Update tailscale action to main

* fix: euledm when using the exp sigma schedule.

* fix-copies

* remove print.

* reduce friction

* yiyi's suggestioms

.github/workflows/build_docker_images.yml

@@ -25,17 +25,17 @@ jobs:
     steps:
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v1
-      
+
       - name: Check out code
         uses: actions/checkout@v3
-      
+
       - name: Find Changed Dockerfiles
         id: file_changes
         uses: jitterbit/get-changed-files@v1
         with:
           format: 'space-delimited'
           token: ${{ secrets.GITHUB_TOKEN }}
-      
+
       - name: Build Changed Docker Images
         run: |
           CHANGED_FILES="${{ steps.file_changes.outputs.all }}"
@@ -52,7 +52,7 @@ jobs:
   build-and-push-docker-images:
     runs-on: [ self-hosted, intel-cpu, 8-cpu, ci ]
     if: github.event_name != 'pull_request'
-    
+
     permissions:
       contents: read
       packages: write
@@ -69,6 +69,7 @@ jobs:
           - diffusers-flax-tpu
           - diffusers-onnxruntime-cpu
           - diffusers-onnxruntime-cuda
+          - diffusers-doc-builder
 
     steps:
       - name: Checkout repository

.github/workflows/build_documentation.yml

@@ -21,7 +21,7 @@ jobs:
       package: diffusers
       notebook_folder: diffusers_doc
       languages: en ko zh ja pt
-
+      custom_container: diffusers/diffusers-doc-builder
     secrets:
       token: ${{ secrets.HUGGINGFACE_PUSH }}
       hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/build_pr_documentation.yml

@@ -20,3 +20,4 @@ jobs:
       install_libgl1: true
       package: diffusers
       languages: en ko zh ja pt
+      custom_container: diffusers/diffusers-doc-builder

.github/workflows/nightly_tests.yml

@@ -59,7 +59,7 @@ jobs:
     runs-on: [single-gpu, nvidia-gpu, t4, ci]
     container:
       image: diffusers/diffusers-pytorch-cuda
-      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --gpus 0
+      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface/diffusers:/mnt/cache/ --gpus 0
     steps:
       - name: Checkout diffusers
         uses: actions/checkout@v3

.github/workflows/pr_test_peft_backend.yml

@@ -111,3 +111,21 @@ jobs:
           -s -v \
           --make-reports=tests_${{ matrix.config.report }} \
           tests/lora/
+        python -m pytest -n 4 --max-worker-restart=0 --dist=loadfile \
+          -s -v \
+          --make-reports=tests_models_lora_${{ matrix.config.report }} \
+          tests/models/ -k "lora"
+        
+    
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_${{ matrix.config.report }}_failures_short.txt
+        cat reports/tests_models_lora_${{ matrix.config.report }}_failures_short.txt
+    
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: pr_${{ matrix.config.report }}_test_reports
+        path: reports

.github/workflows/push_tests.yml

@@ -62,7 +62,7 @@ jobs:
     runs-on: [single-gpu, nvidia-gpu, t4, ci]
     container:
       image: diffusers/diffusers-pytorch-cuda
-      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface/diffusers:/mnt/cache/ --gpus 0 --privileged
+      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface/diffusers:/mnt/cache/ --gpus 0
     steps:
       - name: Checkout diffusers
         uses: actions/checkout@v3
@@ -71,12 +71,6 @@ jobs:
       - name: NVIDIA-SMI
         run: |
           nvidia-smi
-      - name: Tailscale
-        uses: huggingface/tailscale-action@v1
-        with:
-          authkey: ${{ secrets.TAILSCALE_SSH_AUTHKEY }}
-          slackChannel: ${{ secrets.SLACK_CIFEEDBACK_CHANNEL }}
-          slackToken: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
       - name: Install dependencies
         run: |
           python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
@@ -95,18 +89,11 @@ jobs:
             -s -v -k "not Flax and not Onnx" \
             --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
             tests/pipelines/${{ matrix.module }}
-      - name: Tailscale Wait
-        if: ${{ failure() || runner.debug == '1' }}
-        uses: huggingface/tailscale-action@v1
-        with:
-           waitForSSH: true
-           authkey: ${{ secrets.TAILSCALE_SSH_AUTHKEY }}
       - name: Failure short reports
         if: ${{ failure() }}
         run: |
           cat reports/tests_pipeline_${{ matrix.module }}_cuda_stats.txt
           cat reports/tests_pipeline_${{ matrix.module }}_cuda_failures_short.txt
-
       - name: Test suite reports artifacts
         if: ${{ always() }}
         uses: actions/upload-artifact@v2
@@ -202,12 +189,17 @@ jobs:
           -s -v -k "not Flax and not Onnx and not PEFTLoRALoading" \
           --make-reports=tests_peft_cuda \
           tests/lora/
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "lora and not Flax and not Onnx and not PEFTLoRALoading" \
+          --make-reports=tests_peft_cuda_models_lora \
+          tests/models/
 
     - name: Failure short reports
       if: ${{ failure() }}
       run: |
         cat reports/tests_peft_cuda_stats.txt
         cat reports/tests_peft_cuda_failures_short.txt
+        cat reports/tests_peft_cuda_models_lora_failures_short.txt
 
     - name: Test suite reports artifacts
       if: ${{ always() }}

.github/workflows/run_tests_from_a_pr.yml

@@ -0,0 +1,73 @@
+name: Check running SLOW tests from a PR (only GPU)
+
+on:
+  workflow_dispatch:
+    inputs:
+      docker_image:
+        default: 'diffusers/diffusers-pytorch-cuda'
+        description: 'Name of the Docker image'
+        required: true
+      branch: 
+        description: 'PR Branch to test on'
+        required: true
+      test:
+        description: 'Tests to run (e.g.: `tests/models`).'
+        required: true
+
+env:
+  DIFFUSERS_IS_CI: yes
+  IS_GITHUB_CI: "1"
+  HF_HOME: /mnt/cache
+  OMP_NUM_THREADS: 8
+  MKL_NUM_THREADS: 8
+  PYTEST_TIMEOUT: 600
+  RUN_SLOW: yes
+
+jobs:
+  run_tests:
+    name: "Run a test on our runner from a PR"
+    runs-on: [single-gpu, nvidia-gpu, t4, ci]
+    container:
+      image: ${{ github.event.inputs.docker_image }}
+      options: --gpus 0 --privileged --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+
+    steps:
+      - name: Validate test files input
+        id: validate_test_files
+        env: 
+          PY_TEST: ${{ github.event.inputs.test }}
+        run: |
+          if [[ ! "$PY_TEST" =~ ^tests/ ]]; then
+            echo "Error: The input string must start with 'tests/'."
+            exit 1
+          fi
+          
+          if [[ ! "$PY_TEST" =~ ^tests/(models|pipelines) ]]; then
+            echo "Error: The input string must contain either 'models' or 'pipelines' after 'tests/'."
+            exit 1
+          fi
+          
+          if [[ "$PY_TEST" == *";"* ]]; then
+            echo "Error: The input string must not contain ';'."
+            exit 1
+          fi
+          echo "$PY_TEST"
+
+      - name: Checkout PR branch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.inputs.branch }}
+          repository: ${{ github.event.pull_request.head.repo.full_name }}
+
+
+      - name: Install pytest 
+        run: | 
+          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+          python -m uv pip install -e [quality,test]
+          python -m uv pip install peft
+      
+      - name: Run tests
+        env: 
+            PY_TEST: ${{ github.event.inputs.test }}
+        run: |
+          pytest "$PY_TEST"

.github/workflows/ssh-runner.yml

@@ -25,7 +25,7 @@ jobs:
     runs-on: [single-gpu, nvidia-gpu, "${{ github.event.inputs.runner_type }}", ci]
     container:
       image: ${{ github.event.inputs.docker_image }}
-      options: --gpus all --privileged --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface/diffusers:/mnt/cache/ --gpus 0 --privileged
 
     steps:
       - name: Checkout diffusers
@@ -38,7 +38,7 @@ jobs:
           nvidia-smi
 
       - name: Tailscale # In order to be able to SSH when a test fails
-        uses: huggingface/tailscale-action@v1
+        uses: huggingface/tailscale-action@main
         with:
           authkey: ${{ secrets.TAILSCALE_SSH_AUTHKEY }}
           slackChannel: ${{ secrets.SLACK_CIFEEDBACK_CHANNEL }}

README.md

@@ -77,7 +77,7 @@ Please refer to the [How to use Stable Diffusion in Apple Silicon](https://huggi
 
 ## Quickstart
 
-Generating outputs is super easy with 🤗 Diffusers. To generate an image from text, use the `from_pretrained` method to load any pretrained diffusion model (browse the [Hub](https://huggingface.co/models?library=diffusers&sort=downloads) for 22000+ checkpoints):
+Generating outputs is super easy with 🤗 Diffusers. To generate an image from text, use the `from_pretrained` method to load any pretrained diffusion model (browse the [Hub](https://huggingface.co/models?library=diffusers&sort=downloads) for 25.000+ checkpoints):
 
 ```python
 from diffusers import DiffusionPipeline
@@ -219,7 +219,7 @@ Also, say 👋 in our public Discord channel <a href="https://discord.gg/G7tWnz9
 - https://github.com/deep-floyd/IF
 - https://github.com/bentoml/BentoML
 - https://github.com/bmaltais/kohya_ss
-- +9000 other amazing GitHub repositories 💪
+- +11.000 other amazing GitHub repositories 💪
 
 Thank you for using us ❤️.
 

docker/diffusers-doc-builder/Dockerfile

@@ -0,0 +1,52 @@
+FROM ubuntu:20.04
+LABEL maintainer="Hugging Face"
+LABEL repository="diffusers"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt-get -y update \
+    && apt-get install -y software-properties-common \
+    && add-apt-repository ppa:deadsnakes/ppa
+
+RUN apt install -y bash \
+                   build-essential \
+                   git \
+                   git-lfs \
+                   curl \
+                   ca-certificates \
+                   libsndfile1-dev \
+                   python3.10 \
+                   python3-pip \
+                   libgl1 \
+                   zip \
+                   wget \
+                   python3.10-venv && \
+    rm -rf /var/lib/apt/lists
+
+# make sure to use venv
+RUN python3.10 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
+    python3.10 -m uv pip install --no-cache-dir \
+        torch \
+        torchvision \
+        torchaudio \
+        invisible_watermark \
+        --extra-index-url https://download.pytorch.org/whl/cpu && \
+    python3.10 -m uv pip install --no-cache-dir \
+        accelerate \
+        datasets \
+        hf-doc-builder \
+        huggingface-hub \
+        Jinja2 \
+        librosa \
+        numpy \
+        scipy \
+        tensorboard \
+        transformers \
+        matplotlib \
+        setuptools==69.5.1
+
+CMD ["/bin/bash"]

docs/source/en/_toctree.yml

@@ -29,10 +29,8 @@
     title: Load community pipelines and components
   - local: using-diffusers/schedulers
     title: Load schedulers and models
-  - local: using-diffusers/using_safetensors
-    title: Load safetensors
   - local: using-diffusers/other-formats
-    title: Load different Stable Diffusion formats
+    title: Model files and layouts
   - local: using-diffusers/loading_adapters
     title: Load adapters
   - local: using-diffusers/push_to_hub
@@ -59,6 +57,8 @@
     title: Distributed inference with multiple GPUs
   - local: using-diffusers/merge_loras
     title: Merge LoRAs
+  - local: using-diffusers/scheduler_features
+    title: Scheduler features
   - local: using-diffusers/callback
     title: Pipeline callbacks
   - local: using-diffusers/reusing_seeds
@@ -68,6 +68,10 @@
   - local: using-diffusers/weighted_prompts
     title: Prompt techniques
   title: Inference techniques
+- sections:
+  - local: advanced_inference/outpaint
+    title: Outpainting
+  title: Advanced inference
 - sections:
   - local: using-diffusers/sdxl
     title: Stable Diffusion XL
@@ -93,6 +97,8 @@
     title: Trajectory Consistency Distillation-LoRA
   - local: using-diffusers/svd
     title: Stable Video Diffusion
+  - local: using-diffusers/marigold_usage
+    title: Marigold Computer Vision
   title: Specific pipeline examples
 - sections:
   - local: training/overview
@@ -231,13 +237,19 @@
     - local: api/models/consistency_decoder_vae
       title: ConsistencyDecoderVAE
     - local: api/models/transformer2d
-      title: Transformer2D
+      title: Transformer2DModel
+    - local: api/models/pixart_transformer2d
+      title: PixArtTransformer2DModel
+    - local: api/models/dit_transformer2d
+      title: DiTTransformer2DModel
+    - local: api/models/hunyuan_transformer2d
+      title: HunyuanDiT2DModel
     - local: api/models/transformer_temporal
-      title: Transformer Temporal
+      title: TransformerTemporalModel
     - local: api/models/prior_transformer
-      title: Prior Transformer
+      title: PriorTransformer
     - local: api/models/controlnet
-      title: ControlNet
+      title: ControlNetModel
     title: Models
     isExpanded: false
   - sections:
@@ -279,6 +291,8 @@
       title: DiffEdit
     - local: api/pipelines/dit
       title: DiT
+    - local: api/pipelines/hunyuandit
+      title: Hunyuan-DiT
     - local: api/pipelines/i2vgenxl
       title: I2VGen-XL
     - local: api/pipelines/pix2pix
@@ -295,6 +309,8 @@
       title: Latent Diffusion
     - local: api/pipelines/ledits_pp
       title: LEDITS++
+    - local: api/pipelines/marigold
+      title: Marigold
     - local: api/pipelines/panorama
       title: MultiDiffusion
     - local: api/pipelines/musicldm

docs/source/en/advanced_inference/outpaint.md

@@ -0,0 +1,231 @@
+<!--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.
+-->
+
+# Outpainting
+
+Outpainting extends an image beyond its original boundaries, allowing you to add, replace, or modify visual elements in an image while preserving the original image. Like [inpainting](../using-diffusers/inpaint), you want to fill the white area (in this case, the area outside of the original image) with new visual elements while keeping the original image (represented by a mask of black pixels). There are a couple of ways to outpaint, such as with a [ControlNet](https://hf.co/blog/OzzyGT/outpainting-controlnet) or with [Differential Diffusion](https://hf.co/blog/OzzyGT/outpainting-differential-diffusion).
+
+This guide will show you how to outpaint with an inpainting model, ControlNet, and a ZoeDepth estimator.
+
+Before you begin, make sure you have the [controlnet_aux](https://github.com/huggingface/controlnet_aux) library installed so you can use the ZoeDepth estimator.
+
+```py
+!pip install -q controlnet_aux
+```
+
+## Image preparation
+
+Start by picking an image to outpaint with and remove the background with a Space like [BRIA-RMBG-1.4](https://hf.co/spaces/briaai/BRIA-RMBG-1.4).
+
+<iframe
+	src="https://briaai-bria-rmbg-1-4.hf.space"
+	frameborder="0"
+	width="850"
+	height="450"
+></iframe>
+
+For example, remove the background from this image of a pair of shoes.
+
+<div class="flex flex-row gap-4">
+  <div class="flex-1">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/original-jordan.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">original image</figcaption>
+  </div>
+  <div class="flex-1">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/no-background-jordan.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">background removed</figcaption>
+  </div>
+</div>
+
+[Stable Diffusion XL (SDXL)](../using-diffusers/sdxl) models work best with 1024x1024 images, but you can resize the image to any size as long as your hardware has enough memory to support it. The transparent background in the image should also be replaced with a white background. Create a function (like the one below) that scales and pastes the image onto a white background.
+
+```py
+import random
+
+import requests
+import torch
+from controlnet_aux import ZoeDetector
+from PIL import Image, ImageOps
+
+from diffusers import (
+    AutoencoderKL,
+    ControlNetModel,
+    StableDiffusionXLControlNetPipeline,
+    StableDiffusionXLInpaintPipeline,
+)
+
+def scale_and_paste(original_image):
+    aspect_ratio = original_image.width / original_image.height
+
+    if original_image.width > original_image.height:
+        new_width = 1024
+        new_height = round(new_width / aspect_ratio)
+    else:
+        new_height = 1024
+        new_width = round(new_height * aspect_ratio)
+
+    resized_original = original_image.resize((new_width, new_height), Image.LANCZOS)
+    white_background = Image.new("RGBA", (1024, 1024), "white")
+    x = (1024 - new_width) // 2
+    y = (1024 - new_height) // 2
+    white_background.paste(resized_original, (x, y), resized_original)
+
+    return resized_original, white_background
+
+original_image = Image.open(
+    requests.get(
+        "https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/no-background-jordan.png",
+        stream=True,
+    ).raw
+).convert("RGBA")
+resized_img, white_bg_image = scale_and_paste(original_image)
+```
+
+To avoid adding unwanted extra details, use the ZoeDepth estimator to provide additional guidance during generation and to ensure the shoes remain consistent with the original image.
+
+```py
+zoe = ZoeDetector.from_pretrained("lllyasviel/Annotators")
+image_zoe = zoe(white_bg_image, detect_resolution=512, image_resolution=1024)
+image_zoe
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/zoedepth-jordan.png"/>
+</div>
+
+## Outpaint
+
+Once your image is ready, you can generate content in the white area around the shoes with [controlnet-inpaint-dreamer-sdxl](https://hf.co/destitech/controlnet-inpaint-dreamer-sdxl), a SDXL ControlNet trained for inpainting.
+
+Load the inpainting ControlNet, ZoeDepth model, VAE and pass them to the [`StableDiffusionXLControlNetPipeline`]. Then you can create an optional `generate_image` function (for convenience) to outpaint an initial image.
+
+```py
+controlnets = [
+    ControlNetModel.from_pretrained(
+        "destitech/controlnet-inpaint-dreamer-sdxl", torch_dtype=torch.float16, variant="fp16"
+    ),
+    ControlNetModel.from_pretrained(
+        "diffusers/controlnet-zoe-depth-sdxl-1.0", torch_dtype=torch.float16
+    ),
+]
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to("cuda")
+pipeline = StableDiffusionXLControlNetPipeline.from_pretrained(
+    "SG161222/RealVisXL_V4.0", torch_dtype=torch.float16, variant="fp16", controlnet=controlnets, vae=vae
+).to("cuda")
+
+def generate_image(prompt, negative_prompt, inpaint_image, zoe_image, seed: int = None):
+    if seed is None:
+        seed = random.randint(0, 2**32 - 1)
+
+    generator = torch.Generator(device="cpu").manual_seed(seed)
+
+    image = pipeline(
+        prompt,
+        negative_prompt=negative_prompt,
+        image=[inpaint_image, zoe_image],
+        guidance_scale=6.5,
+        num_inference_steps=25,
+        generator=generator,
+        controlnet_conditioning_scale=[0.5, 0.8],
+        control_guidance_end=[0.9, 0.6],
+    ).images[0]
+
+    return image
+
+prompt = "nike air jordans on a basketball court"
+negative_prompt = ""
+
+temp_image = generate_image(prompt, negative_prompt, white_bg_image, image_zoe, 908097)
+```
+
+Paste the original image over the initial outpainted image. You'll improve the outpainted background in a later step.
+
+```py
+x = (1024 - resized_img.width) // 2
+y = (1024 - resized_img.height) // 2
+temp_image.paste(resized_img, (x, y), resized_img)
+temp_image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/initial-outpaint.png"/>
+</div>
+
+> [!TIP]
+> Now is a good time to free up some memory if you're running low!
+>
+> ```py
+> pipeline=None
+> torch.cuda.empty_cache()
+> ```
+
+Now that you have an initial outpainted image, load the [`StableDiffusionXLInpaintPipeline`] with the [RealVisXL](https://hf.co/SG161222/RealVisXL_V4.0) model to generate the final outpainted image with better quality.
+
+```py
+pipeline = StableDiffusionXLInpaintPipeline.from_pretrained(
+    "OzzyGT/RealVisXL_V4.0_inpainting",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    vae=vae,
+).to("cuda")
+```
+
+Prepare a mask for the final outpainted image. To create a more natural transition between the original image and the outpainted background, blur the mask to help it blend better.
+
+```py
+mask = Image.new("L", temp_image.size)
+mask.paste(resized_img.split()[3], (x, y))
+mask = ImageOps.invert(mask)
+final_mask = mask.point(lambda p: p > 128 and 255)
+mask_blurred = pipeline.mask_processor.blur(final_mask, blur_factor=20)
+mask_blurred
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/blurred-mask.png"/>
+</div>
+
+Create a better prompt and pass it to the `generate_outpaint` function to generate the final outpainted image. Again, paste the original image over the final outpainted background.
+
+```py
+def generate_outpaint(prompt, negative_prompt, image, mask, seed: int = None):
+    if seed is None:
+        seed = random.randint(0, 2**32 - 1)
+
+    generator = torch.Generator(device="cpu").manual_seed(seed)
+
+    image = pipeline(
+        prompt,
+        negative_prompt=negative_prompt,
+        image=image,
+        mask_image=mask,
+        guidance_scale=10.0,
+        strength=0.8,
+        num_inference_steps=30,
+        generator=generator,
+    ).images[0]
+
+    return image
+
+prompt = "high quality photo of nike air jordans on a basketball court, highly detailed"
+negative_prompt = ""
+
+final_image = generate_outpaint(prompt, negative_prompt, temp_image, mask_blurred, 7688778)
+x = (1024 - resized_img.width) // 2
+y = (1024 - resized_img.height) // 2
+final_image.paste(resized_img, (x, y), resized_img)
+final_image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/final-outpaint.png"/>
+</div>

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

@@ -12,9 +12,9 @@ specific language governing permissions and limitations under the License.
 
 # Loading Pipelines and Models via `from_single_file`
 
-The `from_single_file` method allows you to load supported pipelines using a single checkpoint file as opposed to the folder format used by Diffusers. This is useful if you are working with many of the Stable Diffusion Web UI's (such as A1111) that extensively rely on a single file to distribute all the components of a diffusion model.
+The `from_single_file` method allows you to load supported pipelines using a single checkpoint file as opposed to Diffusers' multiple folders format. This is useful if you are working with Stable Diffusion Web UI's (such as A1111) that rely on a single file format to distribute all the components of a model.
 
-The `from_single_file` method also supports loading models in their originally distributed format. This means that supported models that have been finetuned with other services can be loaded directly into supported Diffusers model objects and pipelines.
+The `from_single_file` method also supports loading models in their originally distributed format. This means that supported models that have been finetuned with other services can be loaded directly into Diffusers model objects and pipelines.
 
 ## Pipelines that currently support `from_single_file` loading
 
@@ -59,7 +59,7 @@ pipe = StableDiffusionXLPipeline.from_single_file(ckpt_path)
 
 ## Setting components in a Pipeline using `from_single_file`
 
-Swap components of the pipeline by passing them directly to the `from_single_file` method. e.g If you would like use a different scheduler than the pipeline default.
+Set components of a pipeline by passing them directly to the `from_single_file` method. For example, here we are swapping out the pipeline's default scheduler with the `DDIMScheduler`.
 
 ```python
 from diffusers import StableDiffusionXLPipeline, DDIMScheduler
@@ -71,13 +71,15 @@ pipe = StableDiffusionXLPipeline.from_single_file(ckpt_path, scheduler=scheduler
 
 ```
 
+Here we are passing in a ControlNet model to the `StableDiffusionControlNetPipeline`.
+
 ```python
-from diffusers import StableDiffusionPipeline, ControlNetModel
+from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
 
 ckpt_path = "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.safetensors"
 
-controlnet = ControlNetModel.from_pretrained("https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.safetensors")
-pipe = StableDiffusionPipeline.from_single_file(ckpt_path, controlnet=controlnet)
+controlnet = ControlNetModel.from_pretrained("lllyasviel/control_v11p_sd15_canny")
+pipe = StableDiffusionControlNetPipeline.from_single_file(ckpt_path, controlnet=controlnet)
 
 ```
 
@@ -93,7 +95,7 @@ model = StableCascadeUNet.from_single_file(ckpt_path)
 
 ## Using a Diffusers model repository to configure single file loading
 
-Under the hood, `from_single_file` will try to determine a model repository to use to configure the components of the pipeline. You can also pass in a repository id to the `config` argument of the `from_single_file` method to explicitly set the repository to use.
+Under the hood, `from_single_file` will try to automatically determine a model repository to use to configure the components of a pipeline. You can also explicitly set the model repository to configure the pipeline with the `config` argument.
 
