+- The parameter `low_cpu_mem_usage` can be set to `True` to reduce CPU memory usage when using streams for group offloading. This is useful when the CPU memory is the bottleneck, but it may counteract the benefits of using streams and increase the overall execution time. The CPU memory savings come from creating pinned-tensors on-the-fly instead of pre-pinning them. This parameter is better suited for using `leaf_level` offloading.
+- When using `use_stream=True`, users can additionally specify `record_stream=True` to get better speedups at the expense of slightly increased memory usage. Refer to the [official PyTorch docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html) to know more about this.
+
+For more information about available parameters and an explanation of how group offloading works, refer to [`~hooks.group_offloading.apply_group_offloading`].
+
+</Tip>
+
 ## FP8 layerwise weight-casting

 PyTorch supports `torch.float8_e4m3fn` and `torch.float8_e5m2` as weight storage dtypes, but they can't be used for computation in many different tensor operations due to unimplemented kernel support. However, you can use these dtypes to store model weights in fp8 precision and upcast them on-the-fly when the layers are used in the forward pass. This is known as layerwise weight-casting.
@@ -235,6 +251,14 @@ In the above example, layerwise casting is enabled on the transformer component

 However, you gain more control and flexibility by directly utilizing the [`~hooks.layerwise_casting.apply_layerwise_casting`] function instead of [`~ModelMixin.enable_layerwise_casting`].

+<Tip>
+
+- Layerwise casting may not work with all models out-of-the-box. Sometimes, the forward implementations of the model might contain internal typecasting of weight values. Such implementations are not supported due to the currently simplistic implementation of layerwise casting, which assumes that the forward pass is independent of the weight precision and that the input dtypes are always in `compute_dtype`. An example of an incompatible implementation can be found [here](https://github.com/huggingface/transformers/blob/7f5077e53682ca855afc826162b204ebf809f1f9/src/transformers/models/t5/modeling_t5.py#L294-L299).
+- Layerwise casting may fail on custom modeling implementations that make use of [PEFT](https://github.com/huggingface/peft) layers. Some minimal checks to handle this case is implemented but is not extensively tested or guaranteed to work in all cases.
+- It can be also be applied partially to specific layers of a model. Partially applying layerwise casting can either be done manually by calling the `apply_layerwise_casting` function on specific internal modules, or by specifying the `skip_modules_pattern` and `skip_modules_classes` parameters for a root module. These parameters are particularly useful for layers such as normalization and modulation.
+
+</Tip>
+
 ## Channels-last memory format

 The channels-last memory format is an alternative way of ordering NCHW tensors in memory to preserve dimension ordering. Channels-last tensors are ordered in such a way that the channels become the densest dimension (storing images pixel-per-pixel). Since not all operators currently support the channels-last format, it may result in worst performance but you should still try and see if it works for your model.
@@ -12,6 +12,9 @@ specific language governing permissions and limitations under the License.

 # Metal Performance Shaders (MPS)

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

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

 <Tip warning={true}>

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

 </Tip>

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

 ## Troubleshoot

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

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

 Attention slicing performs the costly attention operation in multiple steps instead of all at once. It usually improves performance by ~20% in computers without universal memory, but we've observed *better performance* in most Apple silicon computers unless you have 64GB of RAM or more.
+
+### Batch inference
+
+Generating multiple prompts in a batch can crash or fail to work reliably. If this is the case, try iterating instead of batching.
@@ -95,6 +95,23 @@ Use the Space below to gauge a pipeline's memory requirements before you downloa
    ></iframe>
 </div>

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

 To load a pipeline locally, use [git-lfs](https://git-lfs.github.com/) to manually download a checkpoint to your local disk.
@@ -194,6 +194,59 @@ Currently, [`~loaders.StableDiffusionLoraLoaderMixin.set_adapters`] only support