 ```python
 from diffusers import StableDiffusionXLPipeline
@@ -105,9 +107,19 @@ pipe = StableDiffusionXLPipeline.from_single_file(ckpt_path, config=repo_id)
 
 ```
 
+In the example above, since we explicitly passed `repo_id="segmind/SSD-1B"` to the `config` argument, it will use this [configuration file](https://huggingface.co/segmind/SSD-1B/blob/main/unet/config.json) from the `unet` subfolder in `"segmind/SSD-1B"` to configure the `unet` component of the pipeline; Similarly, it will use the `config.json` file from `vae` subfolder to configure the `vae` model, `config.json` file from `text_encoder` folder to configure `text_encoder` and so on.
+
+<Tip>
+
+Most of the time you do not need to explicitly set a `config` argument. `from_single_file` will automatically map the checkpoint to the appropriate model repository. However, this option can be useful in cases where model components in the checkpoint might have been changed from what was originally distributed, or in cases where a checkpoint file might not have the necessary metadata to correctly determine the configuration to use for the pipeline.
+
+</Tip>
+
 ## Override configuration options when using single file loading
 
-Override the default model or pipeline configuration options when using `from_single_file` by passing in the relevant arguments directly to the `from_single_file` method. Any argument that is supported by the model or pipeline class can be configured in this way:
+Override the default model or pipeline configuration options by providing the relevant arguments directly to the `from_single_file` method. Any argument supported by the model or pipeline class can be configured in this way:
+
+### Setting a pipeline configuration option
 
 ```python
 from diffusers import StableDiffusionXLInstructPix2PixPipeline
@@ -117,6 +129,8 @@ pipe = StableDiffusionXLInstructPix2PixPipeline.from_single_file(ckpt_path, conf
 
 ```
 
+### Setting a model configuration option
+
 ```python
 from diffusers import UNet2DConditionModel
 
@@ -125,10 +139,6 @@ model = UNet2DConditionModel.from_single_file(ckpt_path, upcast_attention=True)
 
 ```
 
-In the example above, since we explicitly passed `repo_id="segmind/SSD-1B"`, it will use this [configuration file](https://huggingface.co/segmind/SSD-1B/blob/main/unet/config.json) from the "unet" subfolder in `"segmind/SSD-1B"` to configure the unet component included in the checkpoint; Similarly, it will use the `config.json` file from `"vae"` subfolder to configure the vae model, `config.json` file from text_encoder folder to configure text_encoder and so on.
-
-Note that most of the time you do not need to explicitly a `config` argument, `from_single_file` will automatically map the checkpoint to a repo id (we will discuss this in more details in next section). However, this can be useful in cases where model components might have been changed from what was originally distributed or in cases where a checkpoint file might not have the necessary metadata to correctly determine the configuration to use for the pipeline.
-
 <Tip>
 
 To learn more about how to load single file weights, see the [Load different Stable Diffusion formats](../../using-diffusers/other-formats) loading guide.
@@ -137,9 +147,11 @@ To learn more about how to load single file weights, see the [Load different Sta
 
 ## Working with local files
 
-As of `diffusers>=0.28.0` the `from_single_file` method will attempt to configure a pipeline or model by first inferring the model type from the checkpoint file and then using the model type to determine the appropriate model repo configuration to use from the Hugging Face Hub. For example, any single file checkpoint based on the Stable Diffusion XL base model will use the [`stabilityai/stable-diffusion-xl-base-1.0`](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) model repo to configure the pipeline.
+As of `diffusers>=0.28.0` the `from_single_file` method will attempt to configure a pipeline or model by first inferring the model type from the keys in the checkpoint file. This inferred model type is then used to determine the appropriate model repository on the Hugging Face Hub to configure the model or pipeline.
 
-If you are working in an environment with restricted internet access, it is recommended to download the config files and checkpoints for the model to your preferred directory and pass the local paths to the `pretrained_model_link_or_path` and `config` arguments of the `from_single_file` method.
+For example, any single file checkpoint based on the Stable Diffusion XL base model will use the [`stabilityai/stable-diffusion-xl-base-1.0`](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) model repository to configure the pipeline.
+
+If you are working in an environment with restricted internet access, it is recommended that you download the config files and checkpoints for the model to your preferred directory and pass the local paths to the `pretrained_model_link_or_path` and `config` arguments of the `from_single_file` method.
 
 ```python
 from huggingface_hub import hf_hub_download, snapshot_download
@@ -211,13 +223,14 @@ pipe = StableDiffusionXLPipeline.from_single_file(my_local_checkpoint_path, conf
 ```
 
 <Tip>
-Disabling symlinking means that the `huggingface_hub` caching mechanism has no way to determine whether a file has already been downloaded to the local directory. This means that the `hf_hub_download` and `snapshot_download` functions will download files to the local directory each time they are executed. If you are disabling symlinking, it is recommended that you separate the model download and loading steps to avoid downloading the same file multiple times.
+
+As of `huggingface_hub>=0.23.0` the `local_dir_use_symlinks` argument isn't necessary for the `hf_hub_download` and `snapshot_download` functions.
 
 </Tip>
 
 ## Using the original configuration file of a model
 
-If you would like to configure the parameters of the model components in the pipeline using the orignal YAML configuration file, you can pass a local path or url to the original configuration file to the `original_config` argument of the `from_single_file` method.
+If you would like to configure the model components in a pipeline using the orignal YAML configuration file, you can pass a local path or url to the original configuration file via the `original_config` argument.
 
 ```python
 from diffusers import StableDiffusionXLPipeline
@@ -229,13 +242,12 @@ original_config = "https://raw.githubusercontent.com/Stability-AI/generative-mod
 pipe = StableDiffusionXLPipeline.from_single_file(ckpt_path, original_config=original_config)
 ```
 
-In the example above, the `original_config` file is only used to configure the parameters of the individual model components of the pipeline. For example it will be used to configure parameters such as the `in_channels` of the `vae` model and `unet` model. It is not used to determine the type of component objects in the pipeline.
-
-
 <Tip>
-When using `original_config` with local_files_only=True`, Diffusers will attempt to infer the components based on the type signatures of pipeline class, rather than attempting to fetch the pipeline config from the Hugging Face Hub. This is to prevent backwards breaking changes in existing code that might not be able to connect to the internet to fetch the necessary pipeline config files.
 
-This is not as reliable as providing a path to a local config repo and might lead to errors when configuring the pipeline. To avoid this, please run the pipeline with `local_files_only=False` once to download the appropriate pipeline config files to the local cache.
+When using `original_config` with `local_files_only=True`, Diffusers will attempt to infer the components of the pipeline based on the type signatures of pipeline class, rather than attempting to fetch the configuration files from a model repository on the Hugging Face Hub. This is to prevent backward breaking changes in existing code that might not be able to connect to the internet to fetch the necessary configuration files.
+
+This is not as reliable as providing a path to a local model repository using the `config` argument and might lead to errors when configuring the pipeline. To avoid this, please run the pipeline with `local_files_only=False` once to download the appropriate pipeline configuration files to the local cache.
+
 </Tip>
 
 

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

@@ -10,7 +10,7 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# ControlNet
+# ControlNetModel
 
 The ControlNet model was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, Maneesh Agrawala. It provides a greater degree of control over text-to-image generation by conditioning the model on additional inputs such as edge maps, depth maps, segmentation maps, and keypoints for pose detection.
 

docs/source/en/api/models/dit_transformer2d.md

@@ -0,0 +1,19 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# DiTTransformer2DModel
+
+A Transformer model for image-like data from [DiT](https://huggingface.co/papers/2212.09748).
+
+## DiTTransformer2DModel
+
+[[autodoc]] DiTTransformer2DModel

docs/source/en/api/models/hunyuan_transformer2d.md

@@ -0,0 +1,20 @@
+<!--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.
+-->
+
+# HunyuanDiT2DModel
+
+A Diffusion Transformer model for 2D data from [Hunyuan-DiT](https://github.com/Tencent/HunyuanDiT).
+
+## HunyuanDiT2DModel
+
+[[autodoc]] HunyuanDiT2DModel
+

docs/source/en/api/models/pixart_transformer2d.md

@@ -0,0 +1,19 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# PixArtTransformer2DModel
+
+A Transformer model for image-like data from [PixArt-Alpha](https://huggingface.co/papers/2310.00426) and [PixArt-Sigma](https://huggingface.co/papers/2403.04692). 
+
+## PixArtTransformer2DModel
+
+[[autodoc]] PixArtTransformer2DModel

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

@@ -10,7 +10,7 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Prior Transformer
+# PriorTransformer
 
 The Prior Transformer was originally introduced in [Hierarchical Text-Conditional Image Generation with CLIP Latents](https://huggingface.co/papers/2204.06125) by Ramesh et al. It is used to predict CLIP image embeddings from CLIP text embeddings; image embeddings are predicted through a denoising diffusion process.
 

docs/source/en/api/models/transformer2d.md

@@ -10,7 +10,7 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Transformer2D
+# Transformer2DModel
 
 A Transformer model for image-like data from [CompVis](https://huggingface.co/CompVis) that is based on the [Vision Transformer](https://huggingface.co/papers/2010.11929) introduced by Dosovitskiy et al. The [`Transformer2DModel`] accepts discrete (classes of vector embeddings) or continuous (actual embeddings) inputs.
 

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

@@ -10,7 +10,7 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Transformer Temporal
+# TransformerTemporalModel
 
 A Transformer model for video-like data.
 

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

@@ -24,4 +24,4 @@ The abstract from the paper is:
 
 ## VQEncoderOutput
 
-[[autodoc]] models.vq_model.VQEncoderOutput
+[[autodoc]] models.autoencoders.vq_model.VQEncoderOutput

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

@@ -0,0 +1,41 @@
+<!--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.
+-->
+
+# Hunyuan-DiT
+![chinese elements understanding](https://github.com/gnobitab/diffusers-hunyuan/assets/1157982/39b99036-c3cb-4f16-bb1a-40ec25eda573)
+
+[Hunyuan-DiT : A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding](https://arxiv.org/abs/2405.08748) from Tencent Hunyuan.
+
+The abstract from the paper is:
+
+*We present Hunyuan-DiT, a text-to-image diffusion transformer with fine-grained understanding of both English and Chinese. To construct Hunyuan-DiT, we carefully design the transformer structure, text encoder, and positional encoding. We also build from scratch a whole data pipeline to update and evaluate data for iterative model optimization. For fine-grained language understanding, we train a Multimodal Large Language Model to refine the captions of the images. Finally, Hunyuan-DiT can perform multi-turn multimodal dialogue with users, generating and refining images according to the context. Through our holistic human evaluation protocol with more than 50 professional human evaluators, Hunyuan-DiT sets a new state-of-the-art in Chinese-to-image generation compared with other open-source models.*
+
+
+You can find the original codebase at [Tencent/HunyuanDiT](https://github.com/Tencent/HunyuanDiT) and all the available checkpoints at [Tencent-Hunyuan](https://huggingface.co/Tencent-Hunyuan/HunyuanDiT).
+
+**Highlights**: HunyuanDiT supports Chinese/English-to-image, multi-resolution generation.
+
+HunyuanDiT has the following components:
+* It uses a diffusion transformer as the backbone
+* It combines two text encoders, a bilingual CLIP and a multilingual T5 encoder
+
+
+## Memory optimization
+
+By loading the T5 text encoder in 8 bits, you can run the pipeline in just under 6 GBs of GPU VRAM. Refer to [this script](https://gist.github.com/sayakpaul/3154605f6af05b98a41081aaba5ca43e) for details. 
+
+## HunyuanDiTPipeline
+
+[[autodoc]] HunyuanDiTPipeline
+	- all
+	- __call__
+	

docs/source/en/api/pipelines/i2vgenxl.md

@@ -47,6 +47,7 @@ Sample output with I2VGenXL:
 * Unlike SVD, it additionally accepts text prompts as inputs.
 * It can generate higher resolution videos.
 * When using the [`DDIMScheduler`] (which is default for this pipeline), less than 50 steps for inference leads to bad results.
+* This implementation is 1-stage variant of I2VGenXL. The main figure in the [I2VGen-XL](https://arxiv.org/abs/2311.04145) paper shows a 2-stage variant, however, 1-stage variant works well. See [this discussion](https://github.com/huggingface/diffusers/discussions/7952) for more details.
 
 ## I2VGenXLPipeline
 [[autodoc]] I2VGenXLPipeline

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

@@ -0,0 +1,76 @@
+<!--Copyright 2024 Marigold authors 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.
+-->
+
+# Marigold Pipelines for Computer Vision Tasks
+
+![marigold](https://marigoldmonodepth.github.io/images/teaser_collage_compressed.jpg)
+
+Marigold was proposed in [Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation](https://huggingface.co/papers/2312.02145), a CVPR 2024 Oral paper by [Bingxin Ke](http://www.kebingxin.com/), [Anton Obukhov](https://www.obukhov.ai/), [Shengyu Huang](https://shengyuh.github.io/), [Nando Metzger](https://nandometzger.github.io/), [Rodrigo Caye Daudt](https://rcdaudt.github.io/), and [Konrad Schindler](https://scholar.google.com/citations?user=FZuNgqIAAAAJ&hl=en). 
+The idea is to repurpose the rich generative prior of Text-to-Image Latent Diffusion Models (LDMs) for traditional computer vision tasks. 
+Initially, this idea was explored to fine-tune Stable Diffusion for Monocular Depth Estimation, as shown in the teaser above. 
+Later, 
+- [Tianfu Wang](https://tianfwang.github.io/) trained the first Latent Consistency Model (LCM) of Marigold, which unlocked fast single-step inference;
+- [Kevin Qu](https://www.linkedin.com/in/kevin-qu-b3417621b/?locale=en_US) extended the approach to Surface Normals Estimation;
+- [Anton Obukhov](https://www.obukhov.ai/) contributed the pipelines and documentation into diffusers (enabled and supported by [YiYi Xu](https://yiyixuxu.github.io/) and [Sayak Paul](https://sayak.dev/)).
+
+The abstract from the paper is:
+
+*Monocular depth estimation is a fundamental computer vision task. Recovering 3D depth from a single image is geometrically ill-posed and requires scene understanding, so it is not surprising that the rise of deep learning has led to a breakthrough. The impressive progress of monocular depth estimators has mirrored the growth in model capacity, from relatively modest CNNs to large Transformer architectures. Still, monocular depth estimators tend to struggle when presented with images with unfamiliar content and layout, since their knowledge of the visual world is restricted by the data seen during training, and challenged by zero-shot generalization to new domains. This motivates us to explore whether the extensive priors captured in recent generative diffusion models can enable better, more generalizable depth estimation. We introduce Marigold, a method for affine-invariant monocular depth estimation that is derived from Stable Diffusion and retains its rich prior knowledge. The estimator can be fine-tuned in a couple of days on a single GPU using only synthetic training data. It delivers state-of-the-art performance across a wide range of datasets, including over 20% performance gains in specific cases. Project page: https://marigoldmonodepth.github.io.*
+
+## Available Pipelines
+
+Each pipeline supports one Computer Vision task, which takes an input RGB image as input and produces a *prediction* of the modality of interest, such as a depth map of the input image. 
+Currently, the following tasks are implemented:
+
+| Pipeline                                                                                                                                    | Predicted Modalities                                                                                             |                                                                       Demos                                                                        |
+|---------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------|:--------------------------------------------------------------------------------------------------------------------------------------------------:|
+| [MarigoldDepthPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_depth.py)     | [Depth](https://en.wikipedia.org/wiki/Depth_map), [Disparity](https://en.wikipedia.org/wiki/Binocular_disparity) | [Fast Demo (LCM)](https://huggingface.co/spaces/prs-eth/marigold-lcm), [Slow Original Demo (DDIM)](https://huggingface.co/spaces/prs-eth/marigold) |
+| [MarigoldNormalsPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_normals.py) | [Surface normals](https://en.wikipedia.org/wiki/Normal_mapping)                                                  |                                   [Fast Demo (LCM)](https://huggingface.co/spaces/prs-eth/marigold-normals-lcm)                                    |
+
+
+## Available Checkpoints
+
+The original checkpoints can be found under the [PRS-ETH](https://huggingface.co/prs-eth/) Hugging Face organization. 
+
+<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-components-across-pipelines) section to learn how to efficiently load the same components into multiple pipelines. Also, to know more about reducing the memory usage of this pipeline, refer to the ["Reduce memory usage"] section [here](../../using-diffusers/svd#reduce-memory-usage).
+
+</Tip>
+
+<Tip warning={true}>
+
+Marigold pipelines were designed and tested only with `DDIMScheduler` and `LCMScheduler`. 
+Depending on the scheduler, the number of inference steps required to get reliable predictions varies, and there is no universal value that works best across schedulers.
+Because of that, the default value of `num_inference_steps` in the `__call__` method of the pipeline is set to `None` (see the API reference). 
+Unless set explicitly, its value will be taken from the checkpoint configuration `model_index.json`. 
+This is done to ensure high-quality predictions when calling the pipeline with just the `image` argument. 
+
+</Tip>
+
+See also Marigold [usage examples](marigold_usage).
+
+## MarigoldDepthPipeline
+[[autodoc]] MarigoldDepthPipeline
+	- all
+	- __call__
+
+## MarigoldNormalsPipeline
+[[autodoc]] MarigoldNormalsPipeline
+	- all
+	- __call__
+
+## MarigoldDepthOutput
+[[autodoc]] pipelines.marigold.pipeline_marigold_depth.MarigoldDepthOutput
+
+## MarigoldNormalsOutput
+[[autodoc]] pipelines.marigold.pipeline_marigold_normals.MarigoldNormalsOutput

docs/source/en/conceptual/philosophy.md

@@ -70,7 +70,7 @@ The following design principles are followed:
 - Pipelines should be used **only** for inference.
 - Pipelines should be very readable, self-explanatory, and easy to tweak.
 - Pipelines should be designed to build on top of each other and be easy to integrate into higher-level APIs.
-- Pipelines are **not** intended to be feature-complete user interfaces. For future complete user interfaces one should rather have a look at [InvokeAI](https://github.com/invoke-ai/InvokeAI), [Diffuzers](https://github.com/abhishekkrthakur/diffuzers), and [lama-cleaner](https://github.com/Sanster/lama-cleaner).
+- Pipelines are **not** intended to be feature-complete user interfaces. For feature-complete user interfaces one should rather have a look at [InvokeAI](https://github.com/invoke-ai/InvokeAI), [Diffuzers](https://github.com/abhishekkrthakur/diffuzers), and [lama-cleaner](https://github.com/Sanster/lama-cleaner).
 - Every pipeline should have one and only one way to run it via a `__call__` method. The naming of the `__call__` arguments should be shared across all pipelines.
 - Pipelines should be named after the task they are intended to solve.
 - In almost all cases, novel diffusion pipelines shall be implemented in a new pipeline folder/file.

docs/source/en/optimization/tgate.md

@@ -6,7 +6,7 @@ Before you begin, make sure you install T-GATE.
 
 ```bash
 pip install tgate
-pip install -U pytorch diffusers transformers accelerate DeepCache
+pip install -U torch diffusers transformers accelerate DeepCache
 ```
 
 
@@ -46,12 +46,12 @@ pipe = TgatePixArtLoader(
 
 image = pipe.tgate(
        "An alpaca made of colorful building blocks, cyberpunk.",
-        gate_step=gate_step,
+       gate_step=gate_step,
        num_inference_steps=inference_step,
 ).images[0]
 ```
 </hfoption>
-<hfoption id="Stable Diffusion XL"> 
+<hfoption id="Stable Diffusion XL">
 
 Accelerate `StableDiffusionXLPipeline` with T-GATE:
 
@@ -78,9 +78,9 @@ pipe = TgateSDXLLoader(
 ).to("cuda")
 
 image = pipe.tgate(
-        "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k.",
-        gate_step=gate_step,
-        num_inference_steps=inference_step
+       "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k.",
+       gate_step=gate_step,
+       num_inference_steps=inference_step
 ).images[0]
 ```
 </hfoption>
@@ -111,9 +111,9 @@ pipe = TgateSDXLDeepCacheLoader(
 ).to("cuda")
 
 image = pipe.tgate(
-        "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k.",
-        gate_step=gate_step,
-        num_inference_steps=inference_step
+       "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k.",
+       gate_step=gate_step,
+       num_inference_steps=inference_step
 ).images[0]
 ```
 </hfoption>
@@ -151,9 +151,9 @@ pipe = TgateSDXLLoader(
 ).to("cuda")
 
 image = pipe.tgate(
-        "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k.",
-        gate_step=gate_step,
-        num_inference_steps=inference_step
+       "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k.",
+       gate_step=gate_step,
+       num_inference_steps=inference_step
 ).images[0]
 ```
 </hfoption>

docs/source/en/training/dreambooth.md

@@ -440,6 +440,198 @@ Stable Diffusion XL (SDXL) is a powerful text-to-image model that generates high
 
 The SDXL training script is discussed in more detail in the [SDXL training](sdxl) guide.
 
+## DeepFloyd IF
+
+DeepFloyd IF is a cascading pixel diffusion model with three stages. The first stage generates a base image and the second and third stages progressively upscales the base image into a high-resolution 1024x1024 image. Use the [train_dreambooth_lora.py](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora.py) or [train_dreambooth.py](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth.py) scripts to train a DeepFloyd IF model with LoRA or the full model.
+
+DeepFloyd IF uses predicted variance, but the Diffusers training scripts uses predicted error so the trained DeepFloyd IF models are switched to a fixed variance schedule. The training scripts will update the scheduler config of the fully trained model for you. However, when you load the saved LoRA weights you must also update the pipeline's scheduler config.
+
+```py
+from diffusers import DiffusionPipeline
+
+pipe = DiffusionPipeline.from_pretrained("DeepFloyd/IF-I-XL-v1.0", use_safetensors=True)
+
+pipe.load_lora_weights("<lora weights path>")
+
+# Update scheduler config to fixed variance schedule
+pipe.scheduler = pipe.scheduler.__class__.from_config(pipe.scheduler.config, variance_type="fixed_small")
+```
+
+The stage 2 model requires additional validation images to upscale. You can download and use a downsized version of the training images for this.
+
+```py
+from huggingface_hub import snapshot_download
+
+local_dir = "./dog_downsized"
+snapshot_download(
+    "diffusers/dog-example-downsized",
+    local_dir=local_dir,
+    repo_type="dataset",
+    ignore_patterns=".gitattributes",
+)
+```
+
+The code samples below provide a brief overview of how to train a DeepFloyd IF model with a combination of DreamBooth and LoRA. Some important parameters to note are:
+
+* `--resolution=64`, a much smaller resolution is required because DeepFloyd IF is a pixel diffusion model and to work on uncompressed pixels, the input images must be smaller
+* `--pre_compute_text_embeddings`, compute the text embeddings ahead of time to save memory because the [`~transformers.T5Model`] can take up a lot of memory
+* `--tokenizer_max_length=77`, you can use a longer default text length with T5 as the text encoder but the default model encoding procedure uses a shorter text length
+* `--text_encoder_use_attention_mask`, to pass the attention mask to the text encoder
+
+<hfoptions id="IF-DreamBooth">
+<hfoption id="Stage 1 LoRA DreamBooth">
+
+Training stage 1 of DeepFloyd IF with LoRA and DreamBooth requires ~28GB of memory.
+
+```bash
+export MODEL_NAME="DeepFloyd/IF-I-XL-v1.0"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="dreambooth_dog_lora"
+
+accelerate launch train_dreambooth_lora.py \
+  --report_to wandb \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --instance_prompt="a sks dog" \
+  --resolution=64 \
+  --train_batch_size=4 \
+  --gradient_accumulation_steps=1 \
+  --learning_rate=5e-6 \
+  --scale_lr \
+  --max_train_steps=1200 \
+  --validation_prompt="a sks dog" \
+  --validation_epochs=25 \
+  --checkpointing_steps=100 \
+  --pre_compute_text_embeddings \
+  --tokenizer_max_length=77 \
+  --text_encoder_use_attention_mask
+```
+
+</hfoption>
+<hfoption id="Stage 2 LoRA DreamBooth">
+
+For stage 2 of DeepFloyd IF with LoRA and DreamBooth, pay attention to these parameters:
+
+* `--validation_images`, the images to upscale during validation
+* `--class_labels_conditioning=timesteps`, to additionally conditional the UNet as needed in stage 2
+* `--learning_rate=1e-6`, a lower learning rate is used compared to stage 1
+* `--resolution=256`, the expected resolution for the upscaler
+
+```bash
+export MODEL_NAME="DeepFloyd/IF-II-L-v1.0"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="dreambooth_dog_upscale"
+export VALIDATION_IMAGES="dog_downsized/image_1.png dog_downsized/image_2.png dog_downsized/image_3.png dog_downsized/image_4.png"
+
+python train_dreambooth_lora.py \
+    --report_to wandb \
+    --pretrained_model_name_or_path=$MODEL_NAME \
+    --instance_data_dir=$INSTANCE_DIR \
+    --output_dir=$OUTPUT_DIR \
+    --instance_prompt="a sks dog" \
+    --resolution=256 \
+    --train_batch_size=4 \
+    --gradient_accumulation_steps=1 \
+    --learning_rate=1e-6 \ 
+    --max_train_steps=2000 \
+    --validation_prompt="a sks dog" \
+    --validation_epochs=100 \
+    --checkpointing_steps=500 \
+    --pre_compute_text_embeddings \
+    --tokenizer_max_length=77 \
+    --text_encoder_use_attention_mask \
+    --validation_images $VALIDATION_IMAGES \
+    --class_labels_conditioning=timesteps
+```
+
+</hfoption>
+<hfoption id="Stage 1 DreamBooth">
+
+For stage 1 of DeepFloyd IF with DreamBooth, pay attention to these parameters:
+
+* `--skip_save_text_encoder`, to skip saving the full T5 text encoder with the finetuned model
+* `--use_8bit_adam`, to use 8-bit Adam optimizer to save memory due to the size of the optimizer state when training the full model
+* `--learning_rate=1e-7`, a really low learning rate should be used for full model training otherwise the model quality is degraded (you can use a higher learning rate with a larger batch size)
+
+Training with 8-bit Adam and a batch size of 4, the full model can be trained with ~48GB of memory.
+
+```bash
+export MODEL_NAME="DeepFloyd/IF-I-XL-v1.0"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="dreambooth_if"
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=64 \
+  --train_batch_size=4 \
+  --gradient_accumulation_steps=1 \
+  --learning_rate=1e-7 \
+  --max_train_steps=150 \
+  --validation_prompt "a photo of sks dog" \
+  --validation_steps 25 \
+  --text_encoder_use_attention_mask \
+  --tokenizer_max_length 77 \
+  --pre_compute_text_embeddings \
+  --use_8bit_adam \
+  --set_grads_to_none \
+  --skip_save_text_encoder \
+  --push_to_hub
+```
+
+</hfoption>
+<hfoption id="Stage 2 DreamBooth">
+
+For stage 2 of DeepFloyd IF with DreamBooth, pay attention to these parameters:
+
+* `--learning_rate=5e-6`, use a lower learning rate with a smaller effective batch size
+* `--resolution=256`, the expected resolution for the upscaler
+* `--train_batch_size=2` and `--gradient_accumulation_steps=6`, to effectively train on images wiht faces requires larger batch sizes
+
+```bash
+export MODEL_NAME="DeepFloyd/IF-II-L-v1.0"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="dreambooth_dog_upscale"
+export VALIDATION_IMAGES="dog_downsized/image_1.png dog_downsized/image_2.png dog_downsized/image_3.png dog_downsized/image_4.png"
+
+accelerate launch train_dreambooth.py \
+  --report_to wandb \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --instance_prompt="a sks dog" \
+  --resolution=256 \
+  --train_batch_size=2 \
+  --gradient_accumulation_steps=6 \
+  --learning_rate=5e-6 \
+  --max_train_steps=2000 \
+  --validation_prompt="a sks dog" \
+  --validation_steps=150 \
+  --checkpointing_steps=500 \
+  --pre_compute_text_embeddings \
+  --tokenizer_max_length=77 \
+  --text_encoder_use_attention_mask \
+  --validation_images $VALIDATION_IMAGES \
+  --class_labels_conditioning timesteps \
+  --push_to_hub
+```
+
+</hfoption>
+</hfoptions>
+
+### Training tips
+
+Training the DeepFloyd IF model can be challenging, but here are some tips that we've found helpful:
+
+- LoRA is sufficient for training the stage 1 model because the model's low resolution makes representing finer details difficult regardless.
+- For common or simple objects, you don't necessarily need to finetune the upscaler. Make sure the prompt passed to the upscaler is adjusted to remove the new token from the instance prompt. For example, if your stage 1 prompt is "a sks dog" then your stage 2 prompt should be "a dog".
+- For finer details like faces, fully training the stage 2 upscaler is better than training the stage 2 model with LoRA. It also helps to use lower learning rates with larger batch sizes.
+- Lower learning rates should be used to train the stage 2 model.
+- The [`DDPMScheduler`] works better than the DPMSolver used in the training scripts.
+
 ## Next steps
 
 Congratulations on training your DreamBooth model! To learn more about how to use your new model, the following guide may be helpful:

docs/source/en/tutorials/basic_training.md

@@ -260,7 +260,7 @@ Then, you'll need a way to evaluate the model. For evaluation, you can use the [
 ...     # The default pipeline output type is `List[PIL.Image]`
 ...     images = pipeline(
 ...         batch_size=config.eval_batch_size,
-...         generator=torch.manual_seed(config.seed),
+...         generator=torch.Generator(device='cpu').manual_seed(config.seed), # Use a separate torch generator to avoid rewinding the random state of the main training loop
 ...     ).images
 
 ...     # Make a grid out of the images

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

@@ -12,54 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Controlling image quality
 
-The components of a diffusion model, like the UNet and scheduler, can be optimized to improve the quality of generated images leading to better image lighting and details. These techniques are especially useful if you don't have the resources to simply use a larger model for inference. You can enable these techniques during inference without any additional training.
+The components of a diffusion model, like the UNet and scheduler, can be optimized to improve the quality of generated images leading to better details. These techniques are especially useful if you don't have the resources to simply use a larger model for inference. You can enable these techniques during inference without any additional training.
 
 This guide will show you how to turn these techniques on in your pipeline and how to configure them to improve the quality of your generated images.
 
-## Lighting
-
-The Stable Diffusion models aren't very good at generating images that are very bright or dark because the scheduler doesn't start sampling from the last timestep and it doesn't enforce a zero signal-to-noise ratio (SNR). The [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://hf.co/papers/2305.08891) paper fixes these issues which are now available in some Diffusers schedulers.
-
-> [!TIP]
-> For inference, you need a model that has been trained with *v_prediction*. To train your own model with *v_prediction*, add the following flag to the [train_text_to_image.py](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py) or [train_text_to_image_lora.py](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_lora.py) scripts.
->
-> ```bash
-> --prediction_type="v_prediction"
-> ```
-
-For example, load the [ptx0/pseudo-journey-v2](https://hf.co/ptx0/pseudo-journey-v2) checkpoint which was trained with `v_prediction` and the [`DDIMScheduler`]. Now you should configure the following parameters in the [`DDIMScheduler`].
-
-* `rescale_betas_zero_snr=True` to rescale the noise schedule to zero SNR
-* `timestep_spacing="trailing"` to start sampling from the last timestep
-
-Set `guidance_rescale` in the pipeline to prevent over-exposure. A lower value increases brightness but some of the details may appear washed out.
-
-```py
-from diffusers import DiffusionPipeline, DDIMScheduler
-
-pipeline = DiffusionPipeline.from_pretrained("ptx0/pseudo-journey-v2", use_safetensors=True)
-
-pipeline.scheduler = DDIMScheduler.from_config(
-    pipeline.scheduler.config, rescale_betas_zero_snr=True, timestep_spacing="trailing"
-)
-pipeline.to("cuda")
-prompt = "cinematic photo of a snowy mountain at night with the northern lights aurora borealis overhead, 35mm photograph, film, professional, 4k, highly detailed"
-generator = torch.Generator(device="cpu").manual_seed(23)
-image = pipeline(prompt, guidance_rescale=0.7, generator=generator).images[0]
-image
-```
-
-<div class="flex gap-4">
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/no-zero-snr.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">default Stable Diffusion v2-1 image</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/zero-snr.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">image with zero SNR and trailing timestep spacing enabled</figcaption>
-  </div>
-</div>
-
 ## Details
 
 [FreeU](https://hf.co/papers/2309.11497) improves image details by rebalancing the UNet's backbone and skip connection weights. The skip connections can cause the model to overlook some of the backbone semantics which may lead to unnatural image details in the generated image. This technique does not require any additional training and can be applied on the fly during inference for tasks like image-to-image and text-to-video.

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

@@ -78,7 +78,7 @@ image = pipe(
     prompt=prompt,
     num_inference_steps=4,
     guidance_scale=0,
-    eta=0.3, 
+    eta=0.3,
     generator=torch.Generator(device=device).manual_seed(0),
 ).images[0]
 ```
@@ -156,14 +156,14 @@ image = pipe(
     prompt=prompt,
     num_inference_steps=8,
     guidance_scale=0,
-    eta=0.3, 
+    eta=0.3,
     generator=torch.Generator(device=device).manual_seed(0),
 ).images[0]
 ```
 
 ![](https://github.com/jabir-zheng/TCD/raw/main/assets/animagine_xl.png)
 
-TCD-LoRA also supports other LoRAs trained on different styles. For example, let's load the [TheLastBen/Papercut_SDXL](https://huggingface.co/TheLastBen/Papercut_SDXL) LoRA and fuse it with the TCD-LoRA with the [`~loaders.UNet2DConditionLoadersMixin.set_adapters`] method. 
+TCD-LoRA also supports other LoRAs trained on different styles. For example, let's load the [TheLastBen/Papercut_SDXL](https://huggingface.co/TheLastBen/Papercut_SDXL) LoRA and fuse it with the TCD-LoRA with the [`~loaders.UNet2DConditionLoadersMixin.set_adapters`] method.
 
 > [!TIP]
 > Check out the [Merge LoRAs](merge_loras) guide to learn more about efficient merging methods.
@@ -171,7 +171,7 @@ TCD-LoRA also supports other LoRAs trained on different styles. For example, let
 ```python
 import torch
 from diffusers import StableDiffusionXLPipeline
-from scheduling_tcd import TCDScheduler 
+from scheduling_tcd import TCDScheduler
 