 </Tip>

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

 Other popular LoRA trainers from the community include those by [Kohya](https://github.com/kohya-ss/sd-scripts/) and [TheLastBen](https://github.com/TheLastBen/fast-stable-diffusion). These trainers create different LoRA checkpoints than those trained by 🤗 Diffusers, but they can still be loaded in the same way.
@@ -66,12 +66,6 @@ from accelerate.utils import write_basic_config
 write_basic_config()
 ```

-## 원을 채우는 데이터셋
-
-원본 데이터셋은 ControlNet [repo](https://huggingface.co/lllyasviel/ControlNet/blob/main/training/fill50k.zip)에 올라와있지만, 우리는 [여기](https://huggingface.co/datasets/fusing/fill50k)에 새롭게 다시 올려서 🤗 Datasets 과 호환가능합니다. 그래서 학습 스크립트 상에서 데이터 불러오기를 다룰 수 있습니다.
-
-우리의 학습 예시는 원래 ControlNet의 학습에 쓰였던 [`stable-diffusion-v1-5/stable-diffusion-v1-5`](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5)을 사용합니다. 그렇지만 ControlNet은 대응되는 어느 Stable Diffusion 모델([`CompVis/stable-diffusion-v1-4`](https://huggingface.co/CompVis/stable-diffusion-v1-4)) 혹은 [`stabilityai/stable-diffusion-2-1`](https://huggingface.co/stabilityai/stable-diffusion-2-1)의 증가를 위해 학습될 수 있습니다.
-
 자체 데이터셋을 사용하기 위해서는 [학습을 위한 데이터셋 생성하기](create_dataset) 가이드를 확인하세요.

 ## 학습
@@ -79,13 +79,13 @@ This command will prompt you for a token. Copy-paste yours from your [settings/t
 ### Target Modules
 When LoRA was first adapted from language models to diffusion models, it was applied to the cross-attention layers in the Unet that relate the image representations with the prompts that describe them. 
 More recently, SOTA text-to-image diffusion models replaced the Unet with a diffusion Transformer(DiT). With this change, we may also want to explore 
-applying LoRA training onto different types of layers and blocks. To allow more flexibility and control over the targeted modules we added `--lora_layers`- in which you can specify in a comma seperated string
+applying LoRA training onto different types of layers and blocks. To allow more flexibility and control over the targeted modules we added `--lora_layers`- in which you can specify in a comma separated string
 the exact modules for LoRA training. Here are some examples of target modules you can provide: 
 - for attention only layers: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0"`
 - to train the same modules as in the fal trainer: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2"`
 - to train the same modules as in ostris ai-toolkit / replicate trainer: `--lora_blocks="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2,norm1_context.linear, norm1.linear,norm.linear,proj_mlp,proj_out"`
 > [!NOTE]
-> `--lora_layers` can also be used to specify which **blocks** to apply LoRA training to. To do so, simply add a block prefix to each layer in the comma seperated string:
+> `--lora_layers` can also be used to specify which **blocks** to apply LoRA training to. To do so, simply add a block prefix to each layer in the comma separated string:
 > **single DiT blocks**: to target the ith single transformer block, add the prefix `single_transformer_blocks.i`, e.g. - `single_transformer_blocks.i.attn.to_k`
 > **MMDiT blocks**: to target the ith MMDiT block, add the prefix `transformer_blocks.i`, e.g. - `transformer_blocks.i.attn.to_k` 
 > [!NOTE]
@@ -1,7 +1,8 @@
-accelerate>=0.16.0
+accelerate>=0.31.0
 torchvision
-transformers>=4.25.1
+transformers>=4.41.2
 ftfy
 tensorboard
 Jinja2
-peft==0.7.0
+peft>=0.11.1
+sentencepiece
@@ -24,7 +24,7 @@ import re
 import shutil
 from contextlib import nullcontext
 from pathlib import Path
-from typing import List, Optional, Union
+from typing import List, Optional

 import numpy as np
 import torch
@@ -228,10 +228,20 @@ def log_validation(

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

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

    for tracker in accelerator.trackers:
        phase_name = "test" if is_final_validation else "validation"
@@ -378,7 +388,7 @@ def parse_args(input_args=None):
        default=None,
        help="the concept to use to initialize the new inserted tokens when training with "
        "--train_text_encoder_ti = True. By default, new tokens (<si><si+1>) are initialized with random value. "
-        "Alternatively, you could specify a different word/words whos value will be used as the starting point for the new inserted tokens. "
+        "Alternatively, you could specify a different word/words whose value will be used as the starting point for the new inserted tokens. "
        "--num_new_tokens_per_abstraction is ignored when initializer_concept is provided",
    )
    parser.add_argument(
@@ -657,15 +667,17 @@ def parse_args(input_args=None):
    parser.add_argument(
        "--adam_weight_decay_text_encoder", type=float, default=1e-03, help="Weight decay to use for text_encoder"
    )
+
    parser.add_argument(
        "--lora_layers",
        type=str,
        default=None,
        help=(
-            "The transformer modules to apply LoRA training on. Please specify the layers in a comma seperated. "
+            "The transformer modules to apply LoRA training on. Please specify the layers in a comma separated. "
            'E.g. - "to_k,to_q,to_v,to_out.0" will result in lora training of attention layers only. For more examples refer to https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/README_flux.md'
        ),
    )
+
    parser.add_argument(
        "--adam_epsilon",
        type=float,
@@ -738,6 +750,15 @@ def parse_args(input_args=None):
            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
        ),
    )
+    parser.add_argument(
+        "--upcast_before_saving",
+        action="store_true",
+        default=False,
+        help=(
+            "Whether to upcast the trained transformer layers to float32 before saving (at the end of training). "
+            "Defaults to precision dtype used for training to save memory"
+        ),
+    )
    parser.add_argument(
        "--prior_generation_precision",
        type=str,
@@ -818,9 +839,9 @@ class TokenEmbeddingsHandler:
        idx = 0
        for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
            assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )

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


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

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

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

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


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

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

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

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

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

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

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

    return prompt_embeds, pooled_prompt_embeds, text_ids


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

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

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

    transformer_lora_parameters = list(filter(lambda p: p.requires_grad, transformer.parameters()))
-
    if args.train_text_encoder:
        text_lora_parameters_one = list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
    # if we use textual inversion, we freeze all parameters except for the token embeddings
@@ -1737,7 +1662,8 @@ def main(args):
        for name, param in text_encoder_one.named_parameters():
            if "token_embedding" in name:
                # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-                param.data = param.to(dtype=torch.float32)
+                if args.mixed_precision == "fp16":
+                    param.data = param.to(dtype=torch.float32)
                param.requires_grad = True
                text_lora_parameters_one.append(param)
            else:
@@ -1747,7 +1673,8 @@ def main(args):
            for name, param in text_encoder_two.named_parameters():
                if "shared" in name:
                    # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-                    param.data = param.to(dtype=torch.float32)
+                    if args.mixed_precision == "fp16":
+                        param.data = param.to(dtype=torch.float32)
                    param.requires_grad = True
                    text_lora_parameters_two.append(param)
                else:
@@ -1828,6 +1755,7 @@ def main(args):
            optimizer_class = bnb.optim.AdamW8bit
        else:
            optimizer_class = torch.optim.AdamW
+
        optimizer = optimizer_class(
            params_to_optimize,
            betas=(args.adam_beta1, args.adam_beta2),
@@ -2021,6 +1949,7 @@ def main(args):
                lr_scheduler,
            )
        else:
+            print("I SHOULD BE HERE")
            transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
                transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler
            )
@@ -2125,7 +2054,7 @@ def main(args):
                if args.train_text_encoder:
                    text_encoder_one.train()
                    # set top parameter requires_grad = True for gradient checkpointing works
-                    accelerator.unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
+                    unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
                elif args.train_text_encoder_ti:  # textual inversion / pivotal tuning
                    text_encoder_one.train()
                if args.enable_t5_ti:
@@ -2137,6 +2066,11 @@ def main(args):
                pivoted_tr = True

        for step, batch in enumerate(train_dataloader):
+            models_to_accumulate = [transformer]
+            if not freeze_text_encoder:
+                models_to_accumulate.extend([text_encoder_one])
+                if args.enable_t5_ti:
+                    models_to_accumulate.extend([text_encoder_two])
            if pivoted_te:
                # stopping optimization of text_encoder params
                optimizer.param_groups[te_idx]["lr"] = 0.0
@@ -2145,7 +2079,7 @@ def main(args):
                logger.info(f"PIVOT TRANSFORMER {epoch}")
                optimizer.param_groups[0]["lr"] = 0.0

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

                # encode batch prompts when custom prompts are provided for each image -
@@ -2189,7 +2123,7 @@ def main(args):
                model_input = (model_input - vae_config_shift_factor) * vae_config_scaling_factor
                model_input = model_input.to(dtype=weight_dtype)

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

                latent_image_ids = FluxPipeline._prepare_latent_image_ids(
                    model_input.shape[0],
@@ -2228,7 +2162,7 @@ def main(args):
                )

                # handle guidance
-                if transformer.config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                    guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                    guidance = guidance.expand(model_input.shape[0])
                else:
@@ -2288,16 +2222,26 @@ def main(args):
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    if not freeze_text_encoder:
-                        if args.train_text_encoder:
+                        if args.train_text_encoder:  # text encoder tuning
                            params_to_clip = itertools.chain(transformer.parameters(), text_encoder_one.parameters())
                        elif pure_textual_inversion:
-                            params_to_clip = itertools.chain(
-                                text_encoder_one.parameters(), text_encoder_two.parameters()
-                            )
+                            if args.enable_t5_ti:
+                                params_to_clip = itertools.chain(
+                                    text_encoder_one.parameters(), text_encoder_two.parameters()
+                                )
+                            else:
+                                params_to_clip = itertools.chain(text_encoder_one.parameters())
                        else:
-                            params_to_clip = itertools.chain(
-                                transformer.parameters(), text_encoder_one.parameters(), text_encoder_two.parameters()
-                            )
+                            if args.enable_t5_ti:
+                                params_to_clip = itertools.chain(
+                                    transformer.parameters(),
+                                    text_encoder_one.parameters(),
+                                    text_encoder_two.parameters(),
+                                )
+                            else:
+                                params_to_clip = itertools.chain(
+                                    transformer.parameters(), text_encoder_one.parameters()
+                                )
                    else:
                        params_to_clip = itertools.chain(transformer.parameters())
                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
@@ -2339,6 +2283,10 @@ def main(args):

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

            logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
@@ -2351,14 +2299,16 @@ def main(args):
        if accelerator.is_main_process:
            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                # create pipeline
-                if freeze_text_encoder:
+                if freeze_text_encoder:  # no text encoder one, two optimizations
                    text_encoder_one, text_encoder_two = load_text_encoders(text_encoder_cls_one, text_encoder_cls_two)
+                    text_encoder_one.to(weight_dtype)
+                    text_encoder_two.to(weight_dtype)
                pipeline = FluxPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
-                    transformer=accelerator.unwrap_model(transformer),
+                    text_encoder=unwrap_model(text_encoder_one),
+                    text_encoder_2=unwrap_model(text_encoder_two),
+                    transformer=unwrap_model(transformer),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
@@ -2372,21 +2322,21 @@ def main(args):
                    epoch=epoch,
                    torch_dtype=weight_dtype,
                )
-                images = None
-                del pipeline
-
                if freeze_text_encoder:
                    del text_encoder_one, text_encoder_two
                    free_memory()
-                elif args.train_text_encoder:
-                    del text_encoder_two
-                    free_memory()
+
+                images = None
+                del pipeline

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

        if args.train_text_encoder:
@@ -2428,8 +2378,8 @@ def main(args):
                accelerator=accelerator,
                pipeline_args=pipeline_args,
                epoch=epoch,
-                torch_dtype=weight_dtype,
                is_final_validation=True,
+                torch_dtype=weight_dtype,
            )

        save_model_card(
@@ -2452,6 +2402,7 @@ def main(args):
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )
+
        images = None
        del pipeline

@@ -200,7 +200,8 @@ Special VAE used for training: {vae_path}.
        "diffusers",
        "diffusers-training",
        lora,
-        "template:sd-lora" "stable-diffusion",
+        "template:sd-lora",
+        "stable-diffusion",
        "stable-diffusion-diffusers",
    ]
    model_card = populate_model_card(model_card, tags=tags)
@@ -662,7 +663,7 @@ def parse_args(input_args=None):
        action="store_true",
        default=False,
        help=(
-            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
+            "Whether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
            "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
        ),
    )
@@ -724,9 +725,9 @@ class TokenEmbeddingsHandler:
        idx = 0
        for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
            assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )

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

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

        lora_state_dict, network_alphas = StableDiffusionPipeline.lora_state_dict(input_dir)

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

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

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

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

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


 def log_validation(
@@ -773,7 +772,7 @@ def parse_args(input_args=None):
        action="store_true",
        default=False,
        help=(
-            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
+            "Whether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
            "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
        ),
    )
@@ -891,9 +890,9 @@ class TokenEmbeddingsHandler:
        idx = 0
        for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
            assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )

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

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

        lora_state_dict, network_alphas = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
        if incompatible_keys is not None:
@@ -1875,7 +1874,7 @@ def main(args):
    # pack the statically computed variables appropriately here. This is so that we don't
    # have to pass them to the dataloader.

-    # if --train_text_encoder_ti we need add_special_tokens to be True fo textual inversion
+    # if --train_text_encoder_ti we need add_special_tokens to be True for textual inversion
    add_special_tokens = True if args.train_text_encoder_ti else False

    if not train_dataset.custom_instance_prompts:
@@ -720,7 +720,7 @@ def main(args):
    # Train!
    logger.info("***** Running training *****")
    logger.info(f"  Num training steps = {args.max_train_steps}")
-    logger.info(f"  Instantaneous batch size per device = { args.train_batch_size}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")

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

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1159,7 +1159,7 @@ def main(args):
        lora_state_dict = CogVideoXPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -0,0 +1,201 @@
+# Training CogView4 Control
+
+This (experimental) example shows how to train Control LoRAs with [CogView4](https://huggingface.co/THUDM/CogView4-6B) by conditioning it with additional structural controls (like depth maps, poses, etc.). We provide a script for full fine-tuning, too, refer to [this section](#full-fine-tuning). To know more about CogView4 Control family, refer to the following resources:
+
+To incorporate additional condition latents, we expand the input features of CogView-4 from 64 to 128. The first 64 channels correspond to the original input latents to be denoised, while the latter 64 channels correspond to control latents. This expansion happens on the `patch_embed` layer, where the combined latents are projected to the expected feature dimension of rest of the network. Inference is performed using the `CogView4ControlPipeline`.
+
+> [!NOTE]
+> **Gated model**
+>
+> As the model is gated, before using it with diffusers you first need to go to the [CogView4 Hugging Face page](https://huggingface.co/THUDM/CogView4-6B), 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
+```
+
+The example command below shows how to launch fine-tuning for pose conditions. The dataset ([`raulc0399/open_pose_controlnet`](https://huggingface.co/datasets/raulc0399/open_pose_controlnet)) being used here already has the pose conditions of the original images, so we don't have to compute them.
+
+```bash
+accelerate launch train_control_lora_cogview4.py \
+  --pretrained_model_name_or_path="THUDM/CogView4-6B" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control-lora" \
+  --mixed_precision="bf16" \
+  --train_batch_size=1 \
+  --rank=64 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=5000 \
+  --validation_image="openpose.png" \
+  --validation_prompt="A couple, 4k photo, highly detailed" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+`openpose.png` comes from [here](https://huggingface.co/Adapter/t2iadapter/resolve/main/openpose.png).
+
+You need to install `diffusers` from the branch of [this PR](https://github.com/huggingface/diffusers/pull/9999). When it's merged, you should install `diffusers` from the `main`.
+
+The training script exposes additional CLI args that might be useful to experiment with:
+
+* `use_lora_bias`: When set, additionally trains the biases of the `lora_B` layer. 
+* `train_norm_layers`: When set, additionally trains the normalization scales. Takes care of saving and loading.
+* `lora_layers`: Specify the layers you want to apply LoRA to. If you specify "all-linear", all the linear layers will be LoRA-attached.
+
+### Training with DeepSpeed
+
+It's possible to train with [DeepSpeed](https://github.com/microsoft/DeepSpeed), specifically leveraging the Zero2 system optimization. To use it, save the following config to an YAML file (feel free to modify as needed):
+
+```yaml
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  gradient_clipping: 1.0
+  offload_optimizer_device: cpu
+  offload_param_device: cpu
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+And then while launching training, pass the config file:
+
+```bash
+accelerate launch --config_file=CONFIG_FILE.yaml ...
+```
+
+### Inference
+
+The pose images in our dataset were computed using the [`controlnet_aux`](https://github.com/huggingface/controlnet_aux) library. Let's install it first:
+
+```bash
+pip install controlnet_aux
+```
+
+And then we are ready:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import CogView4ControlPipeline
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+pipe = CogView4ControlPipeline.from_pretrained("THUDM/CogView4-6B", torch_dtype=torch.bfloat16).to("cuda")
+pipe.load_lora_weights("...") # change this.
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  control_image=image,
+  num_inference_steps=50,
+  joint_attention_kwargs={"scale": 0.9},
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Full fine-tuning
+
+We provide a non-LoRA version of the training script `train_control_cogview4.py`. Here is an example command:
+
+```bash
+accelerate launch --config_file=accelerate_ds2.yaml train_control_cogview4.py \
+  --pretrained_model_name_or_path="THUDM/CogView4-6B" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control" \
+  --mixed_precision="bf16" \
+  --train_batch_size=2 \
+  --dataloader_num_workers=4 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --proportion_empty_prompts=0.2 \
+  --learning_rate=5e-5 \
+  --adam_weight_decay=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="cosine" \
+  --lr_warmup_steps=1000 \
+  --checkpointing_steps=1000 \
+  --max_train_steps=10000 \
+  --validation_steps=200 \
+  --validation_image "2_pose_1024.jpg" "3_pose_1024.jpg" \
+  --validation_prompt "two friends sitting by each other enjoying a day at the park, full hd, cinematic" "person enjoying a day at the park, full hd, cinematic" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+Change the `validation_image` and `validation_prompt` as needed.
+
+For inference, this time, we will run:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import CogView4ControlPipeline, CogView4Transformer2DModel
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+transformer = CogView4Transformer2DModel.from_pretrained("...") # change this.
+pipe = CogView4ControlPipeline.from_pretrained(
+  "THUDM/CogView4-6B",  transformer=transformer, torch_dtype=torch.bfloat16
+).to("cuda")
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  control_image=image,
+  num_inference_steps=50,
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Things to note
+
+* The scripts provided in this directory are experimental and educational. This means we may have to tweak things around to get good results on a given condition. We believe this is best done with the community 🤗
+* The scripts are not memory-optimized but we offload the VAE and the text encoders to CPU when they are not used if `--offload` is specified. 
+* We can extract LoRAs from the fully fine-tuned model. While we currently don't provide any utilities for that, users are welcome to refer to [this script](https://github.com/Stability-AI/stability-ComfyUI-nodes/blob/master/control_lora_create.py) that provides a similar functionality. 