 device = "cuda"
 base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
@@ -191,7 +191,7 @@ image = pipe(
     prompt=prompt,
     num_inference_steps=4,
     guidance_scale=0,
-    eta=0.3, 
+    eta=0.3,
     generator=torch.Generator(device=device).manual_seed(0),
 ).images[0]
 ```
@@ -215,7 +215,7 @@ from PIL import Image
 from transformers import DPTFeatureExtractor, DPTForDepthEstimation
 from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline
 from diffusers.utils import load_image, make_image_grid
-from scheduling_tcd import TCDScheduler 
+from scheduling_tcd import TCDScheduler
 
 device = "cuda"
 depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to(device)
@@ -249,13 +249,13 @@ controlnet = ControlNetModel.from_pretrained(
     controlnet_id,
     torch_dtype=torch.float16,
     variant="fp16",
-).to(device)
+)
 pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
     base_model_id,
     controlnet=controlnet,
     torch_dtype=torch.float16,
     variant="fp16",
-).to(device)
+)
 pipe.enable_model_cpu_offload()
 
 pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
@@ -271,9 +271,9 @@ depth_image = get_depth_map(image)
 controlnet_conditioning_scale = 0.5  # recommended for good generalization
 
 image = pipe(
-    prompt, 
-    image=depth_image, 
-    num_inference_steps=4, 
+    prompt,
+    image=depth_image,
+    num_inference_steps=4,
     guidance_scale=0,
     eta=0.3,
     controlnet_conditioning_scale=controlnet_conditioning_scale,
@@ -290,7 +290,7 @@ grid_image = make_image_grid([depth_image, image], rows=1, cols=2)
 import torch
 from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline
 from diffusers.utils import load_image, make_image_grid
-from scheduling_tcd import TCDScheduler 
+from scheduling_tcd import TCDScheduler
 
 device = "cuda"
 base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
@@ -301,13 +301,13 @@ controlnet = ControlNetModel.from_pretrained(
     controlnet_id,
     torch_dtype=torch.float16,
     variant="fp16",
-).to(device)
+)
 pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
     base_model_id,
     controlnet=controlnet,
     torch_dtype=torch.float16,
     variant="fp16",
-).to(device)
+)
 pipe.enable_model_cpu_offload()
 
 pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
@@ -322,9 +322,9 @@ canny_image = load_image("https://huggingface.co/datasets/hf-internal-testing/di
 controlnet_conditioning_scale = 0.5  # recommended for good generalization
 
 image = pipe(
-    prompt, 
-    image=canny_image, 
-    num_inference_steps=4, 
+    prompt,
+    image=canny_image,
+    num_inference_steps=4,
     guidance_scale=0,
     eta=0.3,
     controlnet_conditioning_scale=controlnet_conditioning_scale,
@@ -336,7 +336,7 @@ grid_image = make_image_grid([canny_image, image], rows=1, cols=2)
 ![](https://github.com/jabir-zheng/TCD/raw/main/assets/controlnet_canny_tcd.png)
 
 <Tip>
-The inference parameters in this example might not work for all examples, so we recommend you to try different values for `num_inference_steps`, `guidance_scale`, `controlnet_conditioning_scale` and `cross_attention_kwargs` parameters and choose the best one. 
+The inference parameters in this example might not work for all examples, so we recommend you to try different values for `num_inference_steps`, `guidance_scale`, `controlnet_conditioning_scale` and `cross_attention_kwargs` parameters and choose the best one.
 </Tip>
 
 </hfoption>
@@ -350,7 +350,7 @@ from diffusers import StableDiffusionXLPipeline
 from diffusers.utils import load_image, make_image_grid
 
 from ip_adapter import IPAdapterXL
-from scheduling_tcd import TCDScheduler 
+from scheduling_tcd import TCDScheduler
 
 device = "cuda"
 base_model_path = "stabilityai/stable-diffusion-xl-base-1.0"
@@ -359,8 +359,8 @@ ip_ckpt = "sdxl_models/ip-adapter_sdxl.bin"
 tcd_lora_id = "h1t/TCD-SDXL-LoRA"
 
 pipe = StableDiffusionXLPipeline.from_pretrained(
-    base_model_path, 
-    torch_dtype=torch.float16, 
+    base_model_path,
+    torch_dtype=torch.float16,
     variant="fp16"
 )
 pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
@@ -375,13 +375,13 @@ ref_image = load_image("https://raw.githubusercontent.com/tencent-ailab/IP-Adapt
 prompt = "best quality, high quality, wearing sunglasses"
 
 image = ip_model.generate(
-    pil_image=ref_image, 
+    pil_image=ref_image,
     prompt=prompt,
     scale=0.5,
-    num_samples=1, 
-    num_inference_steps=4, 
+    num_samples=1,
+    num_inference_steps=4,
     guidance_scale=0,
-    eta=0.3, 
+    eta=0.3,
     seed=0,
 )[0]
 

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

@@ -230,7 +230,7 @@ from diffusers.utils import load_image, make_image_grid
 
 pipeline = AutoPipelineForInpainting.from_pretrained(
     "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, variant="fp16"
-).to("cuda")
+)
 pipeline.enable_model_cpu_offload()
 # remove following line if xFormers is not installed or you have PyTorch 2.0 or higher installed
 pipeline.enable_xformers_memory_efficient_attention()
@@ -255,7 +255,7 @@ from diffusers.utils import load_image, make_image_grid
 
 pipeline = AutoPipelineForInpainting.from_pretrained(
     "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16, variant="fp16"
-).to("cuda")
+)
 pipeline.enable_model_cpu_offload()
 # remove following line if xFormers is not installed or you have PyTorch 2.0 or higher installed
 pipeline.enable_xformers_memory_efficient_attention()
@@ -296,7 +296,7 @@ from diffusers.utils import load_image, make_image_grid
 
 pipeline = AutoPipelineForInpainting.from_pretrained(
     "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, variant="fp16"
-).to("cuda")
+)
 pipeline.enable_model_cpu_offload()
 # remove following line if xFormers is not installed or you have PyTorch 2.0 or higher installed
 pipeline.enable_xformers_memory_efficient_attention()
@@ -319,7 +319,7 @@ from diffusers.utils import load_image, make_image_grid
 
 pipeline = AutoPipelineForInpainting.from_pretrained(
     "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16, variant="fp16"
-).to("cuda")
+)
 pipeline.enable_model_cpu_offload()
 # remove following line if xFormers is not installed or you have PyTorch 2.0 or higher installed
 pipeline.enable_xformers_memory_efficient_attention()

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

@@ -0,0 +1,466 @@
+<!--Copyright 2024 Marigold authors 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.
+-->
+
+# Marigold Pipelines for Computer Vision Tasks
+
+[Marigold](../api/pipelines/marigold) is a novel diffusion-based dense prediction approach, and a set of pipelines for various computer vision tasks, such as monocular depth estimation.
+
+This guide will show you how to use Marigold to obtain fast and high-quality predictions for images and videos.
+
+Each pipeline supports one Computer Vision task, which takes an input RGB image as input and produces a *prediction* of the modality of interest, such as a depth map of the input image.
+Currently, the following tasks are implemented:
+
+| Pipeline                                                                                                                                    | Predicted Modalities                                                                                             |                                                                       Demos                                                                        |
+|---------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------|:--------------------------------------------------------------------------------------------------------------------------------------------------:|
+| [MarigoldDepthPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_depth.py)     | [Depth](https://en.wikipedia.org/wiki/Depth_map), [Disparity](https://en.wikipedia.org/wiki/Binocular_disparity) | [Fast Demo (LCM)](https://huggingface.co/spaces/prs-eth/marigold-lcm), [Slow Original Demo (DDIM)](https://huggingface.co/spaces/prs-eth/marigold) |
+| [MarigoldNormalsPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_normals.py) | [Surface normals](https://en.wikipedia.org/wiki/Normal_mapping)                                                  |                                   [Fast Demo (LCM)](https://huggingface.co/spaces/prs-eth/marigold-normals-lcm)                                    |
+
+The original checkpoints can be found under the [PRS-ETH](https://huggingface.co/prs-eth/) Hugging Face organization.
+These checkpoints are meant to work with diffusers pipelines and the [original codebase](https://github.com/prs-eth/marigold).
+The original code can also be used to train new checkpoints.
+
+| Checkpoint                                                                                    | Modality | Comment                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                |
+|-----------------------------------------------------------------------------------------------|----------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| [prs-eth/marigold-v1-0](https://huggingface.co/prs-eth/marigold-v1-0)                         | Depth    | The first Marigold Depth checkpoint, which predicts *affine-invariant depth* maps. The performance of this checkpoint in benchmarks was studied in the original [paper](https://huggingface.co/papers/2312.02145). Designed to be used with the `DDIMScheduler` at inference, it requires at least 10 steps to get reliable predictions. Affine-invariant depth prediction has a range of values in each pixel between 0 (near plane) and 1 (far plane); both planes are chosen by the model as part of the inference process. See the `MarigoldImageProcessor` reference for visualization utilities. |
+| [prs-eth/marigold-depth-lcm-v1-0](https://huggingface.co/prs-eth/marigold-depth-lcm-v1-0)     | Depth    | The fast Marigold Depth checkpoint, fine-tuned from `prs-eth/marigold-v1-0`. Designed to be used with the `LCMScheduler` at inference, it requires as little as 1 step to get reliable predictions. The prediction reliability saturates at 4 steps and declines after that.                                                                                                                                                                                                                                                                                                                           |
+| [prs-eth/marigold-normals-v0-1](https://huggingface.co/prs-eth/marigold-normals-v0-1)         | Normals  | A preview checkpoint for the Marigold Normals pipeline. Designed to be used with the `DDIMScheduler` at inference, it requires at least 10 steps to get reliable predictions. The surface normals predictions are unit-length 3D vectors with values in the range from -1 to 1. *This checkpoint will be phased out after the release of `v1-0` version.*                                                                                                                                                                                                                                              |
+| [prs-eth/marigold-normals-lcm-v0-1](https://huggingface.co/prs-eth/marigold-normals-lcm-v0-1) | Normals  | The fast Marigold Normals checkpoint, fine-tuned from `prs-eth/marigold-normals-v0-1`. Designed to be used with the `LCMScheduler` at inference, it requires as little as 1 step to get reliable predictions. The prediction reliability saturates at 4 steps and declines after that. *This checkpoint will be phased out after the release of `v1-0` version.*                                                                                                                                                                                                                                       |
+The examples below are mostly given for depth prediction, but they can be universally applied with other supported modalities.
+We showcase the predictions using the same input image of Albert Einstein generated by Midjourney.
+This makes it easier to compare visualizations of the predictions across various modalities and checkpoints.
+
+<div class="flex gap-4" style="justify-content: center; width: 100%;">
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://marigoldmonodepth.github.io/images/einstein.jpg"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Example input image for all Marigold pipelines
+    </figcaption>
+  </div>
+</div>
+
+### Depth Prediction Quick Start
+
+To get the first depth prediction, load `prs-eth/marigold-depth-lcm-v1-0` checkpoint into `MarigoldDepthPipeline` pipeline, put the image through the pipeline, and save the predictions:
+
+```python
+import diffusers
+import torch
+
+pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
+    "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
+).to("cuda")
+
+image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+depth = pipe(image)
+
+vis = pipe.image_processor.visualize_depth(depth.prediction)
+vis[0].save("einstein_depth.png")
+
+depth_16bit = pipe.image_processor.export_depth_to_16bit_png(depth.prediction)
+depth_16bit[0].save("einstein_depth_16bit.png")
+```
+
+The visualization function for depth [`~pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_depth`] applies one of [matplotlib's colormaps](https://matplotlib.org/stable/users/explain/colors/colormaps.html) (`Spectral` by default) to map the predicted pixel values from a single-channel `[0, 1]` depth range into an RGB image.
+With the `Spectral` colormap, pixels with near depth are painted red, and far pixels are assigned blue color.
+The 16-bit PNG file stores the single channel values mapped linearly from the `[0, 1]` range into `[0, 65535]`.
+Below are the raw and the visualized predictions; as can be seen, dark areas (mustache) are easier to distinguish in the visualization:
+
+<div class="flex gap-4">
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_lcm_depth_16bit.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Predicted depth (16-bit PNG)
+    </figcaption>
+  </div>
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_lcm_depth.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Predicted depth visualization (Spectral)
+    </figcaption>
+  </div>
+</div>
+
+### Surface Normals Prediction Quick Start
+
+Load `prs-eth/marigold-normals-lcm-v0-1` checkpoint into `MarigoldNormalsPipeline` pipeline, put the image through the pipeline, and save the predictions:
+
+```python
+import diffusers
+import torch
+
+pipe = diffusers.MarigoldNormalsPipeline.from_pretrained(
+    "prs-eth/marigold-normals-lcm-v0-1", variant="fp16", torch_dtype=torch.float16
+).to("cuda")
+
+image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+normals = pipe(image)
+
+vis = pipe.image_processor.visualize_normals(normals.prediction)
+vis[0].save("einstein_normals.png")
+```
+
+The visualization function for normals [`~pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_normals`] maps the three-dimensional prediction with pixel values in the range `[-1, 1]` into an RGB image.
+The visualization function supports flipping surface normals axes to make the visualization compatible with other choices of the frame of reference.
+Conceptually, each pixel is painted according to the surface normal vector in the frame of reference, where `X` axis points right, `Y` axis points up, and `Z` axis points at the viewer.
+Below is the visualized prediction:
+
+<div class="flex gap-4" style="justify-content: center; width: 100%;">
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_lcm_normals.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Predicted surface normals visualization
+    </figcaption>
+  </div>
+</div>
+
+In this example, the nose tip almost certainly has a point on the surface, in which the surface normal vector points straight at the viewer, meaning that its coordinates are `[0, 0, 1]`.
+This vector maps to the RGB `[128, 128, 255]`, which corresponds to the violet-blue color.
+Similarly, a surface normal on the cheek in the right part of the image has a large `X` component, which increases the red hue.
+Points on the shoulders pointing up with a large `Y` promote green color.
+
+### Speeding up inference
+
+The above quick start snippets are already optimized for speed: they load the LCM checkpoint, use the `fp16` variant of weights and computation, and perform just one denoising diffusion step.
+The `pipe(image)` call completes in 280ms on RTX 3090 GPU.
+Internally, the input image is encoded with the Stable Diffusion VAE encoder, then the U-Net performs one denoising step, and finally, the prediction latent is decoded with the VAE decoder into pixel space.
+In this case, two out of three module calls are dedicated to converting between pixel and latent space of LDM.
+Because Marigold's latent space is compatible with the base Stable Diffusion, it is possible to speed up the pipeline call by more than 3x (85ms on RTX 3090) by using a [lightweight replacement of the SD VAE](../api/models/autoencoder_tiny):
+
+```diff
+  import diffusers
+  import torch
+  
+  pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
+      "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
+  ).to("cuda")
+  
++ pipe.vae = diffusers.AutoencoderTiny.from_pretrained(
++     "madebyollin/taesd", torch_dtype=torch.float16
++ ).cuda()
+  
+  image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+  depth = pipe(image)
+```
+
+As suggested in [Optimizations](../optimization/torch2.0#torch.compile), adding `torch.compile` may squeeze extra performance depending on the target hardware:
+
+```diff
+  import diffusers
+  import torch
+  
+  pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
+      "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
+  ).to("cuda")
+  
++ pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+  
+  image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+  depth = pipe(image)
+```
+
+## Qualitative Comparison with Depth Anything
+
+With the above speed optimizations, Marigold delivers predictions with more details and faster than [Depth Anything](https://huggingface.co/docs/transformers/main/en/model_doc/depth_anything) with the largest checkpoint [LiheYoung/depth-anything-large-hf](https://huggingface.co/LiheYoung/depth-anything-large-hf):
+
+<div class="flex gap-4">
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_lcm_depth.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Marigold LCM fp16 with Tiny AutoEncoder
+    </figcaption>
+  </div>
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/einstein_depthanything_large.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Depth Anything Large
+    </figcaption>
+  </div>
+</div>
+
+## Maximizing Precision and Ensembling
+
+Marigold pipelines have a built-in ensembling mechanism combining multiple predictions from different random latents.
+This is a brute-force way of improving the precision of predictions, capitalizing on the generative nature of diffusion.
+The ensembling path is activated automatically when the `ensemble_size` argument is set greater than `1`.
+When aiming for maximum precision, it makes sense to adjust `num_inference_steps` simultaneously with `ensemble_size`.
+The recommended values vary across checkpoints but primarily depend on the scheduler type.
+The effect of ensembling is particularly well-seen with surface normals:
+
+```python
+import diffusers
+
+model_path = "prs-eth/marigold-normals-v1-0"
+
+model_paper_kwargs = {
+	diffusers.schedulers.DDIMScheduler: {
+		"num_inference_steps": 10,
+		"ensemble_size": 10,
+	},
+	diffusers.schedulers.LCMScheduler: {
+		"num_inference_steps": 4,
+		"ensemble_size": 5,
+	},	
+}
+
+image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+
+pipe = diffusers.MarigoldNormalsPipeline.from_pretrained(model_path).to("cuda")
+pipe_kwargs = model_paper_kwargs[type(pipe.scheduler)]
+
+depth = pipe(image, **pipe_kwargs)
+
+vis = pipe.image_processor.visualize_normals(depth.prediction)
+vis[0].save("einstein_normals.png")
+```
+
+<div class="flex gap-4">
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_lcm_normals.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Surface normals, no ensembling
+    </figcaption>
+  </div>
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_normals.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Surface normals, with ensembling
+    </figcaption>
+  </div>
+</div>
+
+As can be seen, all areas with fine-grained structurers, such as hair, got more conservative and on average more correct predictions.
+Such a result is more suitable for precision-sensitive downstream tasks, such as 3D reconstruction.
+
+## Quantitative Evaluation
+
+To evaluate Marigold quantitatively in standard leaderboards and benchmarks (such as NYU, KITTI, and other datasets), follow the evaluation protocol outlined in the paper: load the full precision fp32 model and use appropriate values for `num_inference_steps` and `ensemble_size`.
+Optionally seed randomness to ensure reproducibility. Maximizing `batch_size` will deliver maximum device utilization.
+
+```python
+import diffusers
+import torch
+
+device = "cuda"
+seed = 2024
+model_path = "prs-eth/marigold-v1-0"
+
+model_paper_kwargs = {
+	diffusers.schedulers.DDIMScheduler: {
+		"num_inference_steps": 50,
+		"ensemble_size": 10,
+	},
+	diffusers.schedulers.LCMScheduler: {
+		"num_inference_steps": 4,
+		"ensemble_size": 10,
+	},	
+}
+
+image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+
+generator = torch.Generator(device=device).manual_seed(seed)
+pipe = diffusers.MarigoldDepthPipeline.from_pretrained(model_path).to(device)
+pipe_kwargs = model_paper_kwargs[type(pipe.scheduler)]
+
+depth = pipe(image, generator=generator, **pipe_kwargs)
+
+# evaluate metrics
+```
+
+## Using Predictive Uncertainty
+
+The ensembling mechanism built into Marigold pipelines combines multiple predictions obtained from different random latents.
+As a side effect, it can be used to quantify epistemic (model) uncertainty; simply specify `ensemble_size` greater than 1 and set `output_uncertainty=True`.
+The resulting uncertainty will be available in the `uncertainty` field of the output.
+It can be visualized as follows:
+
+```python
+import diffusers
+import torch
+
+pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
+    "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
+).to("cuda")
+
+image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+depth = pipe(
+	image,
+	ensemble_size=10,  # any number greater than 1; higher values yield higher precision
+	output_uncertainty=True,
+)
+
+uncertainty = pipe.image_processor.visualize_uncertainty(depth.uncertainty)
+uncertainty[0].save("einstein_depth_uncertainty.png")
+```
+
+<div class="flex gap-4">
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_depth_uncertainty.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Depth uncertainty
+    </figcaption>
+  </div>
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_normals_uncertainty.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Surface normals uncertainty
+    </figcaption>
+  </div>
+</div>
+
+The interpretation of uncertainty is easy: higher values (white) correspond to pixels, where the model struggles to make consistent predictions.
+Evidently, the depth model is the least confident around edges with discontinuity, where the object depth changes drastically.
+The surface normals model is the least confident in fine-grained structures, such as hair, and dark areas, such as the collar.
+
+## Frame-by-frame Video Processing with Temporal Consistency
+
+Due to Marigold's generative nature, each prediction is unique and defined by the random noise sampled for the latent initialization.
+This becomes an obvious drawback compared to traditional end-to-end dense regression networks, as exemplified in the following videos:
+
+<div class="flex gap-4">
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_obama.gif"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">Input video</figcaption>
+  </div>
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_obama_depth_independent.gif"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">Marigold Depth applied to input video frames independently</figcaption>
+  </div>
+</div>
+
+To address this issue, it is possible to pass `latents` argument to the pipelines, which defines the starting point of diffusion.
+Empirically, we found that a convex combination of the very same starting point noise latent and the latent corresponding to the previous frame prediction give sufficiently smooth results, as implemented in the snippet below:
+
+```python
+import imageio
+from PIL import Image
+from tqdm import tqdm
+import diffusers
+import torch
+
+device = "cuda"
+path_in = "obama.mp4"
+path_out = "obama_depth.gif"
+
+pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
+    "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
+).to(device)
+pipe.vae = diffusers.AutoencoderTiny.from_pretrained(
+    "madebyollin/taesd", torch_dtype=torch.float16
+).to(device)
+pipe.set_progress_bar_config(disable=True)
+
+with imageio.get_reader(path_in) as reader:
+    size = reader.get_meta_data()['size']
+    last_frame_latent = None
+    latent_common = torch.randn(
+        (1, 4, 768 * size[1] // (8 * max(size)), 768 * size[0] // (8 * max(size)))
+    ).to(device=device, dtype=torch.float16)
+
+    out = []
+    for frame_id, frame in tqdm(enumerate(reader), desc="Processing Video"):
+        frame = Image.fromarray(frame)
+        latents = latent_common
+        if last_frame_latent is not None:
+            latents = 0.9 * latents + 0.1 * last_frame_latent
+
+        depth = pipe(
+			frame, match_input_resolution=False, latents=latents, output_latent=True
+        )
+        last_frame_latent = depth.latent
+        out.append(pipe.image_processor.visualize_depth(depth.prediction)[0])
+
+    diffusers.utils.export_to_gif(out, path_out, fps=reader.get_meta_data()['fps'])
+```
+
+Here, the diffusion process starts from the given computed latent.
+The pipeline sets `output_latent=True` to access `out.latent` and computes its contribution to the next frame's latent initialization.
+The result is much more stable now:
+
+<div class="flex gap-4">
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_obama_depth_independent.gif"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">Marigold Depth applied to input video frames independently</figcaption>
+  </div>
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_obama_depth_consistent.gif"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">Marigold Depth with forced latents initialization</figcaption>
+  </div>
+</div>
+
+## Marigold for ControlNet
+
+A very common application for depth prediction with diffusion models comes in conjunction with ControlNet.
+Depth crispness plays a crucial role in obtaining high-quality results from ControlNet.
+As seen in comparisons with other methods above, Marigold excels at that task.
+The snippet below demonstrates how to load an image, compute depth, and pass it into ControlNet in a compatible format:
+
+```python
+import torch
+import diffusers
+
+device = "cuda"
+generator = torch.Generator(device=device).manual_seed(2024)
+image = diffusers.utils.load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/controlnet_depth_source.png"
+)
+
+pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
+    "prs-eth/marigold-depth-lcm-v1-0", torch_dtype=torch.float16, variant="fp16"
+).to("cuda")
+
+depth_image = pipe(image, generator=generator).prediction
+depth_image = pipe.image_processor.visualize_depth(depth_image, color_map="binary")
+depth_image[0].save("motorcycle_controlnet_depth.png")
+
+controlnet = diffusers.ControlNetModel.from_pretrained(
+    "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16"
+).to("cuda")
+pipe = diffusers.StableDiffusionXLControlNetPipeline.from_pretrained(
+    "SG161222/RealVisXL_V4.0", torch_dtype=torch.float16, variant="fp16", controlnet=controlnet
+).to("cuda")
+pipe.scheduler = diffusers.DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True)
+
+controlnet_out = pipe(
+    prompt="high quality photo of a sports bike, city",
+    negative_prompt="",
+    guidance_scale=6.5,
+    num_inference_steps=25,
+    image=depth_image,
+    controlnet_conditioning_scale=0.7,
+    control_guidance_end=0.7,
+    generator=generator,
+).images
+controlnet_out[0].save("motorcycle_controlnet_out.png")
+```
+
+<div class="flex gap-4">
+  <div style="flex: 1 1 33%; max-width: 33%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/controlnet_depth_source.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Input image
+    </figcaption>
+  </div>
+  <div style="flex: 1 1 33%; max-width: 33%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/motorcycle_controlnet_depth.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Depth in the format compatible with ControlNet
+    </figcaption>
+  </div>
+  <div style="flex: 1 1 33%; max-width: 33%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/motorcycle_controlnet_out.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      ControlNet generation, conditioned on depth and prompt: "high quality photo of a sports bike, city"
+    </figcaption>
+  </div>
+</div>
+
+Hopefully, you will find Marigold useful for solving your downstream tasks, be it a part of a more broad generative workflow, or a perception task, such as 3D reconstruction.

docs/source/en/using-diffusers/other-formats.md

@@ -10,156 +10,86 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Load different Stable Diffusion formats
+# Model files and layouts
 
 [[open-in-colab]]
 
-Stable Diffusion models are available in different formats depending on the framework they're trained and saved with, and where you download them from. Converting these formats for use in 🤗 Diffusers allows you to use all the features supported by the library, such as [using different schedulers](schedulers) for inference, [building your custom pipeline](write_own_pipeline), and a variety of techniques and methods for [optimizing inference speed](../optimization/opt_overview).
+Diffusion models are saved in various file types and organized in different layouts. Diffusers stores model weights as safetensors files in *Diffusers-multifolder* layout and it also supports loading files (like safetensors and ckpt files) from a *single-file* layout which is commonly used in the diffusion ecosystem.
 
-<Tip>
+Each layout has its own benefits and use cases, and this guide will show you how to load the different files and layouts, and how to convert them.
 
-We highly recommend using the `.safetensors` format because it is more secure than traditional pickled files which are vulnerable and can be exploited to execute any code on your machine (learn more in the [Load safetensors](using_safetensors) guide).
+## Files
 
-</Tip>
+PyTorch model weights are typically saved with Python's [pickle](https://docs.python.org/3/library/pickle.html) utility as ckpt or bin files. However, pickle is not secure and pickled files may contain malicious code that can be executed. This vulnerability is a serious concern given the popularity of model sharing. To address this security issue, the [Safetensors](https://hf.co/docs/safetensors) library was developed as a secure alternative to pickle, which saves models as safetensors files.
 
-This guide will show you how to convert other Stable Diffusion formats to be compatible with 🤗 Diffusers.
+### safetensors
 
-## PyTorch .ckpt
+> [!TIP]
+> Learn more about the design decisions and why safetensor files are preferred for saving and loading model weights in the [Safetensors audited as really safe and becoming the default](https://blog.eleuther.ai/safetensors-security-audit/) blog post.
 
-The checkpoint - or `.ckpt` - format is commonly used to store and save models. The `.ckpt` file contains the entire model and is typically several GBs in size. While you can load and use a `.ckpt` file directly with the [`~StableDiffusionPipeline.from_single_file`] method, it is generally better to convert the `.ckpt` file to 🤗 Diffusers so both formats are available.
+[Safetensors](https://hf.co/docs/safetensors) is a safe and fast file format for securely storing and loading tensors. Safetensors restricts the header size to limit certain types of attacks, supports lazy loading (useful for distributed setups), and has generally faster loading speeds.
 
-There are two options for converting a `.ckpt` file: use a Space to convert the checkpoint or convert the `.ckpt` file with a script.
+Make sure you have the [Safetensors](https://hf.co/docs/safetensors) library installed.
 
-### Convert with a Space
-
-The easiest and most convenient way to convert a `.ckpt` file is to use the [SD to Diffusers](https://huggingface.co/spaces/diffusers/sd-to-diffusers) Space. You can follow the instructions on the Space to convert the `.ckpt` file.
-
-This approach works well for basic models, but it may struggle with more customized models. You'll know the Space failed if it returns an empty pull request or error. In this case, you can try converting the `.ckpt` file with a script.
-
-### Convert with a script
-
-🤗 Diffusers provides a [conversion script](https://github.com/huggingface/diffusers/blob/main/scripts/convert_original_stable_diffusion_to_diffusers.py) for converting `.ckpt` files. This approach is more reliable than the Space above.
-
-Before you start, make sure you have a local clone of 🤗 Diffusers to run the script and log in to your Hugging Face account so you can open pull requests and push your converted model to the Hub.
-
-```bash
-huggingface-cli login
+```py
+!pip install safetensors
 ```
 
-To use the script:
+Safetensors stores weights in a safetensors file. Diffusers loads safetensors files by default if they're available and the Safetensors library is installed. There are two ways safetensors files can be organized:
 
-1. Git clone the repository containing the `.ckpt` file you want to convert. For this example, let's convert this [TemporalNet](https://huggingface.co/CiaraRowles/TemporalNet) `.ckpt` file:
+1. Diffusers-multifolder layout: there may be several separate safetensors files, one for each pipeline component (text encoder, UNet, VAE), organized in subfolders (check out the [runwayml/stable-diffusion-v1-5](https://hf.co/runwayml/stable-diffusion-v1-5/tree/main) repository as an example)
+2. single-file layout: all the model weights may be saved in a single file (check out the [WarriorMama777/OrangeMixs](https://hf.co/WarriorMama777/OrangeMixs/tree/main/Models/AbyssOrangeMix) repository as an example)
 
-```bash
-git lfs install
-git clone https://huggingface.co/CiaraRowles/TemporalNet
-```
+<hfoptions id="safetensors">
+<hfoption id="multifolder">
 
-2. Open a pull request on the repository where you're converting the checkpoint from:
-
-```bash
-cd TemporalNet && git fetch origin refs/pr/13:pr/13
-git checkout pr/13
-```
-
-3. There are several input arguments to configure in the conversion script, but the most important ones are:
-
-    - `checkpoint_path`: the path to the `.ckpt` file to convert.
-    - `original_config_file`: a YAML file defining the configuration of the original architecture. If you can't find this file, try searching for the YAML file in the GitHub repository where you found the `.ckpt` file.
-    - `dump_path`: the path to the converted model.
-
-        For example, you can take the `cldm_v15.yaml` file from the [ControlNet](https://github.com/lllyasviel/ControlNet/tree/main/models) repository because the TemporalNet model is a Stable Diffusion v1.5 and ControlNet model.
-
-4. Now you can run the script to convert the `.ckpt` file:
-
-```bash
-python ../diffusers/scripts/convert_original_stable_diffusion_to_diffusers.py --checkpoint_path temporalnetv3.ckpt --original_config_file cldm_v15.yaml --dump_path ./ --controlnet
-```
-
-5. Once the conversion is done, upload your converted model and test out the resulting [pull request](https://huggingface.co/CiaraRowles/TemporalNet/discussions/13)!
-
-```bash
-git push origin pr/13:refs/pr/13
-```
-
-## Keras .pb or .h5
-
-<Tip warning={true}>
-
-🧪 This is an experimental feature. Only Stable Diffusion v1 checkpoints are supported by the Convert KerasCV Space at the moment.
-
-</Tip>
-
-[KerasCV](https://keras.io/keras_cv/) supports training for [Stable Diffusion](https://github.com/keras-team/keras-cv/blob/master/keras_cv/models/stable_diffusion) v1 and v2. However, it offers limited support for experimenting with Stable Diffusion models for inference and deployment whereas 🤗 Diffusers has a more complete set of features for this purpose, such as different [noise schedulers](https://huggingface.co/docs/diffusers/using-diffusers/schedulers), [flash attention](https://huggingface.co/docs/diffusers/optimization/xformers), and [other
-optimization techniques](https://huggingface.co/docs/diffusers/optimization/fp16).
-
-The [Convert KerasCV](https://huggingface.co/spaces/sayakpaul/convert-kerascv-sd-diffusers) Space converts `.pb` or `.h5` files to PyTorch, and then wraps them in a [`StableDiffusionPipeline`] so it is ready for inference. The converted checkpoint is stored in a repository on the Hugging Face Hub.
-
-For this example, let's convert the [`sayakpaul/textual-inversion-kerasio`](https://huggingface.co/sayakpaul/textual-inversion-kerasio/tree/main) checkpoint which was trained with Textual Inversion. It uses the special token `<my-funny-cat>` to personalize images with cats.
-
-The Convert KerasCV Space allows you to input the following:
-
-* Your Hugging Face token.
-* Paths to download the UNet and text encoder weights from. Depending on how the model was trained, you don't necessarily need to provide the paths to both the UNet and text encoder. For example, Textual Inversion only requires the embeddings from the text encoder and a text-to-image model only requires the UNet weights.
-* Placeholder token is only applicable for textual inversion models.
-* The `output_repo_prefix` is the name of the repository where the converted model is stored.
-
-Click the **Submit** button to automatically convert the KerasCV checkpoint! Once the checkpoint is successfully converted, you'll see a link to the new repository containing the converted checkpoint. Follow the link to the new repository, and you'll see the Convert KerasCV Space generated a model card with an inference widget to try out the converted model.
-
-If you prefer to run inference with code, click on the **Use in Diffusers** button in the upper right corner of the model card to copy and paste the code snippet:
+Use the [`~DiffusionPipeline.from_pretrained`] method to load a model with safetensors files stored in multiple folders.
 
 ```py
 from diffusers import DiffusionPipeline
 
 pipeline = DiffusionPipeline.from_pretrained(
-    "sayakpaul/textual-inversion-cat-kerascv_sd_diffusers_pipeline", use_safetensors=True
+    "runwayml/stable-diffusion-v1-5",
+    use_safetensors=True
 )
 ```
 
-Then, you can generate an image like:
+</hfoption>
+<hfoption id="single file">
+
+Use the [`~loaders.FromSingleFileMixin.from_single_file`] method to load a model with all the weights stored in a single safetensors file.
 
 ```py
-from diffusers import DiffusionPipeline
+from diffusers import StableDiffusionPipeline
 
-pipeline = DiffusionPipeline.from_pretrained(
-    "sayakpaul/textual-inversion-cat-kerascv_sd_diffusers_pipeline", use_safetensors=True
+pipeline = StableDiffusionPipeline.from_single_file(
+    "https://huggingface.co/WarriorMama777/OrangeMixs/blob/main/Models/AbyssOrangeMix/AbyssOrangeMix.safetensors"
 )
-pipeline.to("cuda")
-
-placeholder_token = "<my-funny-cat-token>"
-prompt = f"two {placeholder_token} getting married, photorealistic, high quality"
-image = pipeline(prompt, num_inference_steps=50).images[0]
 ```
 
-## A1111 LoRA files
+</hfoption>
+</hfoptions>
 
-[Automatic1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) (A1111) is a popular web UI for Stable Diffusion that supports model sharing platforms like [Civitai](https://civitai.com/). Models trained with the Low-Rank Adaptation (LoRA) technique are especially popular because they're fast to train and have a much smaller file size than a fully finetuned model. 🤗 Diffusers supports loading A1111 LoRA checkpoints with [`~loaders.LoraLoaderMixin.load_lora_weights`]:
+#### LoRA files
+
+[LoRA](https://hf.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) is a lightweight adapter that is fast and easy to train, making them especially popular for generating images in a certain way or style. These adapters are commonly stored in a safetensors file, and are widely popular on model sharing platforms like [civitai](https://civitai.com/).
+
+LoRAs are loaded into a base model with the [`~loaders.LoraLoaderMixin.load_lora_weights`] method.
 
 ```py
 from diffusers import StableDiffusionXLPipeline
 import torch
 
+# base model
 pipeline = StableDiffusionXLPipeline.from_pretrained(
     "Lykon/dreamshaper-xl-1-0", torch_dtype=torch.float16, variant="fp16"
 ).to("cuda")
-```
 
-Download a LoRA checkpoint from Civitai; this example uses the [Blueprintify SD XL 1.0](https://civitai.com/models/150986/blueprintify-sd-xl-10) checkpoint, but feel free to try out any LoRA checkpoint!
+# download LoRA weights
+!wget https://civitai.com/api/download/models/168776 -O blueprintify.safetensors
 
-```py
-# uncomment to download the safetensor weights
-#!wget https://civitai.com/api/download/models/168776 -O blueprintify.safetensors
-```
-
-Load the LoRA checkpoint into the pipeline with the [`~loaders.LoraLoaderMixin.load_lora_weights`] method:
-
-```py
+# load LoRA weights
 pipeline.load_lora_weights(".", weight_name="blueprintify.safetensors")
-```
-
-Now you can use the pipeline to generate images:
-
-```py
 prompt = "bl3uprint, a highly detailed blueprint of the empire state building, explaining how to build all parts, many txt, blueprint grid backdrop"
 negative_prompt = "lowres, cropped, worst quality, low quality, normal quality, artifacts, signature, watermark, username, blurry, more than one bridge, bad architecture"
 
@@ -174,3 +104,166 @@ image
 <div class="flex justify-center">
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/blueprint-lora.png"/>
 </div>
+
+### ckpt
+
+> [!WARNING]
+> Pickled files may be unsafe because they can be exploited to execute malicious code. It is recommended to use safetensors files instead where possible, or convert the weights to safetensors files.
+
+PyTorch's [torch.save](https://pytorch.org/docs/stable/generated/torch.save.html) function uses Python's [pickle](https://docs.python.org/3/library/pickle.html) utility to serialize and save models. These files are saved as a ckpt file and they contain the entire model's weights.
+
+Use the [`~loaders.FromSingleFileMixin.from_single_file`] method to directly load a ckpt file.
+
+```py
+from diffusers import StableDiffusionPipeline
+
+pipeline = StableDiffusionPipeline.from_single_file(
+    "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned.ckpt"
+)
+```
+
+## Storage layout
+
+There are two ways model files are organized, either in a Diffusers-multifolder layout or in a single-file layout. The Diffusers-multifolder layout is the default, and each component file (text encoder, UNet, VAE) is stored in a separate subfolder. Diffusers also supports loading models from a single-file layout where all the components are bundled together.
+
+### Diffusers-multifolder
+
+The Diffusers-multifolder layout is the default storage layout for Diffusers. Each component's (text encoder, UNet, VAE) weights are stored in a separate subfolder. The weights can be stored as safetensors or ckpt files.
+
+<div class="flex flex-row gap-4">
+  <div class="flex-1">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/multifolder-layout.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">multifolder layout</figcaption>
+  </div>
+  <div class="flex-1">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/multifolder-unet.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">UNet subfolder</figcaption>
+  </div>
+</div>
+
+To load from Diffusers-multifolder layout, use the [`~DiffusionPipeline.from_pretrained`] method.
+
+```py
+from diffusers import DiffusionPipeline
+
+pipeline = DiffusionPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    use_safetensors=True,
+).to("cuda")
+```
+
+Benefits of using the Diffusers-multifolder layout include:
+
+1. Faster to load each component file individually or in parallel.
+2. Reduced memory usage because you only load the components you need. For example, models like [SDXL Turbo](https://hf.co/stabilityai/sdxl-turbo), [SDXL Lightning](https://hf.co/ByteDance/SDXL-Lightning), and [Hyper-SD](https://hf.co/ByteDance/Hyper-SD) have the same components except for the UNet. You can reuse their shared components with the [`~DiffusionPipeline.from_pipe`] method without consuming any additional memory (take a look at the [Reuse a pipeline](./loading#reuse-a-pipeline) guide) and only load the UNet. This way, you don't need to download redundant components and unnecessarily use more memory.
+
+    ```py
+    import torch
+    from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel, EulerDiscreteScheduler
+
+    # download one model
+    sdxl_pipeline = StableDiffusionXLPipeline.from_pretrained(
+        "stabilityai/stable-diffusion-xl-base-1.0",
+        torch_dtype=torch.float16,
+        variant="fp16",
+        use_safetensors=True,
+    ).to("cuda")
+
+    # switch UNet for another model
+    unet = UNet2DConditionModel.from_pretrained(
+        "stabilityai/sdxl-turbo",
+        subfolder="unet",
+        torch_dtype=torch.float16,
+        variant="fp16",
+        use_safetensors=True
+    )
+    # reuse all the same components in new model except for the UNet
+    turbo_pipeline = StableDiffusionXLPipeline.from_pipe(
+        sdxl_pipeline, unet=unet,
+    ).to("cuda")
+    turbo_pipeline.scheduler = EulerDiscreteScheduler.from_config(
+        turbo_pipeline.scheduler.config,
+        timestep+spacing="trailing"
+    )
+    image = turbo_pipeline(
+        "an astronaut riding a unicorn on mars",
+        num_inference_steps=1,
+        guidance_scale=0.0,
+    ).images[0]
+    image
+    ```
+
+3. Reduced storage requirements because if a component, such as the SDXL [VAE](https://hf.co/madebyollin/sdxl-vae-fp16-fix), is shared across multiple models, you only need to download and store a single copy of it instead of downloading and storing it multiple times. For 10 SDXL models, this can save ~3.5GB of storage. The storage savings is even greater for newer models like PixArt Sigma, where the [text encoder](https://hf.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/tree/main/text_encoder) alone is ~19GB!
+4. Flexibility to replace a component in the model with a newer or better version.
+
+    ```py
+    from diffusers import DiffusionPipeline, AutoencoderKL
+
+    vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16, use_safetensors=True)
+    pipeline = DiffusionPipeline.from_pretrained(
+        "stabilityai/stable-diffusion-xl-base-1.0",
+        vae=vae,
+        torch_dtype=torch.float16,
+        variant="fp16",
+        use_safetensors=True,
+    ).to("cuda")
+    ```
+
+5. More visibility and information about a model's components, which are stored in a [config.json](https://hf.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/unet/config.json) file in each component subfolder.
+
+### Single-file
+
+The single-file layout stores all the model weights in a single file. All the model components (text encoder, UNet, VAE) weights are kept together instead of separately in subfolders. This can be a safetensors or ckpt file.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/single-file-layout.png"/>
+</div>
+
+To load from a single-file layout, use the [`~loaders.FromSingleFileMixin.from_single_file`] method.
+
+```py
+import torch
+from diffusers import StableDiffusionXLPipeline
+
+pipeline = StableDiffusionXLPipeline.from_single_file(
+    "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    use_safetensors=True,
+).to("cuda")
+```
+
+Benefits of using a single-file layout include:
+
+1. Easy compatibility with diffusion interfaces such as [ComfyUI](https://github.com/comfyanonymous/ComfyUI) or [Automatic1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) which commonly use a single-file layout.
+2. Easier to manage (download and share) a single file.
+
+## Convert layout and files
+
+Diffusers provides many scripts and methods to convert storage layouts and file formats to enable broader support across the diffusion ecosystem.
+
+Take a look at the [diffusers/scripts](https://github.com/huggingface/diffusers/tree/main/scripts) collection to find a script that fits your conversion needs.
+
+> [!TIP]
+> Scripts that have "`to_diffusers`" appended at the end mean they convert a model to the Diffusers-multifolder layout. Each script has their own specific set of arguments for configuring the conversion, so make sure you check what arguments are available!
+
+For example, to convert a Stable Diffusion XL model stored in Diffusers-multifolder layout to a single-file layout, run the [convert_diffusers_to_original_sdxl.py](https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_sdxl.py) script. Provide the path to the model to convert, and the path to save the converted model to. You can optionally specify whether you want to save the model as a safetensors file and whether to save the model in half-precision.
+
+```bash
+python convert_diffusers_to_original_sdxl.py --model_path path/to/model/to/convert --checkpoint_path path/to/save/model/to --use_safetensors
+```
+
+You can also save a model to Diffusers-multifolder layout with the [`~DiffusionPipeline.save_pretrained`] method. This creates a directory for you if it doesn't already exist, and it also saves the files as a safetensors file by default.
+
+```py
+from diffusers import StableDiffusionXLPipeline
+
+pipeline = StableDiffusionXLPipeline.from_single_file(
+    "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors",
+)
+pipeline.save_pretrained()
+```
+
+Lastly, there are also Spaces, such as [SD To Diffusers](https://hf.co/spaces/diffusers/sd-to-diffusers) and [SD-XL To Diffusers](https://hf.co/spaces/diffusers/sdxl-to-diffusers), that provide a more user-friendly interface for converting models to Diffusers-multifolder layout. This is the easiest and most convenient option for converting layouts, and it'll open a PR on your model repository with the converted files. However, this option is not as reliable as running a script, and the Space may fail for more complicated models.

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

@@ -0,0 +1,235 @@
+<!--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.
+-->
+
+# Scheduler features
+
+The scheduler is an important component of any diffusion model because it controls the entire denoising (or sampling) process. There are many types of schedulers, some are optimized for speed and some for quality. With Diffusers, you can modify the scheduler configuration to use custom noise schedules, sigmas, and rescale the noise schedule. Changing these parameters can have profound effects on inference quality and speed.
+
+This guide will demonstrate how to use these features to improve inference quality.
+
+> [!TIP]
+> Diffusers currently only supports the `timesteps` and `sigmas` parameters for a select list of schedulers and pipelines. Feel free to open a [feature request](https://github.com/huggingface/diffusers/issues/new/choose) if you want to extend these parameters to a scheduler and pipeline that does not currently support it!
+
+## Timestep schedules
+
+The timestep or noise schedule determines the amount of noise at each sampling step. The scheduler uses this to generate an image with the corresponding amount of noise at each step. The timestep schedule is generated from the scheduler's default configuration, but you can customize the scheduler to use new and optimized sampling schedules that aren't in Diffusers yet.
+
+For example, [Align Your Steps (AYS)](https://research.nvidia.com/labs/toronto-ai/AlignYourSteps/) is a method for optimizing a sampling schedule to generate a high-quality image in as little as 10 steps. The optimal [10-step schedule](https://github.com/huggingface/diffusers/blob/a7bf77fc284810483f1e60afe34d1d27ad91ce2e/src/diffusers/schedulers/scheduling_utils.py#L51) for Stable Diffusion XL is:
+
+```py
+from diffusers.schedulers import AysSchedules
+
+sampling_schedule = AysSchedules["StableDiffusionXLTimesteps"]
+print(sampling_schedule)
+"[999, 845, 730, 587, 443, 310, 193, 116, 53, 13]"
+```
+
+You can use the AYS sampling schedule in a pipeline by passing it to the `timesteps` parameter.
+
+```py
+pipeline = StableDiffusionXLPipeline.from_pretrained(
+    "SG161222/RealVisXL_V4.0",
+    torch_dtype=torch.float16,
+    variant="fp16",
+).to("cuda")
+pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, algorithm_type="sde-dpmsolver++")
+
+prompt = "A cinematic shot of a cute little rabbit wearing a jacket and doing a thumbs up"
+generator = torch.Generator(device="cpu").manual_seed(2487854446)
+image = pipeline(
+    prompt=prompt,
+    negative_prompt="",
+    generator=generator,
+    timesteps=sampling_schedule,
+).images[0]
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ays.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">AYS timestep schedule 10 steps</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/10.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">Linearly-spaced timestep schedule 10 steps</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/25.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">Linearly-spaced timestep schedule 25 steps</figcaption>
+  </div>
+</div>
+
+## Timestep spacing
+
+The way sample steps are selected in the schedule can affect the quality of the generated image, especially with respect to [rescaling the noise schedule](#rescale-noise-schedule), which can enable a model to generate much brighter or darker images. Diffusers provides three timestep spacing methods:
+
+- `leading` creates evenly spaced steps
+- `linspace` includes the first and last steps and evenly selects the remaining intermediate steps
+- `trailing` only includes the last step and evenly selects the remaining intermediate steps starting from the end
+
+It is recommended to use the `trailing` spacing method because it generates higher quality images with more details when there are fewer sample steps. But the difference in quality is not as obvious for more standard sample step values.
+
+```py
+import torch
+from diffusers import StableDiffusionXLPipeline, DPMSolverMultistepScheduler
+
+pipeline = StableDiffusionXLPipeline.from_pretrained(
+    "SG161222/RealVisXL_V4.0",
+    torch_dtype=torch.float16,
+    variant="fp16",
+).to("cuda")
+pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, timestep_spacing="trailing")
+
+prompt = "A cinematic shot of a cute little black cat sitting on a pumpkin at night"
+generator = torch.Generator(device="cpu").manual_seed(2487854446)
+image = pipeline(
+    prompt=prompt,
+    negative_prompt="",
+    generator=generator,
+    num_inference_steps=5,
+).images[0]
+image
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/trailing_spacing.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">trailing spacing after 5 steps</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/leading_spacing.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">leading spacing after 5 steps</figcaption>
+  </div>
+</div>
+
+## Sigmas
+
+The `sigmas` parameter is the amount of noise added at each timestep according to the timestep schedule. Like the `timesteps` parameter, you can customize the `sigmas` parameter to control how much noise is added at each step. When you use a custom `sigmas` value, the `timesteps` are calculated from the custom `sigmas` value and the default scheduler configuration is ignored.
+
+For example, you can manually pass the [sigmas](https://github.com/huggingface/diffusers/blob/6529ee67ec02fcf58d2fd9242164ea002b351d75/src/diffusers/schedulers/scheduling_utils.py#L55) for something like the 10-step AYS schedule from before to the pipeline.
+
+```py
+import torch
+
+from diffusers import DiffusionPipeline, EulerDiscreteScheduler
+
+model_id = "stabilityai/stable-diffusion-xl-base-1.0"
+pipeline = DiffusionPipeline.from_pretrained(
+  "stabilityai/stable-diffusion-xl-base-1.0",
+  torch_dtype=torch.float16,
+  variant="fp16",
+).to("cuda")
+pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config)
+
+sigmas = [14.615, 6.315, 3.771, 2.181, 1.342, 0.862, 0.555, 0.380, 0.234, 0.113, 0.0]
+prompt = "anthropomorphic capybara wearing a suit and working with a computer"
+generator = torch.Generator(device='cuda').manual_seed(123)
+image = pipeline(
+    prompt=prompt, 
+    num_inference_steps=10,
+    sigmas=sigmas,
+    generator=generator
+).images[0]
+```
+
+When you take a look at the scheduler's `timesteps` parameter, you'll see that it is the same as the AYS timestep schedule because the `timestep` schedule is calculated from the `sigmas`.
+
+```py
+print(f" timesteps: {pipe.scheduler.timesteps}")
+"timesteps: tensor([999., 845., 730., 587., 443., 310., 193., 116.,  53.,  13.], device='cuda:0')"
+```
+
+### Karras sigmas
+
+> [!TIP]
+> Refer to the scheduler API [overview](../api/schedulers/overview) for a list of schedulers that support Karras sigmas.
+>
+> Karras sigmas should not be used for models that weren't trained with them. For example, the base Stable Diffusion XL model shouldn't use Karras sigmas but the [DreamShaperXL](https://hf.co/Lykon/dreamshaper-xl-1-0) model can since they are trained with Karras sigmas.
+
+Karras scheduler's use the timestep schedule and sigmas from the [Elucidating the Design Space of Diffusion-Based Generative Models](https://hf.co/papers/2206.00364) paper. This scheduler variant applies a smaller amount of noise per step as it approaches the end of the sampling process compared to other schedulers, and can increase the level of details in the generated image.
+
+Enable Karras sigmas by setting `use_karras_sigmas=True` in the scheduler.
+
+```py
+import torch
+from diffusers import StableDiffusionXLPipeline, DPMSolverMultistepScheduler
+
+pipeline = StableDiffusionXLPipeline.from_pretrained(
+    "SG161222/RealVisXL_V4.0",
+    torch_dtype=torch.float16,
+    variant="fp16",
+).to("cuda")
+pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, algorithm_type="sde-dpmsolver++", use_karras_sigmas=True)
+
+prompt = "A cinematic shot of a cute little rabbit wearing a jacket and doing a thumbs up"
+generator = torch.Generator(device="cpu").manual_seed(2487854446)
+image = pipeline(
+    prompt=prompt,
+    negative_prompt="",
+    generator=generator,
+).images[0]
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/karras_sigmas_true.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">Karras sigmas enabled</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/stevhliu/testing-images/resolve/main/karras_sigmas_false.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">Karras sigmas disabled</figcaption>
+  </div>
+</div>
+
+## Rescale noise schedule
+
+In the [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://hf.co/papers/2305.08891) paper, the authors discovered that common noise schedules allowed some signal to leak into the last timestep. This signal leakage at inference can cause models to only generate images with medium brightness. By enforcing a zero signal-to-noise ratio (SNR) for the timstep schedule and sampling from the last timestep, the model can be improved to generate very bright or dark images.
+
+> [!TIP]
+> For inference, you need a model that has been trained with *v_prediction*. To train your own model with *v_prediction*, add the following flag to the [train_text_to_image.py](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py) or [train_text_to_image_lora.py](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_lora.py) scripts.
+>
+> ```bash
+> --prediction_type="v_prediction"
+> ```
+
+For example, load the [ptx0/pseudo-journey-v2](https://hf.co/ptx0/pseudo-journey-v2) checkpoint which was trained with `v_prediction` and the [`DDIMScheduler`]. Configure the following parameters in the [`DDIMScheduler`]:
+
+* `rescale_betas_zero_snr=True` to rescale the noise schedule to zero SNR
+* `timestep_spacing="trailing"` to start sampling from the last timestep
+
+Set `guidance_rescale` in the pipeline to prevent over-exposure. A lower value increases brightness but some of the details may appear washed out.
+
+```py
+from diffusers import DiffusionPipeline, DDIMScheduler
+
+pipeline = DiffusionPipeline.from_pretrained("ptx0/pseudo-journey-v2", use_safetensors=True)
+
+pipeline.scheduler = DDIMScheduler.from_config(
+    pipeline.scheduler.config, rescale_betas_zero_snr=True, timestep_spacing="trailing"
+)
+pipeline.to("cuda")
+prompt = "cinematic photo of a snowy mountain at night with the northern lights aurora borealis overhead, 35mm photograph, film, professional, 4k, highly detailed"
+generator = torch.Generator(device="cpu").manual_seed(23)
+image = pipeline(prompt, guidance_rescale=0.7, generator=generator).images[0]
+image
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/no-zero-snr.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">default Stable Diffusion v2-1 image</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/zero-snr.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">image with zero SNR and trailing timestep spacing enabled</figcaption>
+  </div>
+</div>

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

@@ -212,62 +212,6 @@ images = pipeline(prompt_ids, params, prng_seed, num_inference_steps, jit=True).
 images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:])))
 ```
 
-## Custom Timestep Schedules
-
-With all our schedulers, you can choose one of the popular timestep schedules using configurations such as `timestep_spacing`, `interpolation_type`, and `use_karras_sigmas`. Some schedulers also provide the flexibility to use a custom timestep schedule. You can use any list of arbitrary timesteps, we will use the AYS timestep schedule here as example. It is a set of 10-step optimized timestep schedules released by researchers from Nvidia that can achieve significantly better quality compared to the preset timestep schedules. You can read more about their research [here](https://research.nvidia.com/labs/toronto-ai/AlignYourSteps/). 
-
-```python
-from diffusers.schedulers import AysSchedules
-sampling_schedule = AysSchedules["StableDiffusionXLTimesteps"]
-print(sampling_schedule)
-```
-```
-[999, 845, 730, 587, 443, 310, 193, 116, 53, 13]
-```
-
-You can then create a pipeline and pass this custom timestep schedule to it as `timesteps`.
-
-```python
-pipe = StableDiffusionXLPipeline.from_pretrained(
-    "SG161222/RealVisXL_V4.0",
-    torch_dtype=torch.float16,
-    variant="fp16",
-).to("cuda")
-
-pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, algorithm_type="sde-dpmsolver++")
-
-prompt = "A cinematic shot of a cute little rabbit wearing a jacket and doing a thumbs up"
-
-generator = torch.Generator(device="cpu").manual_seed(2487854446)
-
-image = pipe(
-    prompt=prompt,
-    negative_prompt="",
-    generator=generator,
-    timesteps=sampling_schedule,
-).images[0]
-```
-The generated image has better quality than the default linear timestep schedule for the same number of steps, and it is similar to the default timestep scheduler when running for 25 steps.
-
-<div class="flex gap-4">
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ays.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">AYS timestep schedule 10 steps</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/10.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">Linearly-spaced timestep schedule 10 steps</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/25.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">Linearly-spaced timestep schedule 25 steps</figcaption>
-  </div>
-</div>
-
-> [!TIP]
-> 🤗 Diffusers currently only supports `timesteps` and `sigmas` for a selected list of schedulers and pipelines, but feel free to open a [feature request](https://github.com/huggingface/diffusers/issues/new/choose) if you want to extend feature to a scheduler and pipeline that does not currently support it!
-
-
 ## Models
 
 Models are loaded from the [`ModelMixin.from_pretrained`] method, which downloads and caches the latest version of the model weights and configurations. If the latest files are available in the local cache, [`~ModelMixin.from_pretrained`] reuses files in the cache instead of re-downloading them.

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

@@ -1,84 +0,0 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
--->
-
-# Load safetensors
-
-[[open-in-colab]]
-
-[safetensors](https://github.com/huggingface/safetensors) is a safe and fast file format for storing and loading tensors. Typically, PyTorch model weights are saved or *pickled* into a `.bin` file with Python's [`pickle`](https://docs.python.org/3/library/pickle.html) utility. However, `pickle` is not secure and pickled files may contain malicious code that can be executed. safetensors is a secure alternative to `pickle`, making it ideal for sharing model weights.
-
-This guide will show you how you load `.safetensor` files, and how to convert Stable Diffusion model weights stored in other formats to `.safetensor`. Before you start, make sure you have safetensors installed:
-
-```py
-# uncomment to install the necessary libraries in Colab
-#!pip install safetensors
-```
-
-If you look at the [`runwayml/stable-diffusion-v1-5`](https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main) repository, you'll see weights inside the `text_encoder`, `unet` and `vae` subfolders are stored in the `.safetensors` format. By default, 🤗 Diffusers automatically loads these `.safetensors` files from their subfolders if they're available in the model repository.
-
-For more explicit control, you can optionally set `use_safetensors=True` (if `safetensors` is not installed, you'll get an error message asking you to install it):
-
-```py
-from diffusers import DiffusionPipeline
-
-pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", use_safetensors=True)
-```
-
-However, model weights are not necessarily stored in separate subfolders like in the example above. Sometimes, all the weights are stored in a single `.safetensors` file. In this case, if the weights are Stable Diffusion weights, you can load the file directly with the [`~diffusers.loaders.FromSingleFileMixin.from_single_file`] method:
-
-```py
-from diffusers import StableDiffusionPipeline
-
-pipeline = StableDiffusionPipeline.from_single_file(
-    "https://huggingface.co/WarriorMama777/OrangeMixs/blob/main/Models/AbyssOrangeMix/AbyssOrangeMix.safetensors"
-)
-```
-
-## Convert to safetensors
-
-Not all weights on the Hub are available in the `.safetensors` format, and you may encounter weights stored as `.bin`. In this case, use the [Convert Space](https://huggingface.co/spaces/diffusers/convert) to convert the weights to `.safetensors`. The Convert Space downloads the pickled weights, converts them, and opens a Pull Request to upload the newly converted `.safetensors` file on the Hub. This way, if there is any malicious code contained in the pickled files, they're uploaded to the Hub - which has a [security scanner](https://huggingface.co/docs/hub/security-pickle#hubs-security-scanner) to detect unsafe files and suspicious pickle imports - instead of your computer.
-
-You can use the model with the new `.safetensors` weights by specifying the reference to the Pull Request in the `revision` parameter (you can also test it in this [Check PR](https://huggingface.co/spaces/diffusers/check_pr) Space on the Hub), for example `refs/pr/22`:
-
-```py
-from diffusers import DiffusionPipeline
-
-pipeline = DiffusionPipeline.from_pretrained(
-    "stabilityai/stable-diffusion-2-1", revision="refs/pr/22", use_safetensors=True
-)
-```
-
-## Why use safetensors?
-
-There are several reasons for using safetensors:
-
-- Safety is the number one reason for using safetensors. As open-source and model distribution grows, it is important to be able to trust the model weights you downloaded don't contain any malicious code. The current size of the header in safetensors prevents parsing extremely large JSON files.
-- Loading speed between switching models is another reason to use safetensors, which performs zero-copy of the tensors. It is especially fast compared to `pickle` if you're loading the weights to CPU (the default case), and just as fast if not faster when directly loading the weights to GPU. You'll only notice the performance difference if the model is already loaded, and not if you're downloading the weights or loading the model for the first time.
-
-	The time it takes to load the entire pipeline:
-
-	```py
- 	from diffusers import StableDiffusionPipeline
-
- 	pipeline = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", use_safetensors=True)
- 	"Loaded in safetensors 0:00:02.033658"
- 	"Loaded in PyTorch 0:00:02.663379"
-	```
-
-	But the actual time it takes to load 500MB of the model weights is only:
-
-	```bash
-	safetensors: 3.4873ms
-	PyTorch: 172.7537ms
-	```
-
-- Lazy loading is also supported in safetensors, which is useful in distributed settings to only load some of the tensors. This format allowed the [BLOOM](https://huggingface.co/bigscience/bloom) model to be loaded in 45 seconds on 8 GPUs instead of 10 minutes with regular PyTorch weights.