@@ -0,0 +1,6 @@
+transformers==4.47.0
+wandb
+torch
+torchvision
+accelerate==1.2.0
+peft>=0.14.0
@@ -10,7 +10,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif

 | Example                                                                                                                               | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                                              | Colab                                                                                                                                                                                                              |                                                        Author |
 |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
-|Spatiotemporal Skip Guidance (STG)|[Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664) (CVPR 2025) enhances video diffusion models by generating a weaker model through layer skipping and using it as guidance, improving fidelity in models like HunyuanVideo, LTXVideo, and Mochi.|[Spatiotemporal Skip Guidance](#spatiotemporal-skip-guidance)|-|[Junha Hyung](https://junhahyung.github.io/), [Kinam Kim](https://kinam0252.github.io/)|
+|Spatiotemporal Skip Guidance (STG)|[Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664) (CVPR 2025) enhances video diffusion models by generating a weaker model through layer skipping and using it as guidance, improving fidelity in models like HunyuanVideo, LTXVideo, and Mochi.|[Spatiotemporal Skip Guidance](#spatiotemporal-skip-guidance)|-|[Junha Hyung](https://junhahyung.github.io/), [Kinam Kim](https://kinam0252.github.io/), and [Ednaordinary](https://github.com/Ednaordinary)|
 |Adaptive Mask Inpainting|Adaptive Mask Inpainting algorithm from [Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models](https://github.com/snuvclab/coma) (ECCV '24, Oral) provides a way to insert human inside the scene image without altering the background, by inpainting with adapting mask.|[Adaptive Mask Inpainting](#adaptive-mask-inpainting)|-|[Hyeonwoo Kim](https://sshowbiz.xyz),[Sookwan Han](https://jellyheadandrew.github.io)|
 |Flux with CFG|[Flux with CFG](https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md) provides an implementation of using CFG in [Flux](https://blackforestlabs.ai/announcing-black-forest-labs/).|[Flux with CFG](#flux-with-cfg)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/flux_with_cfg.ipynb)|[Linoy Tsaban](https://github.com/linoytsaban), [Apolinário](https://github.com/apolinario), and [Sayak Paul](https://github.com/sayakpaul)|
 |Differential Diffusion|[Differential Diffusion](https://github.com/exx8/differential-diffusion) modifies an image according to a text prompt, and according to a map that specifies the amount of change in each region.|[Differential Diffusion](#differential-diffusion)|[![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/exx8/differential-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/exx8/differential-diffusion/blob/main/examples/SD2.ipynb)|[Eran Levin](https://github.com/exx8) and [Ohad Fried](https://www.ohadf.com/)|
@@ -24,12 +24,12 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | Long Prompt Weighting Stable Diffusion                                                                                                | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt.                                                                                                                                                                                                                                                                                                                                                                                                  | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion)         | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/long_prompt_weighting_stable_diffusion.ipynb)                                                                                        |                           [SkyTNT](https://github.com/SkyTNT) |
 | Speech to Image                                                                                                                       | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images                                                                                                                                                                                                                                                                                                                                                                                                            | [Speech to Image](#speech-to-image)                                                       |[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/speech_to_image.ipynb)                                                                                                                                                                                                   |             [Mikail Duzenli](https://github.com/MikailINTech)
 | Wild Card Stable Diffusion                                                                                                            | Stable Diffusion Pipeline that supports prompts that contain wildcard terms (indicated by surrounding double underscores), with values instantiated randomly from a corresponding txt file or a dictionary of possible values                                                                                                                                                                                                                                                                            | [Wildcard Stable Diffusion](#wildcard-stable-diffusion)                                   | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/wildcard_stable_diffusion.ipynb)                                                                                                                                                                                 |              [Shyam Sudhakaran](https://github.com/shyamsn97) |
-| [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "&#124;" in prompts (as an AND condition) and weights (separated by "&#124;" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                            | [Composable Stable Diffusion](#composable-stable-diffusion)                               | -                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
+| [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "&#124;" in prompts (as an AND condition) and weights (separated by "&#124;" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                            | [Composable Stable Diffusion](#composable-stable-diffusion)                               | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/composable_stable_diffusion.ipynb)                                                                                                                                                                                                                 |                      [Mark Rich](https://github.com/MarkRich) |
 | Seed Resizing Stable Diffusion                                                                                                        | Stable Diffusion Pipeline that supports resizing an image and retaining the concepts of the 512 by 512 generation.                                                                                                                                                                                                                                                                                                                                                                                       | [Seed Resizing](#seed-resizing)                                                           | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/seed_resizing.ipynb)                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
 | Imagic Stable Diffusion                                                                                                               | Stable Diffusion Pipeline that enables writing a text prompt to edit an existing image                                                                                                                                                                                                                                                                                                                                                                                                                   | [Imagic Stable Diffusion](#imagic-stable-diffusion)                                       | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/imagic_stable_diffusion.ipynb)                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
 | Multilingual Stable Diffusion                                                                                                         | Stable Diffusion Pipeline that supports prompts in 50 different languages.                                                                                                                                                                                                                                                                                                                                                                                                                               | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline)                  | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/multilingual_stable_diffusion.ipynb)                                                                                                                                                                             |          [Juan Carlos Piñeros](https://github.com/juancopi81) |
 | GlueGen Stable Diffusion                                                                                                         | Stable Diffusion Pipeline that supports prompts in different languages using GlueGen adapter.                                                                                                                                                                                                                                                                                                                                                                                                                               | [GlueGen Stable Diffusion](#gluegen-stable-diffusion-pipeline)                  | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/gluegen_stable_diffusion.ipynb)                                                                                                                                                                                                                  |          [Phạm Hồng Vinh](https://github.com/rootonchair) |
-| Image to Image Inpainting Stable Diffusion                                                                                            | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting                                                                                                                                                                                                                                                                                                                                                                                                            | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | -                                                                                                                                                                                                                  |                    [Alex McKinney](https://github.com/vvvm23) |
+| Image to Image Inpainting Stable Diffusion                                                                                            | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting                                                                                                                                                                                                                                                                                                                                                                                                            | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/image_to_image_inpainting_stable_diffusion.ipynb)                                                                                                                                                                                                                  |                    [Alex McKinney](https://github.com/vvvm23) |
 | Text Based Inpainting Stable Diffusion                                                                                                | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting                                                                                                                                                                                                                                                                                                                                                                                              | [Text Based Inpainting Stable Diffusion](#text-based-inpainting-stable-diffusion)     | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/text_based_inpainting_stable_dffusion.ipynb)                                                                                                                                                                                                    |                   [Dhruv Karan](https://github.com/unography) |
 | Bit Diffusion                                                                                                                         | Diffusion on discrete data                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | [Bit Diffusion](#bit-diffusion)                                                           | -  |                       [Stuti R.](https://github.com/kingstut) |
 | K-Diffusion Stable Diffusion                                                                                                          | Run Stable Diffusion with any of [K-Diffusion's samplers](https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py)                                                                                                                                                                                                                                                                                                                                                                  | [Stable Diffusion with K Diffusion](#stable-diffusion-with-k-diffusion)                   | -  |    [Patrick von Platen](https://github.com/patrickvonplaten/) |
@@ -41,7 +41,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | UnCLIP Image Interpolation Pipeline                                                                                                   | Diffusion Pipeline that allows passing two images/image_embeddings and produces images while interpolating between their image-embeddings                                                                                                                                                                                                                                                                                                                                                                | [UnCLIP Image Interpolation Pipeline](#unclip-image-interpolation-pipeline)               | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/unclip_image_interpolation.ipynb)| [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
 | DDIM Noise Comparative Analysis Pipeline                                                                                              | Investigating how the diffusion models learn visual concepts from each noise level (which is a contribution of [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227))                                                                                                                                                                                                                                                                                                                             | [DDIM Noise Comparative Analysis Pipeline](#ddim-noise-comparative-analysis-pipeline)     | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/ddim_noise_comparative_analysis.ipynb)|              [Aengus (Duc-Anh)](https://github.com/aengusng8) |
 | CLIP Guided Img2Img Stable Diffusion Pipeline                                                                                         | Doing CLIP guidance for image to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                  | [CLIP Guided Img2Img Stable Diffusion](#clip-guided-img2img-stable-diffusion)             | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/clip_guided_img2img_stable_diffusion.ipynb) |               [Nipun Jindal](https://github.com/nipunjindal/) |
-| TensorRT Stable Diffusion Text to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
+| TensorRT Stable Diffusion Text to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline)      | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/tensorrt_text2image_stable_diffusion_pipeline.ipynb) |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
 | EDICT Image Editing Pipeline                                                                                                          | Diffusion pipeline for text-guided image editing                                                                                                                                                                                                                                                                                                                                                                                                                                                         | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline)                             | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/edict_image_pipeline.ipynb) |                    [Joqsan Azocar](https://github.com/Joqsan) |
 | Stable Diffusion RePaint                                                                                                              | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.09865) for inpainting.                                                                                                                                                                                                                                                                                                                                                                                                               | [Stable Diffusion RePaint](#stable-diffusion-repaint )|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_repaint.ipynb)|                  [Markus Pobitzer](https://github.com/Markus-Pobitzer) |
 | TensorRT Stable Diffusion Image to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Image2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Image to Image Pipeline](#tensorrt-image2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
@@ -58,7 +58,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | FABRIC - Stable Diffusion with feedback Pipeline | pipeline supports feedback from liked and disliked images | [Stable Diffusion Fabric Pipeline](#stable-diffusion-fabric-pipeline) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_fabric.ipynb)| [Shauray Singh](https://shauray8.github.io/about_shauray/) |
 | sketch inpaint - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion Pipeline](#stable-diffusion-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
 | sketch inpaint xl - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion XL Pipeline](#stable-diffusion-xl-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
-| prompt-to-prompt | change parts of a prompt and retain image structure (see [paper page](https://prompt-to-prompt.github.io/)) | [Prompt2Prompt Pipeline](#prompt2prompt-pipeline) | - | [Umer H. Adil](https://twitter.com/UmerHAdil) |
+| prompt-to-prompt | change parts of a prompt and retain image structure (see [paper page](https://prompt-to-prompt.github.io/)) | [Prompt2Prompt Pipeline](#prompt2prompt-pipeline) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/prompt_2_prompt_pipeline.ipynb) | [Umer H. Adil](https://twitter.com/UmerHAdil) |
 |   Latent Consistency Pipeline                                                                                                    | Implementation of [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://arxiv.org/abs/2310.04378)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Latent Consistency Pipeline](#latent-consistency-pipeline)      | - |              [Simian Luo](https://github.com/luosiallen) |
 |   Latent Consistency Img2img Pipeline                                                                                                    | Img2img pipeline for Latent Consistency Models                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Img2Img Pipeline](#latent-consistency-img2img-pipeline)      | - |              [Logan Zoellner](https://github.com/nagolinc) |
 |   Latent Consistency Interpolation Pipeline                                                                                                    | Interpolate the latent space of Latent Consistency Models with multiple prompts                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Interpolation Pipeline](#latent-consistency-interpolation-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1pK3NrLWJSiJsBynLns1K1-IDTW9zbPvl?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
@@ -85,7 +85,7 @@ PIXART-α Controlnet pipeline | Implementation of the controlnet model for pixar
 | Stable Diffusion XL Attentive Eraser Pipeline |[[AAAI2025 Oral] Attentive Eraser](https://github.com/Anonym0u3/AttentiveEraser) is a novel tuning-free method that enhances object removal capabilities in pre-trained diffusion models.|[Stable Diffusion XL Attentive Eraser Pipeline](#stable-diffusion-xl-attentive-eraser-pipeline)|-|[Wenhao Sun](https://github.com/Anonym0u3) and [Benlei Cui](https://github.com/Benny079)|
 | Perturbed-Attention Guidance |StableDiffusionPAGPipeline is a modification of StableDiffusionPipeline to support Perturbed-Attention Guidance (PAG).|[Perturbed-Attention Guidance](#perturbed-attention-guidance)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/perturbed_attention_guidance.ipynb)|[Hyoungwon Cho](https://github.com/HyoungwonCho)|
 | CogVideoX DDIM Inversion Pipeline | Implementation of DDIM inversion and guided attention-based editing denoising process on CogVideoX. | [CogVideoX DDIM Inversion Pipeline](#cogvideox-ddim-inversion-pipeline) | - | [LittleNyima](https://github.com/LittleNyima) |
-
+| FaithDiff Stable Diffusion XL Pipeline | Implementation of [(CVPR 2025) FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolutionUnleashing Diffusion Priors for Faithful Image Super-resolution](https://arxiv.org/abs/2411.18824) - FaithDiff is a faithful image super-resolution method that leverages latent diffusion models by actively adapting the diffusion prior and jointly fine-tuning its components (encoder and diffusion model) with an alignment module to ensure high fidelity and structural consistency. | [FaithDiff Stable Diffusion XL Pipeline](#faithdiff-stable-diffusion-xl-pipeline) | [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/jychen9811/FaithDiff) | [Junyang Chen, Jinshan Pan, Jiangxin Dong, IMAG Lab, (Adapted by Eliseu Silva)](https://github.com/JyChen9811/FaithDiff) |
 To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.

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

@@ -954,6 +953,7 @@ for i in range(args.num_images):
    images.append(th.from_numpy(np.array(image)).permute(2, 0, 1) / 255.)
 grid = tvu.make_grid(th.stack(images, dim=0), nrow=4, padding=0)
 tvu.save_image(grid, f'{prompt}_{args.weights}' + '.png')
+print("Image saved successfully!")
 ```

 ### Imagic Stable Diffusion
@@ -1269,28 +1269,39 @@ The aim is to overlay two images, then mask out the boundary between `image` and
 For example, this could be used to place a logo on a shirt and make it blend seamlessly.

 ```python
-import PIL
 import torch
-
+import requests
+from PIL import Image
+from io import BytesIO
 from diffusers import DiffusionPipeline

-image_path = "./path-to-image.png"
-inner_image_path = "./path-to-inner-image.png"
-mask_path = "./path-to-mask.png"
+image_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+inner_image_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"

-init_image = PIL.Image.open(image_path).convert("RGB").resize((512, 512))
-inner_image = PIL.Image.open(inner_image_path).convert("RGBA").resize((512, 512))
-mask_image = PIL.Image.open(mask_path).convert("RGB").resize((512, 512))
+def load_image(url, mode="RGB"):
+    response = requests.get(url)
+    if response.status_code == 200:
+        return Image.open(BytesIO(response.content)).convert(mode).resize((512, 512))
+    else:
+        raise FileNotFoundError(f"Could not retrieve image from {url}")
+
+
+init_image = load_image(image_url, mode="RGB")
+inner_image = load_image(inner_image_url, mode="RGBA")
+mask_image = load_image(mask_url, mode="RGB")

 pipe = DiffusionPipeline.from_pretrained(
-    "runwayml/stable-diffusion-inpainting",
+    "stable-diffusion-v1-5/stable-diffusion-inpainting",
    custom_pipeline="img2img_inpainting",
    torch_dtype=torch.float16
 )
 pipe = pipe.to("cuda")

-prompt = "Your prompt here!"
+prompt = "a mecha robot sitting on a bench"
 image = pipe(prompt=prompt, image=init_image, inner_image=inner_image, mask_image=mask_image).images[0]
+
+image.save("output.png")
 ```