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -71,7 +71,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 
@@ -758,7 +758,7 @@ class TokenEmbeddingsHandler:
 
             idx += 1
 
-    # copied from train_dreambooth_lora_sdxl_advanced.py
+    # Copied from train_dreambooth_lora_sdxl_advanced.py
     def save_embeddings(self, file_path: str):
         assert self.train_ids is not None, "Initialize new tokens before saving embeddings."
         tensors = {}

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -78,7 +78,7 @@ from diffusers.utils.torch_utils import is_compiled_module
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/community/README.md

@@ -69,6 +69,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 |   UFOGen Scheduler                                                                                               | Scheduler for UFOGen Model (compatible with Stable Diffusion pipelines)                                                                                                                                                                                                                                                                                                                                                 |  [UFOGen Scheduler](#ufogen-scheduler) | - | [dg845](https://github.com/dg845) |
 | Stable Diffusion XL IPEX Pipeline | Accelerate Stable Diffusion XL inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [Stable Diffusion XL on IPEX](#stable-diffusion-xl-on-ipex) | - | [Dan Li](https://github.com/ustcuna/) |
 | Stable Diffusion BoxDiff Pipeline | Training-free controlled generation with bounding boxes using [BoxDiff](https://github.com/showlab/BoxDiff) | [Stable Diffusion BoxDiff Pipeline](#stable-diffusion-boxdiff) | - | [Jingyang Zhang](https://github.com/zjysteven/) |
+|   FRESCO V2V Pipeline                                                                                                    | Implementation of [[CVPR 2024] FRESCO: Spatial-Temporal Correspondence for Zero-Shot Video Translation](https://arxiv.org/abs/2403.12962)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [FRESCO V2V Pipeline](#fresco)      | - |              [Yifan Zhou](https://github.com/SingleZombie) |
 
 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.
 
@@ -239,12 +240,12 @@ pipeline_output = pipe(
     # denoising_steps=10,     # (optional) Number of denoising steps of each inference pass. Default: 10.
     # ensemble_size=10,       # (optional) Number of inference passes in the ensemble. Default: 10.
     # ------------------------------------------------
-    
+
     # ----- recommended setting for LCM version ------
     # denoising_steps=4,
     # ensemble_size=5,
     # -------------------------------------------------
-    
+
     # processing_res=768,     # (optional) Maximum resolution of processing. If set to 0: will not resize at all. Defaults to 768.
     # match_input_res=True,   # (optional) Resize depth prediction to match input resolution.
     # batch_size=0,           # (optional) Inference batch size, no bigger than `num_ensemble`. If set to 0, the script will automatically decide the proper batch size. Defaults to 0.
@@ -1031,7 +1032,7 @@ image = pipe().images[0]
 
 Make sure you have @crowsonkb's <https://github.com/crowsonkb/k-diffusion> installed:
 
-```
+```sh
 pip install k-diffusion
 ```
 
@@ -1853,13 +1854,13 @@ To use this pipeline, you need to:
 
 You can simply use pip to install IPEX with the latest version.
 
-```python
+```sh
 python -m pip install intel_extension_for_pytorch
 ```
 
 **Note:** To install a specific version, run with the following command:
 
-```
+```sh
 python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
 ```
 
@@ -1957,13 +1958,13 @@ To use this pipeline, you need to:
 
 You can simply use pip to install IPEX with the latest version.
 
-```python
+```sh
 python -m pip install intel_extension_for_pytorch
 ```
 
 **Note:** To install a specific version, run with the following command:
 
-```
+```sh
 python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
 ```
 
@@ -3009,8 +3010,8 @@ This code implements a pipeline for the Stable Diffusion model, enabling the div
 
 ### Sample Code
 
-```
-from from examples.community.regional_prompting_stable_diffusion import RegionalPromptingStableDiffusionPipeline
+```py
+from examples.community.regional_prompting_stable_diffusion import RegionalPromptingStableDiffusionPipeline
 pipe = RegionalPromptingStableDiffusionPipeline.from_single_file(model_path, vae=vae)
 
 rp_args = {
@@ -4035,6 +4036,93 @@ onestep_image = pipe(prompt, num_inference_steps=1).images[0]
 multistep_image = pipe(prompt, num_inference_steps=4).images[0]
 ```
 
+### FRESCO
+
+This is the Diffusers implementation of zero-shot video-to-video translation pipeline [FRESCO](https://github.com/williamyang1991/FRESCO) (without Ebsynth postprocessing and background smooth). To run the code, please install gmflow. Then modify the path in `gmflow_dir`. After that, you can run the pipeline with:
+
+```py
+from PIL import Image
+import cv2
+import torch
+import numpy as np
+
+from diffusers import ControlNetModel,DDIMScheduler, DiffusionPipeline
+import sys
+gmflow_dir = "/path/to/gmflow"
+sys.path.insert(0, gmflow_dir)
+
+def video_to_frame(video_path: str, interval: int):
+    vidcap = cv2.VideoCapture(video_path)
+    success = True
+
+    count = 0
+    res = []
+    while success:
+        count += 1
+        success, image = vidcap.read()
+        if count % interval != 1:
+            continue
+        if image is not None:
+            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+            res.append(image)
+            if len(res) >= 8:
+                break
+
+    vidcap.release()
+    return res
+
+
+input_video_path = 'https://github.com/williamyang1991/FRESCO/raw/main/data/car-turn.mp4'
+output_video_path = 'car.gif'
+
+# You can use any fintuned SD here
+model_path = 'SG161222/Realistic_Vision_V2.0'
+
+prompt = 'a red car turns in the winter'
+a_prompt = ', RAW photo, subject, (high detailed skin:1.2), 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3, '
+n_prompt = '(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime, mutated hands and fingers:1.4), (deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, disconnected limbs, mutation, mutated, ugly, disgusting, amputation'
+
+input_interval = 5
+frames = video_to_frame(
+    input_video_path, input_interval)
+
+control_frames = []
+# get canny image
+for frame in frames:
+    image = cv2.Canny(frame, 50, 100)
+    np_image = np.array(image)
+    np_image = np_image[:, :, None]
+    np_image = np.concatenate([np_image, np_image, np_image], axis=2)
+    canny_image = Image.fromarray(np_image)
+    control_frames.append(canny_image)
+
+# You can use any ControlNet here
+controlnet = ControlNetModel.from_pretrained(
+    "lllyasviel/sd-controlnet-canny").to('cuda')
+
+pipe = DiffusionPipeline.from_pretrained(
+    model_path, controlnet=controlnet, custom_pipeline='fresco_v2v').to('cuda')
+pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+
+generator = torch.manual_seed(0)
+frames = [Image.fromarray(frame) for frame in frames]
+
+output_frames = pipe(
+    prompt + a_prompt,
+    frames,
+    control_frames,
+    num_inference_steps=20,
+    strength=0.75,
+    controlnet_conditioning_scale=0.7,
+    generator=generator,
+    negative_prompt=n_prompt
+).images
+
+output_frames[0].save(output_video_path, save_all=True,
+                 append_images=output_frames[1:], duration=100, loop=0)
+
+```
+
 # Perturbed-Attention Guidance
 
 [Project](https://ku-cvlab.github.io/Perturbed-Attention-Guidance/) / [arXiv](https://arxiv.org/abs/2403.17377) / [GitHub](https://github.com/KU-CVLAB/Perturbed-Attention-Guidance)
@@ -4043,7 +4131,7 @@ This implementation is based on [Diffusers](https://huggingface.co/docs/diffuser
 