 ![2 by 2 grid demonstrating image to image inpainting.](https://user-images.githubusercontent.com/44398246/203506577-ec303be4-887e-4ebd-a773-c83fcb3dd01a.png)
@@ -3252,14 +3263,19 @@ Here's a full example for `ReplaceEdit``:

 ```python
 import torch
-import numpy as np
-import matplotlib.pyplot as plt
 from diffusers import DiffusionPipeline
+import numpy as np
+from PIL import Image

-pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="pipeline_prompt2prompt").to("cuda")
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4", 
+    custom_pipeline="pipeline_prompt2prompt"
+).to("cuda")

-prompts = ["A turtle playing with a ball",
-           "A monkey playing with a ball"]
+prompts = [
+    "A turtle playing with a ball",
+    "A monkey playing with a ball"
+]

 cross_attention_kwargs = {
    "edit_type": "replace",
@@ -3267,7 +3283,15 @@ cross_attention_kwargs = {
    "self_replace_steps": 0.4
 }

-outputs = pipe(prompt=prompts, height=512, width=512, num_inference_steps=50, cross_attention_kwargs=cross_attention_kwargs)
+outputs = pipe(
+    prompt=prompts,
+    height=512,
+    width=512,
+    num_inference_steps=50,
+    cross_attention_kwargs=cross_attention_kwargs
+)
+
+outputs.images[0].save("output_image_0.png")
 ```

 And abbreviated examples for the other edits:
@@ -5309,3 +5333,103 @@ output = pipeline_for_inversion(
 pipeline.export_latents_to_video(output.inverse_latents[-1], "path/to/inverse_video.mp4", fps=8)
 pipeline.export_latents_to_video(output.recon_latents[-1], "path/to/recon_video.mp4", fps=8)
 ```
+# FaithDiff Stable Diffusion XL Pipeline
+
+[Project](https://jychen9811.github.io/FaithDiff_page/) / [GitHub](https://github.com/JyChen9811/FaithDiff/)
+
+This the implementation of the FaithDiff pipeline for SDXL, adapted to use the HuggingFace Diffusers.
+
+For more details see the project links above.
+
+## Example Usage
+
+This example upscale and restores a low-quality image. The input image has a resolution of 512x512 and will be upscaled at a scale of 2x, to a final resolution of 1024x1024. It is possible to upscale to a larger scale, but it is recommended that the input image be at least 1024x1024 in these cases. To upscale this image by 4x, for example, it would be recommended to re-input the result into a new 2x processing, thus performing progressive scaling.
+
+````py
+import random
+import numpy as np
+import torch
+from diffusers import DiffusionPipeline, AutoencoderKL, UniPCMultistepScheduler
+from huggingface_hub import hf_hub_download
+from diffusers.utils import load_image
+from PIL import Image
+
+device = "cuda"
+dtype = torch.float16
+MAX_SEED = np.iinfo(np.int32).max
+
+# Download weights for additional unet layers
+model_file = hf_hub_download(
+    "jychen9811/FaithDiff",
+    filename="FaithDiff.bin", local_dir="./proc_data/faithdiff", local_dir_use_symlinks=False
+)
+
+# Initialize the models and pipeline
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=dtype)
+
+model_id = "SG161222/RealVisXL_V4.0"
+pipe = DiffusionPipeline.from_pretrained(
+    model_id,
+    torch_dtype=dtype,
+    vae=vae,
+    unet=None, #<- Do not load with original model.
+    custom_pipeline="pipeline_faithdiff_stable_diffusion_xl",    
+    use_safetensors=True,
+    variant="fp16",
+).to(device)
+
+# Here we need use pipeline internal unet model
+pipe.unet = pipe.unet_model.from_pretrained(model_id, subfolder="unet", variant="fp16", use_safetensors=True)
+
+# Load aditional layers to the model
+pipe.unet.load_additional_layers(weight_path="proc_data/faithdiff/FaithDiff.bin", dtype=dtype)
+
+# Enable vae tiling
+pipe.set_encoder_tile_settings()
+pipe.enable_vae_tiling()
+
+# Optimization
+pipe.enable_model_cpu_offload()
+
+# Set selected scheduler
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+
+#input params
+prompt = "The image features a woman in her 55s with blonde hair and a white shirt, smiling at the camera. She appears to be in a good mood and is wearing a white scarf around her neck. "
+upscale = 2 # scale here
+start_point = "lr" # or "noise"
+latent_tiled_overlap = 0.5
+latent_tiled_size = 1024
+
+# Load image
+lq_image = load_image("https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/woman.png")
+original_height = lq_image.height
+original_width = lq_image.width
+print(f"Current resolution: H:{original_height} x W:{original_width}")
+
+width = original_width * int(upscale)
+height = original_height * int(upscale)
+print(f"Final resolution: H:{height} x W:{width}")
+
+# Restoration
+image = lq_image.resize((width, height), Image.LANCZOS)
+input_image, width_init, height_init, width_now, height_now = pipe.check_image_size(image)
+
+generator = torch.Generator(device=device).manual_seed(random.randint(0, MAX_SEED))
+gen_image = pipe(lr_img=input_image, 
+                 prompt = prompt,                  
+                 num_inference_steps=20, 
+                 guidance_scale=5, 
+                 generator=generator, 
+                 start_point=start_point, 
+                 height = height_now, 
+                 width=width_now, 
+                 overlap=latent_tiled_overlap, 
+                 target_size=(latent_tiled_size, latent_tiled_size)
+                ).images[0]
+
+cropped_image = gen_image.crop((0, 0, width_init, height_init))
+cropped_image.save("data/result.png")
+````
+### Result
+[<img src="https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/faithdiff_restored.PNG" width="512px" height="512px"/>](https://imgsli.com/MzY1NzE2)
@@ -1103,7 +1103,7 @@ class AdaptiveMaskInpaintPipeline(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `default_mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -686,7 +686,7 @@ class StableDiffusionHDPainterPipeline(StableDiffusionInpaintPipeline):
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -362,7 +362,7 @@ class ImageToImageInpaintingPipeline(DiffusionPipeline):
                f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                " `pipeline.unet` or your `mask_image` or `image` input."
            )

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

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

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

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

            finally:
@@ -1773,7 +1773,7 @@ class SDXLLongPromptWeightingPipeline(
                        f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                        f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                        f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                        f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                        f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                        " `pipeline.unet` or your `mask_image` or `image` input."
                    )
            elif num_channels_unet != 4:
@@ -1924,7 +1924,22 @@ class SDXLLongPromptWeightingPipeline(
                self.upcast_vae()
                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)

-            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]

            # cast back to fp16 if needed
            if needs_upcasting:
@@ -1,4 +1,4 @@
-# Copyright 2025 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The DEVAIEXP Team and The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -1070,32 +1070,32 @@ class StableDiffusionXLTilingPipeline(
                    text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
                else:
                    text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
-                    add_time_ids = self._get_add_time_ids(
-                        original_size,
-                        crops_coords_top_left[row][col],
-                        target_size,
+                add_time_ids = self._get_add_time_ids(
+                    original_size,
+                    crops_coords_top_left[row][col],
+                    target_size,
+                    dtype=prompt_embeds.dtype,
+                    text_encoder_projection_dim=text_encoder_projection_dim,
+                )
+                if negative_original_size is not None and negative_target_size is not None:
+                    negative_add_time_ids = self._get_add_time_ids(
+                        negative_original_size,
+                        negative_crops_coords_top_left[row][col],
+                        negative_target_size,
                        dtype=prompt_embeds.dtype,
                        text_encoder_projection_dim=text_encoder_projection_dim,
                    )
-                    if negative_original_size is not None and negative_target_size is not None:
-                        negative_add_time_ids = self._get_add_time_ids(
-                            negative_original_size,
-                            negative_crops_coords_top_left[row][col],
-                            negative_target_size,
-                            dtype=prompt_embeds.dtype,
-                            text_encoder_projection_dim=text_encoder_projection_dim,
-                        )
-                    else:
-                        negative_add_time_ids = add_time_ids
+                else:
+                    negative_add_time_ids = add_time_ids

-                    if self.do_classifier_free_guidance:
-                        prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
-                        add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
-                        add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+                if self.do_classifier_free_guidance:
+                    prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+                    add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+                    add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)

-                    prompt_embeds = prompt_embeds.to(device)
-                    add_text_embeds = add_text_embeds.to(device)
-                    add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+                prompt_embeds = prompt_embeds.to(device)
+                add_text_embeds = add_text_embeds.to(device)
+                add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
                addition_embed_type_row.append((prompt_embeds, add_text_embeds, add_time_ids))
            embeddings_and_added_time.append(addition_embed_type_row)

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

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

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

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

        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
            raise ValueError(
@@ -2050,7 +2050,7 @@ class StableDiffusionXL_AE_Pipeline(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -1578,7 +1578,7 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
@@ -0,0 +1,661 @@
+# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+import types
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import ftfy
+import regex as re
+import torch
+from transformers import AutoTokenizer, UMT5EncoderModel
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import WanLoraLoaderMixin
+from diffusers.models import AutoencoderKLWan, WanTransformer3DModel
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.wan.pipeline_output import WanPipelineOutput
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from diffusers import AutoencoderKLWan
+        >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
+        >>> from examples.community.pipeline_stg_wan import WanSTGPipeline
+
+        >>> # Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+        >>> model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
+        >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+        >>> pipe = WanSTGPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+        >>> flow_shift = 5.0  # 5.0 for 720P, 3.0 for 480P
+        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+        >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [8] # Layer indices from 0 to 39 for 14b or 0 to 29 for 1.3b
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+
+        >>> output = pipe(
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     height=720,
+        ...     width=1280,
+        ...     num_frames=81,
+        ...     guidance_scale=5.0,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ... ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=16)
+        ```
+"""
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def prompt_clean(text):
+    text = whitespace_clean(basic_clean(text))
+    return text
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    rotary_emb: torch.Tensor,
+) -> torch.Tensor:
+    return hidden_states
+
+
+def forward_without_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    rotary_emb: torch.Tensor,
+) -> torch.Tensor:
+    shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+        self.scale_shift_table + temb.float()
+    ).chunk(6, dim=1)
+
+    # 1. Self-attention
+    norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
+    attn_output = self.attn1(hidden_states=norm_hidden_states, rotary_emb=rotary_emb)
+    hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
+
+    # 2. Cross-attention
+    norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
+    attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+    hidden_states = hidden_states + attn_output
+
+    # 3. Feed-forward
+    norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(hidden_states)
+    ff_output = self.ffn(norm_hidden_states)
+    hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
+
+    return hidden_states
+
+
+class WanSTGPipeline(DiffusionPipeline, WanLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation using Wan.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        tokenizer ([`T5Tokenizer`]):
+            Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer),
+            specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        transformer ([`WanTransformer3DModel`]):
+            Conditional Transformer to denoise the input latents.
+        scheduler ([`UniPCMultistepScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: UMT5EncoderModel,
+        transformer: WanTransformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt = [prompt_clean(u) for u in prompt]
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
+        seq_lens = mask.gt(0).sum(dim=1).long()
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
+        )
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        negative_prompt,
+        height,
+        width,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif negative_prompt is not None and (
+            not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list)
+        ):
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: int = 16,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_latent_frames,
+            int(height) // self.vae_scale_factor_spatial,
+            int(width) // self.vae_scale_factor_spatial,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "np",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+        stg_applied_layers_idx: Optional[List[int]] = [3, 8, 16],
+        stg_scale: Optional[float] = 0.0,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `480`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `832`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `81`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `5.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
+                The dtype to use for the torch.amp.autocast.
+
+        Examples:
+
+        Returns:
+            [`~WanPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where
+                the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._stg_scale = stg_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                latent_model_input = latents.to(transformer_dtype)
+                timestep = t.expand(latents.shape[0])
+
+                if self.do_spatio_temporal_guidance:
+                    for idx, block in enumerate(self.transformer.blocks):
+                        block.forward = types.MethodType(forward_without_stg, block)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    if self.do_spatio_temporal_guidance:
+                        for idx, block in enumerate(self.transformer.blocks):
+                            if idx in stg_applied_layers_idx:
+                                block.forward = types.MethodType(forward_with_stg, block)
+                        noise_perturb = self.transformer(
+                            hidden_states=latent_model_input,
+                            timestep=timestep,
+                            encoder_hidden_states=prompt_embeds,
+                            attention_kwargs=attention_kwargs,
+                            return_dict=False,
+                        )[0]
+                        noise_pred = (
+                            noise_uncond
+                            + guidance_scale * (noise_pred - noise_uncond)
+                            + self._stg_scale * (noise_pred - noise_perturb)
+                        )
+                    else:
+                        noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return WanPipelineOutput(frames=video)
@@ -288,8 +288,7 @@ class UFOGenScheduler(SchedulerMixin, ConfigMixin):

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

            timesteps = np.array(timesteps, dtype=np.int64)
@@ -89,7 +89,7 @@ def get_module_kohya_state_dict(module, prefix: str, dtype: torch.dtype, adapter

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

    return kohya_ss_state_dict
@@ -901,7 +901,7 @@ def main(args):
            unet_ = accelerator.unwrap_model(unet)
            lora_state_dict, _ = StableDiffusionXLPipeline.lora_state_dict(input_dir)
            unet_state_dict = {
-                f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")
+                f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")
            }
            unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
            incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
@@ -95,7 +95,7 @@ def get_module_kohya_state_dict(module, prefix: str, dtype: torch.dtype, adapter

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

    return kohya_ss_state_dict
@@ -152,9 +152,7 @@ def log_validation(
                validation_prompt = log["validation_prompt"]
                validation_image = log["validation_image"]

-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]

                for image in images:
                    formatted_images.append(np.asarray(image))
@@ -929,17 +927,22 @@ def main(args):
    )

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

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )
@@ -964,8 +967,14 @@ def main(args):

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

@@ -166,9 +166,7 @@ def log_validation(
                validation_prompt = log["validation_prompt"]
                validation_image = log["validation_image"]

-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]

                for image in images:
                    formatted_images.append(np.asarray(image))
@@ -17,6 +17,7 @@ import argparse
 import contextlib
 import copy
 import functools
+import gc
 import logging
 import math
 import os
@@ -52,6 +53,7 @@ from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_density_for_timestep_sampling, compute_loss_weighting_for_sd3, free_memory
 from diffusers.utils import check_min_version, is_wandb_available, make_image_grid
 from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
+from diffusers.utils.testing_utils import backend_empty_cache
 from diffusers.utils.torch_utils import is_compiled_module


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

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

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

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

        images = []

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

            images.append(image)
@@ -655,6 +695,7 @@ def make_train_dataset(args, tokenizer_one, tokenizer_two, tokenizer_three, acce
            dataset = load_dataset(
                args.train_data_dir,
                cache_dir=args.cache_dir,
+                trust_remote_code=True,
            )
        # See more about loading custom images at
        # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
@@ -1283,8 +1324,8 @@ def main(args):
                noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise

                # Get the text embedding for conditioning
-                prompt_embeds = batch["prompt_embeds"]
-                pooled_prompt_embeds = batch["pooled_prompt_embeds"]
+                prompt_embeds = batch["prompt_embeds"].to(dtype=weight_dtype)
+                pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(dtype=weight_dtype)

                # controlnet(s) inference
                controlnet_image = batch["conditioning_pixel_values"].to(dtype=weight_dtype)
@@ -157,9 +157,7 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
                validation_prompt = log["validation_prompt"]
                validation_image = log["validation_image"]

-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]

                for image in images:
                    formatted_images.append(np.asarray(image))
@@ -50,9 +50,11 @@ def retrieve(class_prompt, class_data_dir, num_class_images):
    total = 0
    pbar = tqdm(desc="downloading real regularization images", total=num_class_images)

-    with open(f"{class_data_dir}/caption.txt", "w") as f1, open(f"{class_data_dir}/urls.txt", "w") as f2, open(
-        f"{class_data_dir}/images.txt", "w"
-    ) as f3:
+    with (
+        open(f"{class_data_dir}/caption.txt", "w") as f1,
+        open(f"{class_data_dir}/urls.txt", "w") as f2,
+        open(f"{class_data_dir}/images.txt", "w") as f3,
+    ):
        while total < num_class_images:
            images = class_images[count]
            count += 1
@@ -731,18 +731,18 @@ def main(args):
            if not class_images_dir.exists():
                class_images_dir.mkdir(parents=True, exist_ok=True)
            if args.real_prior:
-                assert (
-                    class_images_dir / "images"
-                ).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
-                assert (
-                    len(list((class_images_dir / "images").iterdir())) == args.num_class_images
-                ), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
-                assert (
-                    class_images_dir / "caption.txt"
-                ).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
-                assert (
-                    class_images_dir / "images.txt"
-                ).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
+                assert (class_images_dir / "images").exists(), (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
+                assert len(list((class_images_dir / "images").iterdir())) == args.num_class_images, (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
+                assert (class_images_dir / "caption.txt").exists(), (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
+                assert (class_images_dir / "images.txt").exists(), (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
                concept["class_prompt"] = os.path.join(class_images_dir, "caption.txt")
                concept["class_data_dir"] = os.path.join(class_images_dir, "images.txt")
                args.concepts_list[i] = concept
@@ -1014,7 +1014,7 @@ def main(args):

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

    # Enable TF32 for faster training on Ampere GPUs,
@@ -895,7 +895,10 @@ def _encode_prompt_with_t5(

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

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

    _, seq_len, _ = prompt_embeds.shape
@@ -936,9 +939,13 @@ def _encode_prompt_with_clip(

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

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

    # duplicate text embeddings for each generation per prompt, using mps friendly method
    prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
@@ -958,7 +965,12 @@ def encode_prompt(
 ):
    prompt = [prompt] if isinstance(prompt, str) else prompt
    batch_size = len(prompt)
-    dtype = text_encoders[0].dtype
+
+    if hasattr(text_encoders[0], "module"):
+        dtype = text_encoders[0].module.dtype
+    else:
+        dtype = text_encoders[0].dtype
+
    device = device if device is not None else text_encoders[1].device
    pooled_prompt_embeds = _encode_prompt_with_clip(
        text_encoder=text_encoders[0],
@@ -1590,7 +1602,7 @@ def main(args):
                )

                # handle guidance
-                if accelerator.unwrap_model(transformer).config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                    guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                    guidance = guidance.expand(model_input.shape[0])
                else:
@@ -1716,9 +1728,9 @@ def main(args):
                pipeline = FluxPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one, keep_fp32_wrapper=False),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two, keep_fp32_wrapper=False),
-                    transformer=accelerator.unwrap_model(transformer, keep_fp32_wrapper=False),
+                    text_encoder=unwrap_model(text_encoder_one, keep_fp32_wrapper=False),
+                    text_encoder_2=unwrap_model(text_encoder_two, keep_fp32_wrapper=False),
+                    transformer=unwrap_model(transformer, keep_fp32_wrapper=False),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
@@ -982,7 +982,7 @@ def main(args):

        lora_state_dict, network_alphas = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)

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

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

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

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

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

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

    return images

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

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

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

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

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

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

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

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

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1619,7 +1638,7 @@ def main(args):
        if args.train_text_encoder:
            text_encoder_one.train()
            # set top parameter requires_grad = True for gradient checkpointing works
-            accelerator.unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
+            unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)

        for step, batch in enumerate(train_dataloader):
            models_to_accumulate = [transformer]
@@ -1710,7 +1729,7 @@ def main(args):
                )

                # handle guidance
-                if accelerator.unwrap_model(transformer).config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                    guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                    guidance = guidance.expand(model_input.shape[0])
                else:
@@ -1828,9 +1847,9 @@ def main(args):
                pipeline = FluxPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
-                    transformer=accelerator.unwrap_model(transformer),
+                    text_encoder=unwrap_model(text_encoder_one),
+                    text_encoder_2=unwrap_model(text_encoder_two),
+                    transformer=unwrap_model(transformer),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
@@ -1053,7 +1053,7 @@ def main(args):
        lora_state_dict = Lumina2Text2ImgPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1064,7 +1064,7 @@ def main(args):
        lora_state_dict = SanaPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1355,7 +1355,7 @@ def main(args):
        lora_state_dict = StableDiffusion3Pipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -118,7 +118,7 @@ def save_model_card(
            )

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

 <Gallery />

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

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

        lora_state_dict, network_alphas = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
        if incompatible_keys is not None:
@@ -91,9 +91,9 @@ def log_validation(flux_transformer, args, accelerator, weight_dtype, step, is_f
            torch_dtype=weight_dtype,
        )
        pipeline.load_lora_weights(args.output_dir)
-        assert (
-            pipeline.transformer.config.in_channels == initial_channels * 2
-        ), f"{pipeline.transformer.config.in_channels=}"
+        assert pipeline.transformer.config.in_channels == initial_channels * 2, (
+            f"{pipeline.transformer.config.in_channels=}"
+        )

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

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

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


 def download_image(url):
-    image = PIL.Image.open(requests.get(url, stream=True).raw)
+    image = PIL.Image.open(requests.get(url, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)
    image = PIL.ImageOps.exif_transpose(image)
    image = image.convert("RGB")
    return image
@@ -1081,9 +1081,9 @@ class AutoConfig:
                    f"textual_inversion_path: {search_word} -> {textual_inversion_path.model_status.site_url}"
                )

-                pretrained_model_name_or_paths[
-                    pretrained_model_name_or_paths.index(search_word)
-                ] = textual_inversion_path.model_path
+                pretrained_model_name_or_paths[pretrained_model_name_or_paths.index(search_word)] = (
+                    textual_inversion_path.model_path
+                )

        self.load_textual_inversion(
            pretrained_model_name_or_paths, token=tokens, tokenizer=tokenizer, text_encoder=text_encoder, **kwargs
@@ -1,20 +1,27 @@
-# AnyTextPipeline Pipeline
+# AnyTextPipeline

 Project page: https://aigcdesigngroup.github.io/homepage_anytext

 "AnyText comprises a diffusion pipeline with two primary elements: an auxiliary latent module and a text embedding module. The former uses inputs like text glyph, position, and masked image to generate latent features for text generation or editing. The latter employs an OCR model for encoding stroke data as embeddings, which blend with image caption embeddings from the tokenizer to generate texts that seamlessly integrate with the background. We employed text-control diffusion loss and text perceptual loss for training to further enhance writing accuracy."

-Each text line that needs to be generated should be enclosed in double quotes. For any usage questions, please refer to the [paper](https://arxiv.org/abs/2311.03054).
+> **Note:** Each text line that needs to be generated should be enclosed in double quotes.

+For any usage questions, please refer to the [paper](https://arxiv.org/abs/2311.03054).
+
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/gist/tolgacangoz/b87ec9d2f265b448dd947c9d4a0da389/anytext.ipynb)

 ```py
+# This example requires the `anytext_controlnet.py` file:
+# !git clone --depth 1 https://github.com/huggingface/diffusers.git
+# %cd diffusers/examples/research_projects/anytext
+# Let's choose a font file shared by an HF staff:
+# !wget https://huggingface.co/spaces/ysharma/TranslateQuotesInImageForwards/resolve/main/arial-unicode-ms.ttf
+
 import torch
 from diffusers import DiffusionPipeline
 from anytext_controlnet import AnyTextControlNetModel
 from diffusers.utils import load_image

-# I chose a font file shared by an HF staff:
-# !wget https://huggingface.co/spaces/ysharma/TranslateQuotesInImageForwards/resolve/main/arial-unicode-ms.ttf

 anytext_controlnet = AnyTextControlNetModel.from_pretrained("tolgacangoz/anytext-controlnet", torch_dtype=torch.float16,
                                                            variant="fp16",)
@@ -26,6 +33,7 @@ pipe = DiffusionPipeline.from_pretrained("tolgacangoz/anytext", font_path="arial
 # generate image
 prompt = 'photo of caramel macchiato coffee on the table, top-down perspective, with "Any" "Text" written on it using cream'
 draw_pos = load_image("https://raw.githubusercontent.com/tyxsspa/AnyText/refs/heads/main/example_images/gen9.png")
+# There are two modes: "generate" and "edit". "edit" mode requires `ori_image` parameter for the image to be edited.
 image = pipe(prompt, num_inference_steps=20, mode="generate", draw_pos=draw_pos,
             ).images[0]
 image
@@ -146,14 +146,17 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 EXAMPLE_DOC_STRING = """
    Examples:
        ```py
+        >>> # This example requires the `anytext_controlnet.py` file:
+        >>> # !git clone --depth 1 https://github.com/huggingface/diffusers.git
+        >>> # %cd diffusers/examples/research_projects/anytext
+        >>> # Let's choose a font file shared by an HF staff:
+        >>> # !wget https://huggingface.co/spaces/ysharma/TranslateQuotesInImageForwards/resolve/main/arial-unicode-ms.ttf
+
        >>> import torch
        >>> from diffusers import DiffusionPipeline
        >>> from anytext_controlnet import AnyTextControlNetModel
        >>> from diffusers.utils import load_image

-        >>> # I chose a font file shared by an HF staff:
-        >>> !wget https://huggingface.co/spaces/ysharma/TranslateQuotesInImageForwards/resolve/main/arial-unicode-ms.ttf
-
        >>> anytext_controlnet = AnyTextControlNetModel.from_pretrained("tolgacangoz/anytext-controlnet", torch_dtype=torch.float16,
        ...                                                             variant="fp16",)
        >>> pipe = DiffusionPipeline.from_pretrained("tolgacangoz/anytext", font_path="arial-unicode-ms.ttf",
@@ -165,6 +168,7 @@ EXAMPLE_DOC_STRING = """
        >>> # generate image
        >>> prompt = 'photo of caramel macchiato coffee on the table, top-down perspective, with "Any" "Text" written on it using cream'
        >>> draw_pos = load_image("https://raw.githubusercontent.com/tyxsspa/AnyText/refs/heads/main/example_images/gen9.png")
+        >>> # There are two modes: "generate" and "edit". "edit" mode requires `ori_image` parameter for the image to be edited.
        >>> image = pipe(prompt, num_inference_steps=20, mode="generate", draw_pos=draw_pos,
        ...              ).images[0]
        >>> image
@@ -183,9 +187,9 @@ def get_clip_token_for_string(tokenizer, string):
        return_tensors="pt",
    )
    tokens = batch_encoding["input_ids"]
-    assert (
-        torch.count_nonzero(tokens - 49407) == 2
-    ), f"String '{string}' maps to more than a single token. Please use another string"
+    assert torch.count_nonzero(tokens - 49407) == 2, (
+        f"String '{string}' maps to more than a single token. Please use another string"
+    )
    return tokens[0, 1]


@@ -257,11 +261,11 @@ class EmbeddingManager(ModelMixin, ConfigMixin):
            idx = tokenized_text[i] == self.placeholder_token.to(device)
            if sum(idx) > 0:
                if i >= len(self.text_embs_all):
-                    print("truncation for log images...")
+                    logger.warning("truncation for log images...")
                    break
                text_emb = torch.cat(self.text_embs_all[i], dim=0)
                if sum(idx) != len(text_emb):
-                    print("truncation for long caption...")
+                    logger.warning("truncation for long caption...")
                text_emb = text_emb.to(embedded_text.device)
                embedded_text[i][idx] = text_emb[: sum(idx)]
        return embedded_text
@@ -1058,6 +1062,8 @@ class AuxiliaryLatentModule(ModelMixin, ConfigMixin):
                    raise ValueError(f"Can't read ori_image image from {ori_image}!")
            elif isinstance(ori_image, torch.Tensor):
                ori_image = ori_image.cpu().numpy()
+            elif isinstance(ori_image, PIL.Image.Image):
+                ori_image = np.array(ori_image.convert("RGB"))
            else:
                if not isinstance(ori_image, np.ndarray):
                    raise ValueError(f"Unknown format of ori_image: {type(ori_image)}")
@@ -312,9 +312,9 @@ class PatchEmbed(nn.Module):

    def forward(self, x):
        B, C, H, W = x.shape
-        assert (
-            H == self.img_size[0] and W == self.img_size[1]
-        ), f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        assert H == self.img_size[0] and W == self.img_size[1], (
+            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        )
        x = self.proj(x).flatten(2).permute(0, 2, 1)
        return x

@@ -627,6 +627,7 @@ def main(args):
        ema_vae = EMAModel(vae.parameters(), model_cls=AutoencoderKL, model_config=vae.config)
    perceptual_loss = lpips.LPIPS(net="vgg").eval()
    discriminator = NLayerDiscriminator(input_nc=3, n_layers=3, use_actnorm=False).apply(weights_init)
+    discriminator = torch.nn.SyncBatchNorm.convert_sync_batchnorm(discriminator)

    # Taken from [Sayak Paul's Diffusers PR #6511](https://github.com/huggingface/diffusers/pull/6511/files)
    def unwrap_model(model):
@@ -951,13 +952,20 @@ def main(args):
                    logits_fake = discriminator(reconstructions)
                    disc_loss = hinge_d_loss if args.disc_loss == "hinge" else vanilla_d_loss
                    disc_factor = args.disc_factor if global_step >= args.disc_start else 0.0
-                    disc_loss = disc_factor * disc_loss(logits_real, logits_fake)
+                    d_loss = disc_factor * disc_loss(logits_real, logits_fake)
                    logs = {
-                        "disc_loss": disc_loss.detach().mean().item(),
+                        "disc_loss": d_loss.detach().mean().item(),
                        "logits_real": logits_real.detach().mean().item(),
                        "logits_fake": logits_fake.detach().mean().item(),
                        "disc_lr": disc_lr_scheduler.get_last_lr()[0],
                    }
+                    accelerator.backward(d_loss)
+                    if accelerator.sync_gradients:
+                        params_to_clip = discriminator.parameters()
+                        accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+                    disc_optimizer.step()
+                    disc_lr_scheduler.step()
+                    disc_optimizer.zero_grad(set_to_none=args.set_grads_to_none)
            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                progress_bar.update(1)
@@ -619,7 +619,7 @@ def main(args):

            optimizer.step()
            lr_scheduler.step()
-            logger.info(f"max GPU_mem cost is {torch.cuda.max_memory_allocated()/2**20} MB", ranks=[0])
+            logger.info(f"max GPU_mem cost is {torch.cuda.max_memory_allocated() / 2**20} MB", ranks=[0])
            # Checks if the accelerator has performed an optimization step behind the scenes
            progress_bar.update(1)
            global_step += 1
@@ -381,9 +381,7 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step)
                validation_prompt = log["validation_prompt"]
                validation_image = log["validation_image"]

-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]

                for image in images:
                    formatted_images.append(np.asarray(image))
@@ -805,21 +803,20 @@ def parse_args(input_args=None):
        "--control_type",
        type=str,
        default="canny",
-        help=("The type of controlnet conditioning image to use. One of `canny`, `depth`" " Defaults to `canny`."),
+        help=("The type of controlnet conditioning image to use. One of `canny`, `depth` Defaults to `canny`."),
    )
    parser.add_argument(
        "--transformer_layers_per_block",
        type=str,
        default=None,
-        help=("The number of layers per block in the transformer. If None, defaults to" " `args.transformer_layers`."),
+        help=("The number of layers per block in the transformer. If None, defaults to `args.transformer_layers`."),
    )
    parser.add_argument(
        "--old_style_controlnet",
        action="store_true",
        default=False,
        help=(
-            "Use the old style controlnet, which is a single transformer layer with"
-            " a single head. Defaults to False."
+            "Use the old style controlnet, which is a single transformer layer with a single head. Defaults to False."
        ),
    )

@@ -86,7 +86,7 @@ def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: st


 def log_validation(args, unet, accelerator, weight_dtype, epoch, is_final_validation=False):
-    logger.info(f"Running validation... \n Generating images with prompts:\n" f" {VALIDATION_PROMPTS}.")
+    logger.info(f"Running validation... \n Generating images with prompts:\n {VALIDATION_PROMPTS}.")

    # create pipeline
    pipeline = DiffusionPipeline.from_pretrained(
@@ -91,7 +91,7 @@ def import_model_class_from_model_name_or_path(


 def log_validation(args, unet, vae, accelerator, weight_dtype, epoch, is_final_validation=False):
-    logger.info(f"Running validation... \n Generating images with prompts:\n" f" {VALIDATION_PROMPTS}.")
+    logger.info(f"Running validation... \n Generating images with prompts:\n {VALIDATION_PROMPTS}.")

    if is_final_validation:
        if args.mixed_precision == "fp16":
@@ -91,7 +91,7 @@ def import_model_class_from_model_name_or_path(


 def log_validation(args, unet, vae, accelerator, weight_dtype, epoch, is_final_validation=False):
-    logger.info(f"Running validation... \n Generating images with prompts:\n" f" {VALIDATION_PROMPTS}.")
+    logger.info(f"Running validation... \n Generating images with prompts:\n {VALIDATION_PROMPTS}.")

    if is_final_validation:
        if args.mixed_precision == "fp16":
@@ -683,7 +683,7 @@ def main(args):

        lora_state_dict, network_alphas = StableDiffusionXLLoraLoaderMixin.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
        if incompatible_keys is not None:
@@ -89,7 +89,7 @@ def import_model_class_from_model_name_or_path(


 def log_validation(args, unet, vae, accelerator, weight_dtype, epoch, is_final_validation=False):
-    logger.info(f"Running validation... \n Generating images with prompts:\n" f" {VALIDATION_PROMPTS}.")
+    logger.info(f"Running validation... \n Generating images with prompts:\n {VALIDATION_PROMPTS}.")

    if is_final_validation:
        if args.mixed_precision == "fp16":
@@ -790,7 +790,7 @@ def main(args):

        lora_state_dict, network_alphas = StableDiffusionXLLoraLoaderMixin.lora_state_dict(input_dir)

-        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
        if incompatible_keys is not None:
@@ -783,7 +783,7 @@ def main(args):
        lora_state_dict = FluxPipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -26,8 +26,7 @@
    "%load_ext autoreload\n",
    "%autoreload 2\n",
    "\n",
-    "import torch\n",
-    "from diffusers import StableDiffusionGLIGENTextImagePipeline, StableDiffusionGLIGENPipeline"
+    "from diffusers import StableDiffusionGLIGENPipeline"
   ]
  },
  {
@@ -36,28 +35,25 @@
   "metadata": {},
   "outputs": [],
   "source": [
-    "import os\n",
+    "from transformers import CLIPTextModel, CLIPTokenizer\n",
+    "\n",
    "import diffusers\n",
    "from diffusers import (\n",
    "    AutoencoderKL,\n",
    "    DDPMScheduler,\n",
-    "    UNet2DConditionModel,\n",
-    "    UniPCMultistepScheduler,\n",
    "    EulerDiscreteScheduler,\n",
+    "    UNet2DConditionModel,\n",
    ")\n",
-    "from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer\n",
+    "\n",
+    "\n",
    "# pretrained_model_name_or_path = 'masterful/gligen-1-4-generation-text-box'\n",
    "\n",
-    "pretrained_model_name_or_path = '/root/data/zhizhonghuang/checkpoints/models--masterful--gligen-1-4-generation-text-box/snapshots/d2820dc1e9ba6ca082051ce79cfd3eb468ae2c83'\n",
+    "pretrained_model_name_or_path = \"/root/data/zhizhonghuang/checkpoints/models--masterful--gligen-1-4-generation-text-box/snapshots/d2820dc1e9ba6ca082051ce79cfd3eb468ae2c83\"\n",
    "\n",
    "tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder=\"tokenizer\")\n",
    "noise_scheduler = DDPMScheduler.from_pretrained(pretrained_model_name_or_path, subfolder=\"scheduler\")\n",
-    "text_encoder = CLIPTextModel.from_pretrained(\n",
-    "    pretrained_model_name_or_path, subfolder=\"text_encoder\"\n",
-    ")\n",
-    "vae = AutoencoderKL.from_pretrained(\n",
-    "    pretrained_model_name_or_path, subfolder=\"vae\"\n",
-    ")\n",
+    "text_encoder = CLIPTextModel.from_pretrained(pretrained_model_name_or_path, subfolder=\"text_encoder\")\n",
+    "vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder=\"vae\")\n",
    "# unet = UNet2DConditionModel.from_pretrained(\n",
    "#     pretrained_model_name_or_path, subfolder=\"unet\"\n",
    "# )\n",
@@ -71,9 +67,7 @@
   "metadata": {},
   "outputs": [],
   "source": [
-    "unet = UNet2DConditionModel.from_pretrained(\n",
-    "    '/root/data/zhizhonghuang/ckpt/GLIGEN_Text_Retrain_COCO'\n",
-    ")"
+    "unet = UNet2DConditionModel.from_pretrained(\"/root/data/zhizhonghuang/ckpt/GLIGEN_Text_Retrain_COCO\")"
   ]
  },
  {
@@ -108,6 +102,9 @@
   "metadata": {},
   "outputs": [],
   "source": [
+    "import numpy as np\n",
+    "\n",
+    "\n",
    "# prompt = 'A realistic image of landscape scene depicting a green car parking on the left of a blue truck, with a red air balloon and a bird in the sky'\n",
    "# gen_boxes = [('a green car', [21, 281, 211, 159]), ('a blue truck', [269, 283, 209, 160]), ('a red air balloon', [66, 8, 145, 135]), ('a bird', [296, 42, 143, 100])]\n",
    "\n",
@@ -117,10 +114,8 @@
    "# prompt = 'A realistic scene of three skiers standing in a line on the snow near a palm tree'\n",
    "# gen_boxes = [('a skier', [5, 152, 139, 168]), ('a skier', [278, 192, 121, 158]), ('a skier', [148, 173, 124, 155]), ('a palm tree', [404, 105, 103, 251])]\n",
    "\n",
-    "prompt = 'An oil painting of a pink dolphin jumping on the left of a steam boat on the sea'\n",
-    "gen_boxes = [('a steam boat', [232, 225, 257, 149]), ('a jumping pink dolphin', [21, 249, 189, 123])]\n",
-    "\n",
-    "import numpy as np\n",
+    "prompt = \"An oil painting of a pink dolphin jumping on the left of a steam boat on the sea\"\n",
+    "gen_boxes = [(\"a steam boat\", [232, 225, 257, 149]), (\"a jumping pink dolphin\", [21, 249, 189, 123])]\n",
    "\n",
    "boxes = np.array([x[1] for x in gen_boxes])\n",
    "boxes = boxes / 512\n",
@@ -166,7 +161,7 @@
   "metadata": {},
   "outputs": [],
   "source": [
-    "diffusers.utils.make_image_grid(images, 4, len(images)//4)"
+    "diffusers.utils.make_image_grid(images, 4, len(images) // 4)"
   ]
  },
  {
@@ -179,7 +174,7 @@
 ],
 "metadata": {
  "kernelspec": {
-   "display_name": "densecaption",
+   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
@@ -197,5 +192,5 @@
  }
 },
 "nbformat": 4,
- "nbformat_minor": 2
+ "nbformat_minor": 4
 }
@@ -15,8 +15,8 @@
 # limitations under the License.

 """
-    Script to fine-tune Stable Diffusion for LORA InstructPix2Pix.
-    Base code referred from: https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix.py
+Script to fine-tune Stable Diffusion for LORA InstructPix2Pix.
+Base code referred from: https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix.py
 """

 import argparse
@@ -54,6 +54,7 @@ from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionInstructPix2P
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import EMAModel, cast_training_params
 from diffusers.utils import check_min_version, convert_state_dict_to_diffusers, deprecate, is_wandb_available
+from diffusers.utils.constants import DIFFUSERS_REQUEST_TIMEOUT
 from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module
@@ -475,7 +476,7 @@ def convert_to_np(image, resolution):


 def download_image(url):
-    image = PIL.Image.open(requests.get(url, stream=True).raw)
+    image = PIL.Image.open(requests.get(url, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)
    image = PIL.ImageOps.exif_transpose(image)
    image = image.convert("RGB")
    return image
@@ -763,9 +763,9 @@ def main(args):
        # Parse instance and class inputs, and double check that lengths match
        instance_data_dir = args.instance_data_dir.split(",")
        instance_prompt = args.instance_prompt.split(",")
-        assert all(
-            x == len(instance_data_dir) for x in [len(instance_data_dir), len(instance_prompt)]
-        ), "Instance data dir and prompt inputs are not of the same length."
+        assert all(x == len(instance_data_dir) for x in [len(instance_data_dir), len(instance_prompt)]), (
+            "Instance data dir and prompt inputs are not of the same length."
+        )

        if args.with_prior_preservation:
            class_data_dir = args.class_data_dir.split(",")
@@ -788,9 +788,9 @@ def main(args):
                    negative_validation_prompts.append(None)
            args.validation_negative_prompt = negative_validation_prompts

-            assert num_of_validation_prompts == len(
-                negative_validation_prompts
-            ), "The length of negative prompts for validation is greater than the number of validation prompts."
+            assert num_of_validation_prompts == len(negative_validation_prompts), (
+                "The length of negative prompts for validation is greater than the number of validation prompts."
+            )
            args.validation_inference_steps = [args.validation_inference_steps] * num_of_validation_prompts
            args.validation_guidance_scale = [args.validation_guidance_scale] * num_of_validation_prompts

@@ -830,9 +830,9 @@ def main():
                # Let's make sure we don't update any embedding weights besides the newly added token
                index_no_updates = get_mask(tokenizer, accelerator)
                with torch.no_grad():
-                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[
-                        index_no_updates
-                    ] = orig_embeds_params[index_no_updates]
+                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[index_no_updates] = (
+                        orig_embeds_params[index_no_updates]
+                    )

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
@@ -886,9 +886,9 @@ def main():
                index_no_updates[min(placeholder_token_ids) : max(placeholder_token_ids) + 1] = False

                with torch.no_grad():
-                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[
-                        index_no_updates
-                    ] = orig_embeds_params[index_no_updates]
+                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[index_no_updates] = (
+                        orig_embeds_params[index_no_updates]
+                    )

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
@@ -164,9 +164,7 @@ def log_validation(
                validation_prompt = log["validation_prompt"]
                validation_image = log["validation_image"]

-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]

                for image in images:
                    formatted_images.append(np.asarray(image))
@@ -59,6 +59,7 @@ from diffusers.schedulers import (
    UnCLIPScheduler,
 )
 from diffusers.utils import is_accelerate_available, logging
+from diffusers.utils.constants import DIFFUSERS_REQUEST_TIMEOUT


 if is_accelerate_available():
@@ -1435,7 +1436,7 @@ def download_from_original_stable_diffusion_ckpt(
            config_url = "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/x4-upscaling.yaml"

        if config_url is not None:
-            original_config_file = BytesIO(requests.get(config_url).content)
+            original_config_file = BytesIO(requests.get(config_url, timeout=DIFFUSERS_REQUEST_TIMEOUT).content)
        else:
            with open(original_config_file, "r") as f:
                original_config_file = f.read()
@@ -663,8 +663,7 @@ class PromptDiffusionPipeline(
                    self.check_image(image, prompt, prompt_embeds)
        else:
            raise ValueError(
-                f"You have passed a list of images of length {len(image_pair)}."
-                f"Make sure the list size equals to two."
+                f"You have passed a list of images of length {len(image_pair)}.Make sure the list size equals to two."
            )

        # Check `controlnet_conditioning_scale`
@@ -1,8 +1,6 @@
 # Generating images using Flux and PyTorch/XLA

-The `flux_inference` script shows how to do image generation using Flux on TPU devices using PyTorch/XLA. It uses the pallas kernel for flash attention for faster generation.
-
-It has been tested on [Trillium](https://cloud.google.com/blog/products/compute/introducing-trillium-6th-gen-tpus) TPU versions. No other TPU types have been tested.
+The `flux_inference` script shows how to do image generation using Flux on TPU devices using PyTorch/XLA. It uses the pallas kernel for flash attention for faster generation using custom flash block sizes for better performance on [Trillium](https://cloud.google.com/blog/products/compute/introducing-trillium-6th-gen-tpus) TPU versions. No other TPU types have been tested.

 ## Create TPU

@@ -23,20 +21,23 @@ Verify that PyTorch and PyTorch/XLA were installed correctly:
 python3 -c "import torch; import torch_xla;"
 ```

-Install dependencies
+Clone the diffusers repo and install dependencies

 ```bash
+git clone https://github.com/huggingface/diffusers.git
+cd diffusers
 pip install transformers accelerate sentencepiece structlog
-pushd ../../..
 pip install .
-popd
+cd examples/research_projects/pytorch_xla/inference/flux/
 ```

 ## Run the inference job

 ### Authenticate

-Run the following command to authenticate your token in order to download Flux weights.
+**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
@@ -50,51 +51,116 @@ python flux_inference.py

 The script loads the text encoders onto the CPU and the Flux transformer and VAE models onto the TPU. The first time the script runs, the compilation time is longer, while the cache stores the compiled programs. On subsequent runs, compilation is much faster and the subsequent passes being the fastest. 

-On a Trillium v6e-4, you should expect ~9 sec / 4 images or 2.25 sec / image (as devices run generation in parallel):
+On a Trillium v6e-4, you should expect ~6 sec / 4 images or 1.5 sec / image (as devices run generation in parallel):

 ```bash
 WARNING:root:libtpu.so and TPU device found. Setting PJRT_DEVICE=TPU.
-Loading checkpoint shards: 100%|███████████████████████████████| 2/2 [00:00<00:00,  7.01it/s]
-Loading pipeline components...:  40%|██████████▍               | 2/5 [00:00<00:00,  3.78it/s]You set `add_prefix_space`. The tokenizer needs to be converted from the slow tokenizers
-Loading pipeline components...: 100%|██████████████████████████| 5/5 [00:00<00:00,  6.72it/s]
-2025-01-10 00:51:25 [info     ] loading flux from black-forest-labs/FLUX.1-dev
-2025-01-10 00:51:25 [info     ] loading flux from black-forest-labs/FLUX.1-dev
-2025-01-10 00:51:26 [info     ] loading flux from black-forest-labs/FLUX.1-dev
-2025-01-10 00:51:26 [info     ] loading flux from black-forest-labs/FLUX.1-dev
-Loading pipeline components...: 100%|██████████████████████████| 3/3 [00:00<00:00,  4.29it/s]
-Loading pipeline components...: 100%|██████████████████████████| 3/3 [00:00<00:00,  3.26it/s]
-Loading pipeline components...: 100%|██████████████████████████| 3/3 [00:00<00:00,  3.27it/s]
-Loading pipeline components...: 100%|██████████████████████████| 3/3 [00:00<00:00,  3.25it/s]
-2025-01-10 00:51:34 [info     ] starting compilation run...   
-2025-01-10 00:51:35 [info     ] starting compilation run...   
-2025-01-10 00:51:37 [info     ] starting compilation run...   
-2025-01-10 00:51:37 [info     ] starting compilation run...   
-2025-01-10 00:52:52 [info     ] compilation took 78.5155531649998 sec.
-2025-01-10 00:52:53 [info     ] starting inference run...     
-2025-01-10 00:52:57 [info     ] compilation took 79.52986721400157 sec.
-2025-01-10 00:52:57 [info     ] compilation took 81.91776501700042 sec.
-2025-01-10 00:52:57 [info     ] compilation took 80.24951512600092 sec.
-2025-01-10 00:52:57 [info     ] starting inference run...     
-2025-01-10 00:52:57 [info     ] starting inference run...     
-2025-01-10 00:52:58 [info     ] starting inference run...     
-2025-01-10 00:53:22 [info     ] inference time: 25.112665320000815
-2025-01-10 00:53:30 [info     ] inference time: 7.7019307739992655
-2025-01-10 00:53:38 [info     ] inference time: 7.693858365000779
-2025-01-10 00:53:46 [info     ] inference time: 7.690621814001133
-2025-01-10 00:53:53 [info     ] inference time: 7.679490454000188
-2025-01-10 00:54:01 [info     ] inference time: 7.68949568500102
-2025-01-10 00:54:09 [info     ] inference time: 7.686633744000574
-2025-01-10 00:54:16 [info     ] inference time: 7.696786873999372
-2025-01-10 00:54:24 [info     ] inference time: 7.691988694999964
-2025-01-10 00:54:32 [info     ] inference time: 7.700649563999832
-2025-01-10 00:54:39 [info     ] inference time: 7.684993574001055
-2025-01-10 00:54:47 [info     ] inference time: 7.68343457499941
-2025-01-10 00:54:55 [info     ] inference time: 7.667921153999487
-2025-01-10 00:55:02 [info     ] inference time: 7.683585194001353
-2025-01-10 00:55:06 [info     ] avg. inference over 15 iterations took 8.61202360273334 sec.
-2025-01-10 00:55:07 [info     ] avg. inference over 15 iterations took 8.952725123600006 sec.
-2025-01-10 00:55:10 [info     ] inference time: 7.673799695001435
-2025-01-10 00:55:10 [info     ] avg. inference over 15 iterations took 8.849190365400379 sec.
-2025-01-10 00:55:10 [info     ] saved metric information as /tmp/metrics_report.txt
-2025-01-10 00:55:12 [info     ] avg. inference over 15 iterations took 8.940161458400205 sec.
+Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00,  7.06it/s]
+Loading pipeline components...:  60%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████                                                                                                | 3/5 [00:00<00:00,  6.80it/s]You set `add_prefix_space`. The tokenizer needs to be converted from the slow tokenizers
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00,  6.28it/s]
+2025-03-14 21:17:53 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+2025-03-14 21:17:53 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...:   0%|                                                                                                                                                                                                                                                        | 0/3 [00:00<?, ?it/s]2025-03-14 21:17:53 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+2025-03-14 21:17:53 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...:   0%|                                                                                                                                                                                                                                                        | 0/3 [00:00<?, ?it/s]2025-03-14 21:17:54 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+2025-03-14 21:17:54 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.66it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  4.48it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.32it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.69it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.74it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.10it/s]
+2025-03-14 21:17:56 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...:   0%|                                                                                                                                                                                                                                                        | 0/3 [00:00<?, ?it/s]2025-03-14 21:17:56 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.55it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00,  1.46it/s]
+2025-03-14 21:18:34 [info     ] starting compilation run...   
+2025-03-14 21:18:37 [info     ] starting compilation run...   
+2025-03-14 21:18:38 [info     ] starting compilation run...   
+2025-03-14 21:18:39 [info     ] starting compilation run...   
+2025-03-14 21:18:41 [info     ] starting compilation run...   
+2025-03-14 21:18:41 [info     ] starting compilation run...   
+2025-03-14 21:18:42 [info     ] starting compilation run...   
+2025-03-14 21:18:43 [info     ] starting compilation run...   
+ 82%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▊                                                | 23/28 [13:35<03:04, 36.80s/it]2025-03-14 21:33:42.057559: W torch_xla/csrc/runtime/pjrt_computation_client.cc:667] Failed to deserialize executable: INTERNAL: TfrtTpuExecutable proto deserialization failed while parsing core program!
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:27<00:00, 35.28s/it]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:27<00:00, 35.26s/it]
+2025-03-14 21:36:38 [info     ] compilation took 1079.3314765350078 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:12<00:00, 34.73s/it]
+2025-03-14 21:36:38 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:12<00:00, 34.73s/it]
+2025-03-14 21:36:38 [info     ] compilation took 1081.89390801001 sec.
+2025-03-14 21:36:39 [info     ] starting inference run...     
+2025-03-14 21:36:39 [info     ] compilation took 1077.1543154849933 sec.
+2025-03-14 21:36:39 [info     ] compilation took 1075.7239800530078 sec.
+2025-03-14 21:36:39 [info     ] starting inference run...     
+2025-03-14 21:36:40 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:22<00:00, 35.10s/it]
+2025-03-14 21:36:50 [info     ] compilation took 1088.1632604240003 sec.
+2025-03-14 21:36:50 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:28<00:00, 35.32s/it]
+2025-03-14 21:36:55 [info     ] compilation took 1096.8027802760043 sec.
+2025-03-14 21:36:56 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:59<00:00, 36.40s/it]
+2025-03-14 21:37:08 [info     ] compilation took 1113.8591305939917 sec.
+2025-03-14 21:37:08 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:55<00:00, 36.26s/it]
+2025-03-14 21:37:22 [info     ] compilation took 1120.5590810020076 sec.
+2025-03-14 21:37:22 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.00it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:09<00:00,  2.98it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:06<00:00,  4.08it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:09<00:00,  2.82it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:08<00:00,  3.34it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:06<00:00,  4.22it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:06<00:00,  4.09it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:11<00:00,  2.41it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:06<00:00,  4.50it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.10it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.27it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  4.80it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.39it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.39it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.67it/s]
+ 29%|█████████████████████████████████████████████████████████████████████████████▍                                                                                                                                                                                                 | 8/28 [00:01<00:03,  6.08it/s]/home/jfacevedo_google_com/diffusers/src/diffusers/image_processor.py:147: RuntimeWarning: invalid value encountered in cast
+  images = (images * 255).round().astype("uint8")
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.82it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.93it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.02it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.02it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.98it/s]
+ 71%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▊                                                                             | 20/28 [00:03<00:01,  6.03it/s]2025-03-14 21:38:32 [info     ] inference time: 5.962021178987925
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.89it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.09it/s]
+2025-03-14 21:38:32 [info     ] avg. inference over 5 iterations took 7.2685392687970305 sec.
+2025-03-14 21:38:32 [info     ] avg. inference over 5 iterations took 7.402720856998348 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.01it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.89it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.96it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.06it/s]
+ 71%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▊                                                                             | 20/28 [00:03<00:01,  6.01it/s]2025-03-14 21:38:38 [info     ] inference time: 5.950578948002658
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.87it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.09it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.00it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.86it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.99it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.05it/s]
+2025-03-14 21:38:43 [info     ] avg. inference over 5 iterations took 6.763298449796276 sec.
+ 71%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▊                                                                             | 20/28 [00:03<00:01,  6.04it/s]2025-03-14 21:38:44 [info     ] inference time: 5.949129879008979
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.92it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.10it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.02it/s]
+ 39%|██████████████████████████████████████████████████████████████████████████████████████████████████████████                                                                                                                                                                    | 11/28 [00:01<00:02,  5.98it/s]2025-03-14 21:38:46 [info     ] avg. inference over 5 iterations took 7.221068455604836 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.96it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.08it/s]
+ 93%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▋                   | 26/28 [00:04<00:00,  5.92it/s]2025-03-14 21:38:50 [info     ] inference time: 5.954778069004533
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.90it/s]
+ 11%|█████████████████████████████                                                                                                                                                                                                                                                  | 3/28 [00:00<00:04,  6.03it/s]2025-03-14 21:38:50 [info     ] avg. inference over 5 iterations took 6.05970350120042 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.02it/s]
+ 32%|███████████████████████████████████████████████████████████████████████████████████████                                                                                                                                                                                        | 9/28 [00:01<00:03,  5.99it/s]2025-03-14 21:38:51 [info     ] avg. inference over 5 iterations took 6.018543455796316 sec.
+ 54%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▋                                                                                                                             | 15/28 [00:02<00:02,  6.00it/s]2025-03-14 21:38:52 [info     ] avg. inference over 5 iterations took 5.9609976705978625 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.97it/s]
+2025-03-14 21:38:56 [info     ] inference time: 5.944058528999449
+2025-03-14 21:38:56 [info     ] avg. inference over 5 iterations took 5.952113320800708 sec.
+2025-03-14 21:38:56 [info     ] saved metric information as /tmp/metrics_report.txt
 ```
@@ -9,6 +9,7 @@ import torch_xla.debug.metrics as met
 import torch_xla.debug.profiler as xp
 import torch_xla.distributed.xla_multiprocessing as xmp
 import torch_xla.runtime as xr
+from torch_xla.experimental.custom_kernel import FlashAttention

 from diffusers import FluxPipeline

@@ -36,6 +37,19 @@ def _main(index, args, text_pipe, ckpt_id):
        ckpt_id, text_encoder=None, tokenizer=None, text_encoder_2=None, tokenizer_2=None, torch_dtype=torch.bfloat16
    ).to(device0)
    flux_pipe.transformer.enable_xla_flash_attention(partition_spec=("data", None, None, None), is_flux=True)
+    FlashAttention.DEFAULT_BLOCK_SIZES = {
+        "block_q": 1536,
+        "block_k_major": 1536,
+        "block_k": 1536,
+        "block_b": 1536,
+        "block_q_major_dkv": 1536,
+        "block_k_major_dkv": 1536,
+        "block_q_dkv": 1536,
+        "block_k_dkv": 1536,
+        "block_q_dq": 1536,
+        "block_k_dq": 1536,
+        "block_k_major_dq": 1536,
+    }

    prompt = "photograph of an electronics chip in the shape of a race car with trillium written on its side"
    width = args.width
@@ -69,14 +83,14 @@ def _main(index, args, text_pipe, ckpt_id):
    xm.set_rng_state(seed=unique_seed, device=device0)
    times = []
    logger.info("starting inference run...")
+    with torch.no_grad():
+        prompt_embeds, pooled_prompt_embeds, text_ids = text_pipe.encode_prompt(
+            prompt=prompt, prompt_2=None, max_sequence_length=512
+        )
+    prompt_embeds = prompt_embeds.to(device0)
+    pooled_prompt_embeds = pooled_prompt_embeds.to(device0)
    for _ in range(args.itters):
        ts = perf_counter()
-        with torch.no_grad():
-            prompt_embeds, pooled_prompt_embeds, text_ids = text_pipe.encode_prompt(
-                prompt=prompt, prompt_2=None, max_sequence_length=512
-            )
-        prompt_embeds = prompt_embeds.to(device0)
-        pooled_prompt_embeds = pooled_prompt_embeds.to(device0)

        if args.profile:
            xp.trace_detached(f"localhost:{profiler_port}", str(profile_path), duration_ms=profile_duration)
@@ -92,7 +106,7 @@ def _main(index, args, text_pipe, ckpt_id):
        if index == 0:
            logger.info(f"inference time: {inference_time}")
        times.append(inference_time)
-    logger.info(f"avg. inference over {args.itters} iterations took {sum(times)/len(times)} sec.")
+    logger.info(f"avg. inference over {args.itters} iterations took {sum(times) / len(times)} sec.")
    image.save(f"/tmp/inference_out-{index}.png")
    if index == 0:
        metrics_report = met.metrics_report()
@@ -173,7 +173,7 @@ class TrainSD:
        if not dataloader_exception:
            xm.wait_device_ops()
            total_time = time.time() - last_time
-            print(f"Average step time: {total_time/(self.args.max_train_steps-measure_start_step)}")
+            print(f"Average step time: {total_time / (self.args.max_train_steps - measure_start_step)}")
        else:
            print("dataloader exception happen, skip result")
            return
@@ -622,7 +622,7 @@ def main(args):
    num_devices_per_host = num_devices // num_hosts
    if xm.is_master_ordinal():
        print("***** Running training *****")
-        print(f"Instantaneous batch size per device = {args.train_batch_size // num_devices_per_host }")
+        print(f"Instantaneous batch size per device = {args.train_batch_size // num_devices_per_host}")
        print(
            f"Total train batch size (w. parallel, distributed & accumulation) = {args.train_batch_size * num_hosts}"
        )
@@ -1057,7 +1057,7 @@ def main(args):

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

    # Enable TF32 for faster training on Ampere GPUs,
@@ -1021,7 +1021,7 @@ def main(args):

        lora_state_dict, network_alphas = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)

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

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