 ## Example Usage
 
-```
+```py
 import os
 import torch
 

examples/community/kohya_hires_fix.py

@@ -0,0 +1,468 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from diffusers.configuration_utils import register_to_config
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.models.autoencoders import AutoencoderKL
+from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel, UNet2DConditionOutput
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipeline
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class UNet2DConditionModelHighResFix(UNet2DConditionModel):
+    r"""
+    A conditional 2D UNet model that applies Kohya fix proposed for high resolution image generation.
+
+    This model inherits from [`UNet2DConditionModel`]. Check the superclass documentation for learning about all the parameters.
+
+    Parameters:
+        high_res_fix (`List[Dict]`, *optional*, defaults to `[{'timestep': 600, 'scale_factor': 0.5, 'block_num': 1}]`):
+            Enables Kohya fix for high resolution generation. The activation maps are scaled based on the scale_factor up to the timestep at specified block_num.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(self, high_res_fix: List[Dict] = [{"timestep": 600, "scale_factor": 0.5, "block_num": 1}], **kwargs):
+        super().__init__(**kwargs)
+        if high_res_fix:
+            self.config.high_res_fix = sorted(high_res_fix, key=lambda x: x["timestep"], reverse=True)
+
+    @classmethod
+    def _resize(cls, sample, target=None, scale_factor=1, mode="bicubic"):
+        dtype = sample.dtype
+        if dtype == torch.bfloat16:
+            sample = sample.to(torch.float32)
+
+        if target is not None:
+            if sample.shape[-2:] != target.shape[-2:]:
+                sample = nn.functional.interpolate(sample, size=target.shape[-2:], mode=mode, align_corners=False)
+        elif scale_factor != 1:
+            sample = nn.functional.interpolate(sample, scale_factor=scale_factor, mode=mode, align_corners=False)
+
+        return sample.to(dtype)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_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).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        t_emb = self.get_time_embed(sample=sample, timestep=timestep)
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
+        if class_emb is not None:
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        aug_emb = self.get_aug_embed(
+            emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+        if self.config.addition_embed_type == "image_hint":
+            aug_emb, hint = aug_emb
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        encoder_hidden_states = self.process_encoder_hidden_states(
+            encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # 3. down
+        # we're popping the `scale` instead of getting it because otherwise `scale` will be propagated
+        # to the internal blocks and will raise deprecation warnings. this will be confusing for our users.
+        if cross_attention_kwargs is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            lora_scale = cross_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        for down_i, downsample_block in enumerate(self.down_blocks):
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+            # kohya high res fix
+            if self.config.high_res_fix:
+                for high_res_fix in self.config.high_res_fix:
+                    if timestep > high_res_fix["timestep"] and down_i == high_res_fix["block_num"]:
+                        sample = self.__class__._resize(sample, scale_factor=high_res_fix["scale_factor"])
+                        break
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # up scaling of kohya high res fix
+            if self.config.high_res_fix is not None:
+                if res_samples[0].shape[-2:] != sample.shape[-2:]:
+                    sample = self.__class__._resize(sample, target=res_samples[0])
+                    res_samples_up_sampled = (res_samples[0],)
+                    for res_sample in res_samples[1:]:
+                        res_samples_up_sampled += (self.__class__._resize(res_sample, target=res_samples[0]),)
+                    res_samples = res_samples_up_sampled
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                )
+
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)
+
+    @classmethod
+    def from_unet(cls, unet: UNet2DConditionModel, high_res_fix: list):
+        config = dict((unet.config))
+        config["high_res_fix"] = high_res_fix
+        unet_high_res = cls(**config)
+        unet_high_res.load_state_dict(unet.state_dict())
+        unet_high_res.to(unet.dtype)
+        return unet_high_res
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import DiffusionPipeline
+
+        >>> pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4",
+                                         custom_pipeline="kohya_hires_fix",
+                                         torch_dtype=torch.float16,
+                                         high_res_fix=[{'timestep': 600,
+                                                        'scale_factor': 0.5,
+                                                        'block_num': 1}])
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "a photo of an astronaut riding a horse on mars"
+        >>> image = pipe(prompt, height=1000, width=1600).images[0]
+        ```
+"""
+
+
+class StableDiffusionHighResFixPipeline(StableDiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion with Kohya fix for high resolution generation.
+
+    This model inherits from [`StableDiffusionPipeline`]. Check the superclass documentation for the generic methods.
+
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+        text_encoder ([`~transformers.CLIPTextModel`]):
+            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+        tokenizer ([`~transformers.CLIPTokenizer`]):
+            A `CLIPTokenizer` to tokenize text.
+        unet ([`UNet2DConditionModel`]):
+            A `UNet2DConditionModel` to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+            about a model's potential harms.
+        feature_extractor ([`~transformers.CLIPImageProcessor`]):
+            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+        high_res_fix (`List[Dict]`, *optional*, defaults to `[{'timestep': 600, 'scale_factor': 0.5, 'block_num': 1}]`):
+            Enables Kohya fix for high resolution generation. The activation maps are scaled based on the scale_factor up to the timestep at specified block_num.
+    """
+
+    model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+    _exclude_from_cpu_offload = ["safety_checker"]
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPImageProcessor,
+        image_encoder: CLIPVisionModelWithProjection = None,
+        requires_safety_checker: bool = True,
+        high_res_fix: List[Dict] = [{"timestep": 600, "scale_factor": 0.5, "block_num": 1}],
+    ):
+        super().__init__(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
+            requires_safety_checker=requires_safety_checker,
+        )
+
+        unet = UNet2DConditionModelHighResFix.from_unet(unet=unet, high_res_fix=high_res_fix)
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.register_to_config(requires_safety_checker=requires_safety_checker)

examples/community/latent_consistency_img2img.py

@@ -565,7 +565,7 @@ class LCMSchedulerWithTimestamp(SchedulerMixin, ConfigMixin):
             # Glide cosine schedule
             self.betas = betas_for_alpha_bar(num_train_timesteps)
         else:
-            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
 
         # Rescale for zero SNR
         if rescale_betas_zero_snr:

examples/community/latent_consistency_txt2img.py

@@ -477,7 +477,7 @@ class LCMScheduler(SchedulerMixin, ConfigMixin):
             # Glide cosine schedule
             self.betas = betas_for_alpha_bar(num_train_timesteps)
         else:
-            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
 
         # Rescale for zero SNR
         if rescale_betas_zero_snr:

examples/community/llm_grounded_diffusion.py

@@ -1524,35 +1524,35 @@ class LLMGroundedDiffusionPipeline(
         assert emb.shape == (w.shape[0], embedding_dim)
         return emb
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.guidance_scale
     @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.guidance_scale
     def guidance_scale(self):
         return self._guidance_scale
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.guidance_rescale
     @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.guidance_rescale
     def guidance_rescale(self):
         return self._guidance_rescale
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.clip_skip
     @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.clip_skip
     def clip_skip(self):
         return self._clip_skip
 
     # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
     # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
     # corresponds to doing no classifier free guidance.
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.do_classifier_free_guidance
     @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.do_classifier_free_guidance
     def do_classifier_free_guidance(self):
         return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.cross_attention_kwargs
     @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.cross_attention_kwargs
     def cross_attention_kwargs(self):
         return self._cross_attention_kwargs
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.num_timesteps
     @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.num_timesteps
     def num_timesteps(self):
         return self._num_timesteps

examples/community/marigold_depth_estimation.py

@@ -43,7 +43,7 @@ from diffusers.utils import BaseOutput, check_min_version
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.25.0")
+check_min_version("0.29.0.dev0")
 
 
 class MarigoldDepthOutput(BaseOutput):

examples/community/scheduling_ufogen.py

@@ -218,7 +218,7 @@ class UFOGenScheduler(SchedulerMixin, ConfigMixin):
             betas = torch.linspace(-6, 6, num_train_timesteps)
             self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start
         else:
-            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
 
         # Rescale for zero SNR
         if rescale_betas_zero_snr:

examples/consistency_distillation/train_lcm_distill_lora_sd_wds.py

@@ -73,7 +73,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/consistency_distillation/train_lcm_distill_lora_sdxl.py

@@ -66,7 +66,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/consistency_distillation/train_lcm_distill_lora_sdxl_wds.py

@@ -79,7 +79,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/consistency_distillation/train_lcm_distill_sd_wds.py

@@ -72,7 +72,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/consistency_distillation/train_lcm_distill_sdxl_wds.py

@@ -78,7 +78,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/controlnet/train_controlnet.py

@@ -60,7 +60,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/controlnet/train_controlnet_flax.py

@@ -60,7 +60,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = logging.getLogger(__name__)
 

examples/controlnet/train_controlnet_sdxl.py

@@ -61,7 +61,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 if is_torch_npu_available():

examples/custom_diffusion/train_custom_diffusion.py

@@ -63,7 +63,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/dreambooth/train_dreambooth.py

@@ -63,7 +63,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/dreambooth/train_dreambooth_flax.py

@@ -35,7 +35,7 @@ from diffusers.utils import check_min_version
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 # Cache compiled models across invocations of this script.
 cc.initialize_cache(os.path.expanduser("~/.cache/jax/compilation_cache"))

examples/dreambooth/train_dreambooth_lora.py

@@ -70,7 +70,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -78,7 +78,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/instruct_pix2pix/train_instruct_pix2pix.py

@@ -57,7 +57,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/instruct_pix2pix/train_instruct_pix2pix_sdxl.py

@@ -60,7 +60,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py

@@ -52,7 +52,7 @@ if is_wandb_available():
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 
@@ -896,7 +896,6 @@ def main():
         images = []
         if args.validation_prompts is not None:
             logger.info("Running inference for collecting generated images...")
-            pipeline = pipeline.to(accelerator.device)
             pipeline.torch_dtype = weight_dtype
             pipeline.set_progress_bar_config(disable=True)
             pipeline.enable_model_cpu_offload()

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py

@@ -46,7 +46,7 @@ from diffusers.utils import check_min_version, is_wandb_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py

@@ -46,7 +46,7 @@ from diffusers.utils import check_min_version, is_wandb_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/kandinsky2_2/text_to_image/train_text_to_image_prior.py

@@ -51,7 +51,7 @@ if is_wandb_available():
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/t2i_adapter/train_t2i_adapter_sdxl.py

@@ -60,7 +60,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/text_to_image/train_text_to_image.py

@@ -57,7 +57,7 @@ if is_wandb_available():
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/text_to_image/train_text_to_image_flax.py

@@ -49,7 +49,7 @@ from diffusers.utils import check_min_version
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = logging.getLogger(__name__)
 

examples/text_to_image/train_text_to_image_lora.py

@@ -52,8 +52,11 @@ from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module
 
 
+if is_wandb_available():
+    import wandb
+
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 
@@ -99,6 +102,48 @@ These are LoRA adaption weights for {base_model}. The weights were fine-tuned on
     model_card.save(os.path.join(repo_folder, "README.md"))
 
 
+def log_validation(
+    pipeline,
+    args,
+    accelerator,
+    epoch,
+    is_final_validation=False,
+):
+    logger.info(
+        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
+        f" {args.validation_prompt}."
+    )
+    pipeline = pipeline.to(accelerator.device)
+    pipeline.set_progress_bar_config(disable=True)
+    generator = torch.Generator(device=accelerator.device)
+    if args.seed is not None:
+        generator = generator.manual_seed(args.seed)
+    images = []
+    if torch.backends.mps.is_available():
+        autocast_ctx = nullcontext()
+    else:
+        autocast_ctx = torch.autocast(accelerator.device.type)
+
+    with autocast_ctx:
+        for _ in range(args.num_validation_images):
+            images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0])
+
+    for tracker in accelerator.trackers:
+        phase_name = "test" if is_final_validation else "validation"
+        if tracker.name == "tensorboard":
+            np_images = np.stack([np.asarray(img) for img in images])
+            tracker.writer.add_images(phase_name, np_images, epoch, dataformats="NHWC")
+        if tracker.name == "wandb":
+            tracker.log(
+                {
+                    phase_name: [
+                        wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
+                    ]
+                }
+            )
+    return images
+
+
 def parse_args():
     parser = argparse.ArgumentParser(description="Simple example of a training script.")
     parser.add_argument(
@@ -414,11 +459,6 @@ def main():
     if torch.backends.mps.is_available():
         accelerator.native_amp = False
 
-    if args.report_to == "wandb":
-        if not is_wandb_available():
-            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
-        import wandb
-
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -657,17 +697,22 @@ def main():
     )
 
     # Scheduler and math around the number of training steps.
-    overrode_max_train_steps = False
-    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * num_update_steps_per_epoch * accelerator.num_processes
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
     )
 
     # Prepare everything with our `accelerator`.
@@ -677,8 +722,14 @@ def main():
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps * accelerator.num_processes:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 
@@ -864,10 +915,6 @@ def main():
 
         if accelerator.is_main_process:
             if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
-                logger.info(
-                    f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
-                    f" {args.validation_prompt}."
-                )
                 # create pipeline
                 pipeline = DiffusionPipeline.from_pretrained(
                     args.pretrained_model_name_or_path,
@@ -876,38 +923,7 @@ def main():
                     variant=args.variant,
                     torch_dtype=weight_dtype,
                 )
-                pipeline = pipeline.to(accelerator.device)
-                pipeline.set_progress_bar_config(disable=True)
-
-                # run inference
-                generator = torch.Generator(device=accelerator.device)
-                if args.seed is not None:
-                    generator = generator.manual_seed(args.seed)
-                images = []
-                if torch.backends.mps.is_available():
-                    autocast_ctx = nullcontext()
-                else:
-                    autocast_ctx = torch.autocast(accelerator.device.type)
-
-                with autocast_ctx:
-                    for _ in range(args.num_validation_images):
-                        images.append(
-                            pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]
-                        )
-
-                for tracker in accelerator.trackers:
-                    if tracker.name == "tensorboard":
-                        np_images = np.stack([np.asarray(img) for img in images])
-                        tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
-                    if tracker.name == "wandb":
-                        tracker.log(
-                            {
-                                "validation": [
-                                    wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
-                                    for i, image in enumerate(images)
-                                ]
-                            }
-                        )
+                images = log_validation(pipeline, args, accelerator, epoch)
 
                 del pipeline
                 torch.cuda.empty_cache()
@@ -925,6 +941,22 @@ def main():
             safe_serialization=True,
         )
 
+        # Final inference
+        # Load previous pipeline
+        if args.validation_prompt is not None:
+            pipeline = DiffusionPipeline.from_pretrained(
+                args.pretrained_model_name_or_path,
+                revision=args.revision,
+                variant=args.variant,
+                torch_dtype=weight_dtype,
+            )
+
+            # load attention processors
+            pipeline.load_lora_weights(args.output_dir)
+
+            # run inference
+            images = log_validation(pipeline, args, accelerator, epoch, is_final_validation=True)
+
         if args.push_to_hub:
             save_model_card(
                 repo_id,
@@ -940,51 +972,6 @@ def main():
                 ignore_patterns=["step_*", "epoch_*"],
             )
 
-        # Final inference
-        # Load previous pipeline
-        if args.validation_prompt is not None:
-            pipeline = DiffusionPipeline.from_pretrained(
-                args.pretrained_model_name_or_path,
-                revision=args.revision,
-                variant=args.variant,
-                torch_dtype=weight_dtype,
-            )
-            pipeline = pipeline.to(accelerator.device)
-
-            # load attention processors
-            pipeline.load_lora_weights(args.output_dir)
-
-            # run inference
-            generator = torch.Generator(device=accelerator.device)
-            if args.seed is not None:
-                generator = generator.manual_seed(args.seed)
-            images = []
-            if torch.backends.mps.is_available():
-                autocast_ctx = nullcontext()
-            else:
-                autocast_ctx = torch.autocast(accelerator.device.type)
-
-            with autocast_ctx:
-                for _ in range(args.num_validation_images):
-                    images.append(
-                        pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]
-                    )
-
-            for tracker in accelerator.trackers:
-                if len(images) != 0:
-                    if tracker.name == "tensorboard":
-                        np_images = np.stack([np.asarray(img) for img in images])
-                        tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
-                    if tracker.name == "wandb":
-                        tracker.log(
-                            {
-                                "test": [
-                                    wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
-                                    for i, image in enumerate(images)
-                                ]
-                            }
-                        )
-
     accelerator.end_training()
 
 

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -64,8 +64,11 @@ from diffusers.utils.import_utils import is_torch_npu_available, is_xformers_ava
 from diffusers.utils.torch_utils import is_compiled_module
 
 
+if is_wandb_available():
+    import wandb
+
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 if is_torch_npu_available():
@@ -119,6 +122,47 @@ Special VAE used for training: {vae_path}.
     model_card.save(os.path.join(repo_folder, "README.md"))
 
 
+def log_validation(
+    pipeline,
+    args,
+    accelerator,
+    epoch,
+    is_final_validation=False,
+):
+    logger.info(
+        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
+        f" {args.validation_prompt}."
+    )
+    pipeline = pipeline.to(accelerator.device)
+    pipeline.set_progress_bar_config(disable=True)
+
+    # run inference
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    pipeline_args = {"prompt": args.validation_prompt}
+    if torch.backends.mps.is_available():
+        autocast_ctx = nullcontext()
+    else:
+        autocast_ctx = torch.autocast(accelerator.device.type)
+
+    with autocast_ctx:
+        images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
+
+    for tracker in accelerator.trackers:
+        phase_name = "test" if is_final_validation else "validation"
+        if tracker.name == "tensorboard":
+            np_images = np.stack([np.asarray(img) for img in images])
+            tracker.writer.add_images(phase_name, np_images, epoch, dataformats="NHWC")
+        if tracker.name == "wandb":
+            tracker.log(
+                {
+                    phase_name: [
+                        wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
+                    ]
+                }
+            )
+    return images
+
+
 def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
@@ -523,11 +567,6 @@ def main(args):
         kwargs_handlers=[kwargs],
     )
 
-    if args.report_to == "wandb":
-        if not is_wandb_available():
-            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
-        import wandb
-
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -1196,10 +1235,6 @@ def main(args):
 
         if accelerator.is_main_process:
             if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
-                logger.info(
-                    f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
-                    f" {args.validation_prompt}."
-                )
                 # create pipeline
                 pipeline = StableDiffusionXLPipeline.from_pretrained(
                     args.pretrained_model_name_or_path,
@@ -1212,36 +1247,7 @@ def main(args):
                     torch_dtype=weight_dtype,
                 )
 
-                pipeline = pipeline.to(accelerator.device)
-                pipeline.set_progress_bar_config(disable=True)
-
-                # run inference
-                generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
-                pipeline_args = {"prompt": args.validation_prompt}
-                if torch.backends.mps.is_available():
-                    autocast_ctx = nullcontext()
-                else:
-                    autocast_ctx = torch.autocast(accelerator.device.type)
-
-                with autocast_ctx:
-                    images = [
-                        pipeline(**pipeline_args, generator=generator).images[0]
-                        for _ in range(args.num_validation_images)
-                    ]
-
-                for tracker in accelerator.trackers:
-                    if tracker.name == "tensorboard":
-                        np_images = np.stack([np.asarray(img) for img in images])
-                        tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
-                    if tracker.name == "wandb":
-                        tracker.log(
-                            {
-                                "validation": [
-                                    wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
-                                    for i, image in enumerate(images)
-                                ]
-                            }
-                        )
+                images = log_validation(pipeline, args, accelerator, epoch)
 
                 del pipeline
                 torch.cuda.empty_cache()
@@ -1288,33 +1294,13 @@ def main(args):
             variant=args.variant,
             torch_dtype=weight_dtype,
         )
-        pipeline = pipeline.to(accelerator.device)
 
         # load attention processors
         pipeline.load_lora_weights(args.output_dir)
 
         # run inference
-        images = []
         if args.validation_prompt and args.num_validation_images > 0:
-            generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
-            images = [
-                pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
-                for _ in range(args.num_validation_images)
-            ]
-
-            for tracker in accelerator.trackers:
-                if tracker.name == "tensorboard":
-                    np_images = np.stack([np.asarray(img) for img in images])
-                    tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
-                if tracker.name == "wandb":
-                    tracker.log(
-                        {
-                            "test": [
-                                wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
-                                for i, image in enumerate(images)
-                            ]
-                        }
-                    )
+            images = log_validation(pipeline, args, accelerator, epoch, is_final_validation=True)
 
         if args.push_to_hub:
             save_model_card(

examples/text_to_image/train_text_to_image_sdxl.py

@@ -55,7 +55,7 @@ from diffusers.utils.torch_utils import is_compiled_module
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 if is_torch_npu_available():

examples/textual_inversion/textual_inversion.py

@@ -81,7 +81,7 @@ else:
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/textual_inversion/textual_inversion_flax.py

@@ -56,7 +56,7 @@ else:
 # ------------------------------------------------------------------------------
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = logging.getLogger(__name__)
 

examples/textual_inversion/textual_inversion_sdxl.py

@@ -76,7 +76,7 @@ else:
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__)
 

examples/unconditional_image_generation/train_unconditional.py

@@ -29,7 +29,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/vqgan/train_vqgan.py

@@ -50,7 +50,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.27.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/wuerstchen/text_to_image/train_text_to_image_lora_prior.py

@@ -50,7 +50,7 @@ if is_wandb_available():
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/wuerstchen/text_to_image/train_text_to_image_prior.py

@@ -51,7 +51,7 @@ if is_wandb_available():
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.29.0.dev0")
 
 logger = get_logger(__name__, log_level="INFO")
 

setup.py

@@ -254,14 +254,14 @@ version_range_max = max(sys.version_info[1], 10) + 1
 
 setup(
     name="diffusers",
-    version="0.28.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="0.29.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     description="State-of-the-art diffusion in PyTorch and JAX.",
     long_description=open("README.md", "r", encoding="utf-8").read(),
     long_description_content_type="text/markdown",
     keywords="deep learning diffusion jax pytorch stable diffusion audioldm",
     license="Apache 2.0 License",
     author="The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/diffusers/graphs/contributors)",
-    author_email="patrick@huggingface.co",
+    author_email="diffusers@huggingface.co",
     url="https://github.com/huggingface/diffusers",
     package_dir={"": "src"},
     packages=find_packages("src"),

src/diffusers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.28.0.dev0"
+__version__ = "0.29.0.dev0"
 
 from typing import TYPE_CHECKING
 
@@ -82,11 +82,14 @@ else:
             "ConsistencyDecoderVAE",
             "ControlNetModel",
             "ControlNetXSAdapter",
+            "DiTTransformer2DModel",
+            "HunyuanDiT2DModel",
             "I2VGenXLUNet",
             "Kandinsky3UNet",
             "ModelMixin",
             "MotionAdapter",
             "MultiAdapter",
+            "PixArtTransformer2DModel",
             "PriorTransformer",
             "StableCascadeUNet",
             "T2IAdapter",
@@ -227,6 +230,7 @@ else:
             "BlipDiffusionPipeline",
             "CLIPImageProjection",
             "CycleDiffusionPipeline",
+            "HunyuanDiTPipeline",
             "I2VGenXLPipeline",
             "IFImg2ImgPipeline",
             "IFImg2ImgSuperResolutionPipeline",
@@ -259,6 +263,8 @@ else:
             "LDMTextToImagePipeline",
             "LEditsPPPipelineStableDiffusion",
             "LEditsPPPipelineStableDiffusionXL",
+            "MarigoldDepthPipeline",
+            "MarigoldNormalsPipeline",
             "MusicLDMPipeline",
             "PaintByExamplePipeline",
             "PIAPipeline",
@@ -482,11 +488,14 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ConsistencyDecoderVAE,
             ControlNetModel,
             ControlNetXSAdapter,
+            DiTTransformer2DModel,
+            HunyuanDiT2DModel,
             I2VGenXLUNet,
             Kandinsky3UNet,
             ModelMixin,
             MotionAdapter,
             MultiAdapter,
+            PixArtTransformer2DModel,
             PriorTransformer,
             T2IAdapter,
             T5FilmDecoder,
@@ -605,6 +614,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDMPipeline,
             CLIPImageProjection,
             CycleDiffusionPipeline,
+            HunyuanDiTPipeline,
             I2VGenXLPipeline,
             IFImg2ImgPipeline,
             IFImg2ImgSuperResolutionPipeline,
@@ -637,6 +647,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LDMTextToImagePipeline,
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
+            MarigoldDepthPipeline,
+            MarigoldNormalsPipeline,
             MusicLDMPipeline,
             PaintByExamplePipeline,
             PIAPipeline,

src/diffusers/commands/env.py

@@ -105,6 +105,8 @@ class EnvironmentCommand(BaseDiffusersCLICommand):
 
             xformers_version = xformers.__version__
 
+        platform_info = platform.platform()
+
         is_notebook_str = "Yes" if is_notebook() else "No"
 
         is_google_colab_str = "Yes" if is_google_colab() else "No"
@@ -152,7 +154,7 @@ class EnvironmentCommand(BaseDiffusersCLICommand):
 
         info = {
             "🤗 Diffusers version": version,
-            "Platform": f"{platform.freedesktop_os_release().get('PRETTY_NAME', None)} - {platform.platform()}",
+            "Platform": platform_info,
             "Running on a notebook?": is_notebook_str,
             "Running on Google Colab?": is_google_colab_str,
             "Python version": platform.python_version(),

src/diffusers/configuration_utils.py

@@ -706,3 +706,20 @@ def flax_register_to_config(cls):
 
     cls.__init__ = init
     return cls
+
+
+class LegacyConfigMixin(ConfigMixin):
+    r"""
+    A subclass of `ConfigMixin` to resolve class mapping from legacy classes (like `Transformer2DModel`) to more
+    pipeline-specific classes (like `DiTTransformer2DModel`).
+    """
+
+    @classmethod
+    def from_config(cls, config: Union[FrozenDict, Dict[str, Any]] = None, return_unused_kwargs=False, **kwargs):
+        # To prevent depedency import problem.
+        from .models.model_loading_utils import _fetch_remapped_cls_from_config
+
+        # resolve remapping
+        remapped_class = _fetch_remapped_cls_from_config(config, cls)
+
+        return remapped_class.from_config(config, return_unused_kwargs, **kwargs)

src/diffusers/loaders/lora.py

@@ -22,17 +22,14 @@ import torch
 from huggingface_hub import model_info
 from huggingface_hub.constants import HF_HUB_OFFLINE
 from huggingface_hub.utils import validate_hf_hub_args
-from packaging import version
 from torch import nn
 
-from .. import __version__
-from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_state_dict
+from ..models.modeling_utils import load_state_dict
 from ..utils import (
     USE_PEFT_BACKEND,
     _get_model_file,
     convert_state_dict_to_diffusers,
     convert_state_dict_to_peft,
-    convert_unet_state_dict_to_peft,
     delete_adapter_layers,
     get_adapter_name,
     get_peft_kwargs,
@@ -119,13 +116,10 @@ class LoraLoaderMixin:
         if not is_correct_format:
             raise ValueError("Invalid LoRA checkpoint.")
 
-        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
-
         self.load_lora_into_unet(
             state_dict,
             network_alphas=network_alphas,
             unet=getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet,
-            low_cpu_mem_usage=low_cpu_mem_usage,
             adapter_name=adapter_name,
             _pipeline=self,
         )
@@ -136,7 +130,6 @@ class LoraLoaderMixin:
             if not hasattr(self, "text_encoder")
             else self.text_encoder,
             lora_scale=self.lora_scale,
-            low_cpu_mem_usage=low_cpu_mem_usage,
             adapter_name=adapter_name,
             _pipeline=self,
         )
@@ -193,16 +186,8 @@ class LoraLoaderMixin:
                 allowed by Git.
             subfolder (`str`, *optional*, defaults to `""`):
                 The subfolder location of a model file within a larger model repository on the Hub or locally.
-            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
-                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
-                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
-                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
-                argument to `True` will raise an error.
-            mirror (`str`, *optional*):
-                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
-                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
-                information.
-
+            weight_name (`str`, *optional*, defaults to None):
+                Name of the serialized state dict file.
         """
         # Load the main state dict first which has the LoRA layers for either of
         # UNet and text encoder or both.
@@ -383,9 +368,7 @@ class LoraLoaderMixin:
         return (is_model_cpu_offload, is_sequential_cpu_offload)
 
     @classmethod
-    def load_lora_into_unet(
-        cls, state_dict, network_alphas, unet, low_cpu_mem_usage=None, adapter_name=None, _pipeline=None
-    ):
+    def load_lora_into_unet(cls, state_dict, network_alphas, unet, adapter_name=None, _pipeline=None):
         """
         This will load the LoRA layers specified in `state_dict` into `unet`.
 
@@ -395,14 +378,11 @@ class LoraLoaderMixin:
                 into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                 encoder lora layers.
             network_alphas (`Dict[str, float]`):
-                See `LoRALinearLayer` for more details.
+                The value of the network alpha used for stable learning and preventing underflow. This value has the
+                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
+                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
             unet (`UNet2DConditionModel`):
                 The UNet model to load the LoRA layers into.
-            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
-                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
-                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
-                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
-                argument to `True` will raise an error.
             adapter_name (`str`, *optional*):
                 Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                 `default_{i}` where i is the total number of adapters being loaded.
@@ -410,94 +390,18 @@ class LoraLoaderMixin:
         if not USE_PEFT_BACKEND:
             raise ValueError("PEFT backend is required for this method.")
 
-        from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
-
-        low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else _LOW_CPU_MEM_USAGE_DEFAULT
         # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
         # then the `state_dict` keys should have `cls.unet_name` and/or `cls.text_encoder_name` as
         # their prefixes.
         keys = list(state_dict.keys())
+        only_text_encoder = all(key.startswith(cls.text_encoder_name) for key in keys)
 
-        if all(key.startswith(cls.unet_name) or key.startswith(cls.text_encoder_name) for key in keys):
+        if any(key.startswith(cls.unet_name) for key in keys) and not only_text_encoder:
             # Load the layers corresponding to UNet.
             logger.info(f"Loading {cls.unet_name}.")
-
-            unet_keys = [k for k in keys if k.startswith(cls.unet_name)]
-            state_dict = {k.replace(f"{cls.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
-
-            if network_alphas is not None:
-                alpha_keys = [k for k in network_alphas.keys() if k.startswith(cls.unet_name)]
-                network_alphas = {
-                    k.replace(f"{cls.unet_name}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
-                }
-
-        else:
-            # Otherwise, we're dealing with the old format. This means the `state_dict` should only
-            # contain the module names of the `unet` as its keys WITHOUT any prefix.
-            if not USE_PEFT_BACKEND:
-                warn_message = "You have saved the LoRA weights using the old format. To convert the old LoRA weights to the new format, you can first load them in a dictionary and then create a new dictionary like the following: `new_state_dict = {f'unet.{module_name}': params for module_name, params in old_state_dict.items()}`."
-                logger.warning(warn_message)
-
-        if len(state_dict.keys()) > 0:
-            if adapter_name in getattr(unet, "peft_config", {}):
-                raise ValueError(
-                    f"Adapter name {adapter_name} already in use in the Unet - please select a new adapter name."
-                )
-
-            state_dict = convert_unet_state_dict_to_peft(state_dict)
-
-            if network_alphas is not None:
-                # The alphas state dict have the same structure as Unet, thus we convert it to peft format using
-                # `convert_unet_state_dict_to_peft` method.
-                network_alphas = convert_unet_state_dict_to_peft(network_alphas)
-
-            rank = {}
-            for key, val in state_dict.items():
-                if "lora_B" in key:
-                    rank[key] = val.shape[1]
-
-            lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict, is_unet=True)
-            if "use_dora" in lora_config_kwargs:
-                if lora_config_kwargs["use_dora"]:
-                    if is_peft_version("<", "0.9.0"):
-                        raise ValueError(
-                            "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
-                        )
-                else:
-                    if is_peft_version("<", "0.9.0"):
-                        lora_config_kwargs.pop("use_dora")
-            lora_config = LoraConfig(**lora_config_kwargs)
-
-            # adapter_name
-            if adapter_name is None:
-                adapter_name = get_adapter_name(unet)
-
-            # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
-            # otherwise loading LoRA weights will lead to an error
-            is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
-
-            inject_adapter_in_model(lora_config, unet, adapter_name=adapter_name)
-            incompatible_keys = set_peft_model_state_dict(unet, state_dict, adapter_name)
-
-            if incompatible_keys is not None:
-                # check only for unexpected keys
-                unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
-                if unexpected_keys:
-                    logger.warning(
-                        f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
-                        f" {unexpected_keys}. "
-                    )
-
-            # Offload back.
-            if is_model_cpu_offload:
-                _pipeline.enable_model_cpu_offload()
-            elif is_sequential_cpu_offload:
-                _pipeline.enable_sequential_cpu_offload()
-            # Unsafe code />
-
-        unet.load_attn_procs(
-            state_dict, network_alphas=network_alphas, low_cpu_mem_usage=low_cpu_mem_usage, _pipeline=_pipeline
-        )
+            unet.load_attn_procs(
+                state_dict, network_alphas=network_alphas, adapter_name=adapter_name, _pipeline=_pipeline
+            )
 
     @classmethod
     def load_lora_into_text_encoder(
@@ -507,7 +411,6 @@ class LoraLoaderMixin:
         text_encoder,
         prefix=None,
         lora_scale=1.0,
-        low_cpu_mem_usage=None,
         adapter_name=None,
         _pipeline=None,
     ):
@@ -527,11 +430,6 @@ class LoraLoaderMixin:
             lora_scale (`float`):
                 How much to scale the output of the lora linear layer before it is added with the output of the regular
                 lora layer.
-            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
-                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
-                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
-                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
-                argument to `True` will raise an error.
             adapter_name (`str`, *optional*):
                 Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                 `default_{i}` where i is the total number of adapters being loaded.
@@ -541,8 +439,6 @@ class LoraLoaderMixin:
 
         from peft import LoraConfig
 
-        low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else _LOW_CPU_MEM_USAGE_DEFAULT
-
         # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
         # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
         # their prefixes.
@@ -625,9 +521,7 @@ class LoraLoaderMixin:
                 # Unsafe code />
 
     @classmethod
-    def load_lora_into_transformer(
-        cls, state_dict, network_alphas, transformer, low_cpu_mem_usage=None, adapter_name=None, _pipeline=None
-    ):
+    def load_lora_into_transformer(cls, state_dict, network_alphas, transformer, adapter_name=None, _pipeline=None):
         """
         This will load the LoRA layers specified in `state_dict` into `transformer`.
 
@@ -640,19 +534,12 @@ class LoraLoaderMixin:
                 See `LoRALinearLayer` for more details.
             unet (`UNet2DConditionModel`):
                 The UNet model to load the LoRA layers into.
-            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
-                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
-                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
-                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
-                argument to `True` will raise an error.
             adapter_name (`str`, *optional*):
                 Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                 `default_{i}` where i is the total number of adapters being loaded.
         """
         from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
 
-        low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else _LOW_CPU_MEM_USAGE_DEFAULT
-
         keys = list(state_dict.keys())
 
         transformer_keys = [k for k in keys if k.startswith(cls.transformer_name)]
@@ -846,22 +733,11 @@ class LoraLoaderMixin:
         >>> ...
         ```
         """
-        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
-
         if not USE_PEFT_BACKEND:
-            if version.parse(__version__) > version.parse("0.23"):
-                logger.warning(
-                    "You are using `unload_lora_weights` to disable and unload lora weights. If you want to iteratively enable and disable adapter weights,"
-                    "you can use `pipe.enable_lora()` or `pipe.disable_lora()`. After installing the latest version of PEFT."
-                )
+            raise ValueError("PEFT backend is required for this method.")
 
-            for _, module in unet.named_modules():
-                if hasattr(module, "set_lora_layer"):
-                    module.set_lora_layer(None)
-        else:
-            recurse_remove_peft_layers(unet)
-            if hasattr(unet, "peft_config"):
-                del unet.peft_config
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        unet.unload_lora()
 
         # Safe to call the following regardless of LoRA.
         self._remove_text_encoder_monkey_patch()

src/diffusers/loaders/lora_conversion_utils.py

@@ -226,6 +226,8 @@ def _convert_kohya_lora_to_diffusers(state_dict, unet_name="unet", text_encoder_
             diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
             diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
             diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+            diffusers_name = diffusers_name.replace("text.projection", "text_projection")
+
             if "self_attn" in diffusers_name:
                 if lora_name.startswith(("lora_te_", "lora_te1_")):
                     te_state_dict[diffusers_name] = state_dict.pop(key)
@@ -243,6 +245,10 @@ def _convert_kohya_lora_to_diffusers(state_dict, unet_name="unet", text_encoder_
                 else:
                     te2_state_dict[diffusers_name] = state_dict.pop(key)
                     te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            # OneTrainer specificity
+            elif "text_projection" in diffusers_name and lora_name.startswith("lora_te2_"):
+                te2_state_dict[diffusers_name] = state_dict.pop(key)
+                te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
 
             if (is_te_dora_lora or is_te2_dora_lora) and lora_name.startswith(("lora_te_", "lora_te1_", "lora_te2_")):
                 dora_scale_key_to_replace_te = (
@@ -270,7 +276,7 @@ def _convert_kohya_lora_to_diffusers(state_dict, unet_name="unet", text_encoder_
             network_alphas.update({new_name: alpha})
 
     if len(state_dict) > 0:
-        raise ValueError(f"The following keys have not been correctly be renamed: \n\n {', '.join(state_dict.keys())}")
+        raise ValueError(f"The following keys have not been correctly renamed: \n\n {', '.join(state_dict.keys())}")
 
     logger.info("Kohya-style checkpoint detected.")
     unet_state_dict = {f"{unet_name}.{module_name}": params for module_name, params in unet_state_dict.items()}

src/diffusers/loaders/single_file.py

@@ -340,7 +340,7 @@ class FromSingleFileMixin:
             deprecate("original_config_file", "1.0.0", deprecation_message)
             original_config = original_config_file
 
-        resume_download = kwargs.pop("resume_download", False)
+        resume_download = kwargs.pop("resume_download", None)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
         token = kwargs.pop("token", None)

src/diffusers/loaders/single_file_model.py

@@ -166,7 +166,7 @@ class FromOriginalModelMixin:
                 "`from_single_file` cannot accept both `config` and `original_config` arguments. Please provide only one of these arguments"
             )
 
-        resume_download = kwargs.pop("resume_download", False)
+        resume_download = kwargs.pop("resume_download", None)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
         token = kwargs.pop("token", None)

src/diffusers/loaders/single_file_utils.py

@@ -63,7 +63,7 @@ CHECKPOINT_KEY_NAMES = {
     "controlnet": "control_model.time_embed.0.weight",
     "playground-v2-5": "edm_mean",
     "inpainting": "model.diffusion_model.input_blocks.0.0.weight",
-    "clip": "cond_stage_model.transformer.text_model.embeddings.position_ids",
+    "clip": "cond_stage_model.transformer.text_model.embeddings.position_embedding.weight",
     "clip_sdxl": "conditioner.embedders.0.transformer.text_model.embeddings.position_embedding.weight",
     "open_clip": "cond_stage_model.model.token_embedding.weight",
     "open_clip_sdxl": "conditioner.embedders.1.model.positional_embedding",

src/diffusers/loaders/unet.py

@@ -33,34 +33,32 @@ from ..models.embeddings import (
     IPAdapterPlusImageProjection,
     MultiIPAdapterImageProjection,
 )
-from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta, load_state_dict
+from ..models.modeling_utils import load_model_dict_into_meta, load_state_dict
 from ..utils import (
     USE_PEFT_BACKEND,
     _get_model_file,
+    convert_unet_state_dict_to_peft,
     delete_adapter_layers,
+    get_adapter_name,
+    get_peft_kwargs,
     is_accelerate_available,
+    is_peft_version,
     is_torch_version,
     logging,
     set_adapter_layers,
     set_weights_and_activate_adapters,
 )
+from .lora import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE, TEXT_ENCODER_NAME, UNET_NAME
 from .unet_loader_utils import _maybe_expand_lora_scales
 from .utils import AttnProcsLayers
 
 
 if is_accelerate_available():
-    from accelerate import init_empty_weights
     from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
 
 logger = logging.get_logger(__name__)
 
 
-TEXT_ENCODER_NAME = "text_encoder"
-UNET_NAME = "unet"
-
-LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
-LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
-
 CUSTOM_DIFFUSION_WEIGHT_NAME = "pytorch_custom_diffusion_weights.bin"
 CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE = "pytorch_custom_diffusion_weights.safetensors"
 
@@ -79,7 +77,8 @@ class UNet2DConditionLoadersMixin:
         Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
         defined in
         [`attention_processor.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py)
-        and be a `torch.nn.Module` class.
+        and be a `torch.nn.Module` class. Currently supported: LoRA, Custom Diffusion. For LoRA, one must install
+        `peft`: `pip install -U peft`.
 
         Parameters:
             pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
@@ -110,20 +109,20 @@ class UNet2DConditionLoadersMixin:
             token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
-            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
-                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
-                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
-                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
-                argument to `True` will raise an error.
             revision (`str`, *optional*, defaults to `"main"`):
                 The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                 allowed by Git.
             subfolder (`str`, *optional*, defaults to `""`):
                 The subfolder location of a model file within a larger model repository on the Hub or locally.
-            mirror (`str`, *optional*):
-                Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
-                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
-                information.
+            network_alphas (`Dict[str, float]`):
+                The value of the network alpha used for stable learning and preventing underflow. This value has the
+                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
+                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
+            adapter_name (`str`, *optional*, defaults to None):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            weight_name (`str`, *optional*, defaults to None):
+                Name of the serialized state dict file.
 
         Example:
 
@@ -139,9 +138,6 @@ class UNet2DConditionLoadersMixin:
         )
         ```
         """
-        from ..models.attention_processor import CustomDiffusionAttnProcessor
-        from ..models.lora import LoRACompatibleConv, LoRACompatibleLinear, LoRAConv2dLayer, LoRALinearLayer
-
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", None)
@@ -152,15 +148,9 @@ class UNet2DConditionLoadersMixin:
         subfolder = kwargs.pop("subfolder", None)
         weight_name = kwargs.pop("weight_name", None)
         use_safetensors = kwargs.pop("use_safetensors", None)
-        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
-        # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
-        # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
-        network_alphas = kwargs.pop("network_alphas", None)
-
+        adapter_name = kwargs.pop("adapter_name", None)
         _pipeline = kwargs.pop("_pipeline", None)
-
-        is_network_alphas_none = network_alphas is None
-
+        network_alphas = kwargs.pop("network_alphas", None)
         allow_pickle = False
 
         if use_safetensors is None:
@@ -216,198 +206,196 @@ class UNet2DConditionLoadersMixin:
         else:
             state_dict = pretrained_model_name_or_path_or_dict
 
-        # fill attn processors
-        lora_layers_list = []
-
-        is_lora = all(("lora" in k or k.endswith(".alpha")) for k in state_dict.keys()) and not USE_PEFT_BACKEND
         is_custom_diffusion = any("custom_diffusion" in k for k in state_dict.keys())
+        is_lora = all(("lora" in k or k.endswith(".alpha")) for k in state_dict.keys())
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
 
-        if is_lora:
-            # correct keys
-            state_dict, network_alphas = self.convert_state_dict_legacy_attn_format(state_dict, network_alphas)
-
-            if network_alphas is not None:
-                network_alphas_keys = list(network_alphas.keys())
-                used_network_alphas_keys = set()
-
-            lora_grouped_dict = defaultdict(dict)
-            mapped_network_alphas = {}
-
-            all_keys = list(state_dict.keys())
-            for key in all_keys:
-                value = state_dict.pop(key)
-                attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
-                lora_grouped_dict[attn_processor_key][sub_key] = value
-
-                # Create another `mapped_network_alphas` dictionary so that we can properly map them.
-                if network_alphas is not None:
-                    for k in network_alphas_keys:
-                        if k.replace(".alpha", "") in key:
-                            mapped_network_alphas.update({attn_processor_key: network_alphas.get(k)})
-                            used_network_alphas_keys.add(k)
-
-            if not is_network_alphas_none:
-                if len(set(network_alphas_keys) - used_network_alphas_keys) > 0:
-                    raise ValueError(
-                        f"The `network_alphas` has to be empty at this point but has the following keys \n\n {', '.join(network_alphas.keys())}"
-                    )
-
-            if len(state_dict) > 0:
-                raise ValueError(
-                    f"The `state_dict` has to be empty at this point but has the following keys \n\n {', '.join(state_dict.keys())}"
-                )
-
-            for key, value_dict in lora_grouped_dict.items():
-                attn_processor = self
-                for sub_key in key.split("."):
-                    attn_processor = getattr(attn_processor, sub_key)
-
-                # Process non-attention layers, which don't have to_{k,v,q,out_proj}_lora layers
-                # or add_{k,v,q,out_proj}_proj_lora layers.
-                rank = value_dict["lora.down.weight"].shape[0]
-
-                if isinstance(attn_processor, LoRACompatibleConv):
-                    in_features = attn_processor.in_channels
-                    out_features = attn_processor.out_channels
-                    kernel_size = attn_processor.kernel_size
-
-                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
-                    with ctx():
-                        lora = LoRAConv2dLayer(
-                            in_features=in_features,
-                            out_features=out_features,
-                            rank=rank,
-                            kernel_size=kernel_size,
-                            stride=attn_processor.stride,
-                            padding=attn_processor.padding,
-                            network_alpha=mapped_network_alphas.get(key),
-                        )
-                elif isinstance(attn_processor, LoRACompatibleLinear):
-                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
-                    with ctx():
-                        lora = LoRALinearLayer(
-                            attn_processor.in_features,
-                            attn_processor.out_features,
-                            rank,
-                            mapped_network_alphas.get(key),
-                        )
-                else:
-                    raise ValueError(f"Module {key} is not a LoRACompatibleConv or LoRACompatibleLinear module.")
-
-                value_dict = {k.replace("lora.", ""): v for k, v in value_dict.items()}
-                lora_layers_list.append((attn_processor, lora))
-
-                if low_cpu_mem_usage:
-                    device = next(iter(value_dict.values())).device
-                    dtype = next(iter(value_dict.values())).dtype
-                    load_model_dict_into_meta(lora, value_dict, device=device, dtype=dtype)
-                else:
-                    lora.load_state_dict(value_dict)
-
-        elif is_custom_diffusion:
-            attn_processors = {}
-            custom_diffusion_grouped_dict = defaultdict(dict)
-            for key, value in state_dict.items():
-                if len(value) == 0:
-                    custom_diffusion_grouped_dict[key] = {}
-                else:
-                    if "to_out" in key:
-                        attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
-                    else:
-                        attn_processor_key, sub_key = ".".join(key.split(".")[:-2]), ".".join(key.split(".")[-2:])
-                    custom_diffusion_grouped_dict[attn_processor_key][sub_key] = value
-
-            for key, value_dict in custom_diffusion_grouped_dict.items():
-                if len(value_dict) == 0:
-                    attn_processors[key] = CustomDiffusionAttnProcessor(
-                        train_kv=False, train_q_out=False, hidden_size=None, cross_attention_dim=None
-                    )
-                else:
-                    cross_attention_dim = value_dict["to_k_custom_diffusion.weight"].shape[1]
-                    hidden_size = value_dict["to_k_custom_diffusion.weight"].shape[0]
-                    train_q_out = True if "to_q_custom_diffusion.weight" in value_dict else False
-                    attn_processors[key] = CustomDiffusionAttnProcessor(
-                        train_kv=True,
-                        train_q_out=train_q_out,
-                        hidden_size=hidden_size,
-                        cross_attention_dim=cross_attention_dim,
-                    )
-                    attn_processors[key].load_state_dict(value_dict)
-        elif USE_PEFT_BACKEND:
-            # In that case we have nothing to do as loading the adapter weights is already handled above by `set_peft_model_state_dict`
-            # on the Unet
-            pass
+        if is_custom_diffusion:
+            attn_processors = self._process_custom_diffusion(state_dict=state_dict)
+        elif is_lora:
+            is_model_cpu_offload, is_sequential_cpu_offload = self._process_lora(
+                state_dict=state_dict,
+                unet_identifier_key=self.unet_name,
+                network_alphas=network_alphas,
+                adapter_name=adapter_name,
+                _pipeline=_pipeline,
+            )
         else:
             raise ValueError(
-                f"{model_file} does not seem to be in the correct format expected by LoRA or Custom Diffusion training."
+                f"{model_file} does not seem to be in the correct format expected by Custom Diffusion training."
             )
 
         # <Unsafe code
         # We can be sure that the following works as it just sets attention processors, lora layers and puts all in the same dtype
-        # Now we remove any existing hooks to
+        # Now we remove any existing hooks to `_pipeline`.
+
+        # For LoRA, the UNet is already offloaded at this stage as it is handled inside `_process_lora`.
+        if is_custom_diffusion and _pipeline is not None:
+            is_model_cpu_offload, is_sequential_cpu_offload = self._optionally_disable_offloading(_pipeline=_pipeline)
+
+            # only custom diffusion needs to set attn processors
+            self.set_attn_processor(attn_processors)
+            self.to(dtype=self.dtype, device=self.device)
+
+        # Offload back.
+        if is_model_cpu_offload:
+            _pipeline.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            _pipeline.enable_sequential_cpu_offload()
+        # Unsafe code />
+
+    def _process_custom_diffusion(self, state_dict):
+        from ..models.attention_processor import CustomDiffusionAttnProcessor
+
+        attn_processors = {}
+        custom_diffusion_grouped_dict = defaultdict(dict)
+        for key, value in state_dict.items():
+            if len(value) == 0:
+                custom_diffusion_grouped_dict[key] = {}
+            else:
+                if "to_out" in key:
+                    attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+                else:
+                    attn_processor_key, sub_key = ".".join(key.split(".")[:-2]), ".".join(key.split(".")[-2:])
+                custom_diffusion_grouped_dict[attn_processor_key][sub_key] = value
+
+        for key, value_dict in custom_diffusion_grouped_dict.items():
+            if len(value_dict) == 0:
+                attn_processors[key] = CustomDiffusionAttnProcessor(
+                    train_kv=False, train_q_out=False, hidden_size=None, cross_attention_dim=None
+                )
+            else:
+                cross_attention_dim = value_dict["to_k_custom_diffusion.weight"].shape[1]
+                hidden_size = value_dict["to_k_custom_diffusion.weight"].shape[0]
+                train_q_out = True if "to_q_custom_diffusion.weight" in value_dict else False
+                attn_processors[key] = CustomDiffusionAttnProcessor(
+                    train_kv=True,
+                    train_q_out=train_q_out,
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                )
+                attn_processors[key].load_state_dict(value_dict)
+
+        return attn_processors
+
+    def _process_lora(self, state_dict, unet_identifier_key, network_alphas, adapter_name, _pipeline):
+        # This method does the following things:
+        # 1. Filters the `state_dict` with keys matching  `unet_identifier_key` when using the non-legacy
+        #    format. For legacy format no filtering is applied.
+        # 2. Converts the `state_dict` to the `peft` compatible format.
+        # 3. Creates a `LoraConfig` and then injects the converted `state_dict` into the UNet per the
+        #    `LoraConfig` specs.
+        # 4. It also reports if the underlying `_pipeline` has any kind of offloading inside of it.
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
+
+        keys = list(state_dict.keys())
+
+        unet_keys = [k for k in keys if k.startswith(unet_identifier_key)]
+        unet_state_dict = {
+            k.replace(f"{unet_identifier_key}.", ""): v for k, v in state_dict.items() if k in unet_keys
+        }
+
+        if network_alphas is not None:
+            alpha_keys = [k for k in network_alphas.keys() if k.startswith(unet_identifier_key)]
+            network_alphas = {
+                k.replace(f"{unet_identifier_key}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+            }
+
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        state_dict_to_be_used = unet_state_dict if len(unet_state_dict) > 0 else state_dict
+
+        if len(state_dict_to_be_used) > 0:
+            if adapter_name in getattr(self, "peft_config", {}):
+                raise ValueError(
+                    f"Adapter name {adapter_name} already in use in the Unet - please select a new adapter name."
+                )
+
+            state_dict = convert_unet_state_dict_to_peft(state_dict_to_be_used)
+
+            if network_alphas is not None:
+                # The alphas state dict have the same structure as Unet, thus we convert it to peft format using
+                # `convert_unet_state_dict_to_peft` method.
+                network_alphas = convert_unet_state_dict_to_peft(network_alphas)
+
+            rank = {}
+            for key, val in state_dict.items():
+                if "lora_B" in key:
+                    rank[key] = val.shape[1]
+
+            lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict, is_unet=True)
+            if "use_dora" in lora_config_kwargs:
+                if lora_config_kwargs["use_dora"]:
+                    if is_peft_version("<", "0.9.0"):
+                        raise ValueError(
+                            "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
+                        )
+                else:
+                    if is_peft_version("<", "0.9.0"):
+                        lora_config_kwargs.pop("use_dora")
+            lora_config = LoraConfig(**lora_config_kwargs)
+
+            # adapter_name
+            if adapter_name is None:
+                adapter_name = get_adapter_name(self)
+
+            # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
+            # otherwise loading LoRA weights will lead to an error
+            is_model_cpu_offload, is_sequential_cpu_offload = self._optionally_disable_offloading(_pipeline)
+
+            inject_adapter_in_model(lora_config, self, adapter_name=adapter_name)
+            incompatible_keys = set_peft_model_state_dict(self, state_dict, adapter_name)
+
+            if incompatible_keys is not None:
+                # check only for unexpected keys
+                unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+                if unexpected_keys:
+                    logger.warning(
+                        f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                        f" {unexpected_keys}. "
+                    )
+
+        return is_model_cpu_offload, is_sequential_cpu_offload
+
+    @classmethod
+    # Copied from diffusers.loaders.lora.LoraLoaderMixin._optionally_disable_offloading
+    def _optionally_disable_offloading(cls, _pipeline):
+        """
+        Optionally removes offloading in case the pipeline has been already sequentially offloaded to CPU.
+
+        Args:
+            _pipeline (`DiffusionPipeline`):
+                The pipeline to disable offloading for.
+
+        Returns:
+            tuple:
+                A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
+        """
         is_model_cpu_offload = False
         is_sequential_cpu_offload = False
 
-        # For PEFT backend the Unet is already offloaded at this stage as it is handled inside `load_lora_weights_into_unet`
-        if not USE_PEFT_BACKEND:
-            if _pipeline is not None:
-                for _, component in _pipeline.components.items():
-                    if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
-                        is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+        if _pipeline is not None and _pipeline.hf_device_map is None:
+            for _, component in _pipeline.components.items():
+                if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
+                    if not is_model_cpu_offload:
+                        is_model_cpu_offload = isinstance(component._hf_hook, CpuOffload)
+                    if not is_sequential_cpu_offload:
                         is_sequential_cpu_offload = (
-                            isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                            isinstance(component._hf_hook, AlignDevicesHook)
                             or hasattr(component._hf_hook, "hooks")
                             and isinstance(component._hf_hook.hooks[0], AlignDevicesHook)
                         )
 
-                        logger.info(
-                            "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
-                        )
-                        remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+                    logger.info(
+                        "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+                    )
+                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
 
-            # only custom diffusion needs to set attn processors
-            if is_custom_diffusion:
-                self.set_attn_processor(attn_processors)
-
-            # set lora layers
-            for target_module, lora_layer in lora_layers_list:
-                target_module.set_lora_layer(lora_layer)
-
-            self.to(dtype=self.dtype, device=self.device)
-
-            # Offload back.
-            if is_model_cpu_offload:
-                _pipeline.enable_model_cpu_offload()
-            elif is_sequential_cpu_offload:
-                _pipeline.enable_sequential_cpu_offload()
-            # Unsafe code />
-
-    def convert_state_dict_legacy_attn_format(self, state_dict, network_alphas):
-        is_new_lora_format = all(
-            key.startswith(self.unet_name) or key.startswith(self.text_encoder_name) for key in state_dict.keys()
-        )
-        if is_new_lora_format:
-            # Strip the `"unet"` prefix.
-            is_text_encoder_present = any(key.startswith(self.text_encoder_name) for key in state_dict.keys())
-            if is_text_encoder_present:
-                warn_message = "The state_dict contains LoRA params corresponding to the text encoder which are not being used here. To use both UNet and text encoder related LoRA params, use [`pipe.load_lora_weights()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights)."
-                logger.warning(warn_message)
-            unet_keys = [k for k in state_dict.keys() if k.startswith(self.unet_name)]
-            state_dict = {k.replace(f"{self.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
-
-        # change processor format to 'pure' LoRACompatibleLinear format
-        if any("processor" in k.split(".") for k in state_dict.keys()):
-
-            def format_to_lora_compatible(key):
-                if "processor" not in key.split("."):
-                    return key
-                return key.replace(".processor", "").replace("to_out_lora", "to_out.0.lora").replace("_lora", ".lora")
-
-            state_dict = {format_to_lora_compatible(k): v for k, v in state_dict.items()}
-
-            if network_alphas is not None:
-                network_alphas = {format_to_lora_compatible(k): v for k, v in network_alphas.items()}
-        return state_dict, network_alphas
+        return (is_model_cpu_offload, is_sequential_cpu_offload)
 
     def save_attn_procs(
         self,
@@ -460,6 +448,23 @@ class UNet2DConditionLoadersMixin:
             logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
             return
 
+        is_custom_diffusion = any(
+            isinstance(
+                x,
+                (CustomDiffusionAttnProcessor, CustomDiffusionAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor),
+            )
+            for (_, x) in self.attn_processors.items()
+        )
+        if is_custom_diffusion:
+            state_dict = self._get_custom_diffusion_state_dict()
+        else:
+            if not USE_PEFT_BACKEND:
+                raise ValueError("PEFT backend is required for saving LoRAs using the `save_attn_procs()` method.")
+
+            from peft.utils import get_peft_model_state_dict
+
+            state_dict = get_peft_model_state_dict(self)
+
         if save_function is None:
             if safe_serialization:
 
@@ -471,36 +476,6 @@ class UNet2DConditionLoadersMixin:
 
         os.makedirs(save_directory, exist_ok=True)
 
-        is_custom_diffusion = any(
-            isinstance(
-                x,
-                (CustomDiffusionAttnProcessor, CustomDiffusionAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor),
-            )
-            for (_, x) in self.attn_processors.items()
-        )
-        if is_custom_diffusion:
-            model_to_save = AttnProcsLayers(
-                {
-                    y: x
-                    for (y, x) in self.attn_processors.items()
-                    if isinstance(
-                        x,
-                        (
-                            CustomDiffusionAttnProcessor,
-                            CustomDiffusionAttnProcessor2_0,
-                            CustomDiffusionXFormersAttnProcessor,
-                        ),
-                    )
-                }
-            )
-            state_dict = model_to_save.state_dict()
-            for name, attn in self.attn_processors.items():
-                if len(attn.state_dict()) == 0:
-                    state_dict[name] = {}
-        else:
-            model_to_save = AttnProcsLayers(self.attn_processors)
-            state_dict = model_to_save.state_dict()
-
         if weight_name is None:
             if safe_serialization:
                 weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE if is_custom_diffusion else LORA_WEIGHT_NAME_SAFE
@@ -512,56 +487,84 @@ class UNet2DConditionLoadersMixin:
         save_function(state_dict, save_path)
         logger.info(f"Model weights saved in {save_path}")
 
+    def _get_custom_diffusion_state_dict(self):
+        from ..models.attention_processor import (
+            CustomDiffusionAttnProcessor,
+            CustomDiffusionAttnProcessor2_0,
+            CustomDiffusionXFormersAttnProcessor,
+        )
+
+        model_to_save = AttnProcsLayers(
+            {
+                y: x
+                for (y, x) in self.attn_processors.items()
+                if isinstance(
+                    x,
+                    (
+                        CustomDiffusionAttnProcessor,
+                        CustomDiffusionAttnProcessor2_0,
+                        CustomDiffusionXFormersAttnProcessor,
+                    ),
+                )
+            }
+        )
+        state_dict = model_to_save.state_dict()
+        for name, attn in self.attn_processors.items():
+            if len(attn.state_dict()) == 0:
+                state_dict[name] = {}
+
+        return state_dict
+
     def fuse_lora(self, lora_scale=1.0, safe_fusing=False, adapter_names=None):
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for `fuse_lora()`.")
+
         self.lora_scale = lora_scale
         self._safe_fusing = safe_fusing
         self.apply(partial(self._fuse_lora_apply, adapter_names=adapter_names))
 
     def _fuse_lora_apply(self, module, adapter_names=None):
-        if not USE_PEFT_BACKEND:
-            if hasattr(module, "_fuse_lora"):
-                module._fuse_lora(self.lora_scale, self._safe_fusing)
+        from peft.tuners.tuners_utils import BaseTunerLayer
 
-            if adapter_names is not None:
+        merge_kwargs = {"safe_merge": self._safe_fusing}
+
+        if isinstance(module, BaseTunerLayer):
+            if self.lora_scale != 1.0:
+                module.scale_layer(self.lora_scale)
+
+            # For BC with prevous PEFT versions, we need to check the signature
+            # of the `merge` method to see if it supports the `adapter_names` argument.
+            supported_merge_kwargs = list(inspect.signature(module.merge).parameters)
+            if "adapter_names" in supported_merge_kwargs:
+                merge_kwargs["adapter_names"] = adapter_names
+            elif "adapter_names" not in supported_merge_kwargs and adapter_names is not None:
                 raise ValueError(
-                    "The `adapter_names` argument is not supported in your environment. Please switch"
-                    " to PEFT backend to use this argument by installing latest PEFT and transformers."
-                    " `pip install -U peft transformers`"
+                    "The `adapter_names` argument is not supported with your PEFT version. Please upgrade"
+                    " to the latest version of PEFT. `pip install -U peft`"
                 )
-        else:
-            from peft.tuners.tuners_utils import BaseTunerLayer
 
-            merge_kwargs = {"safe_merge": self._safe_fusing}
-
-            if isinstance(module, BaseTunerLayer):
-                if self.lora_scale != 1.0:
-                    module.scale_layer(self.lora_scale)
-
-                # For BC with prevous PEFT versions, we need to check the signature
-                # of the `merge` method to see if it supports the `adapter_names` argument.
-                supported_merge_kwargs = list(inspect.signature(module.merge).parameters)
-                if "adapter_names" in supported_merge_kwargs:
-                    merge_kwargs["adapter_names"] = adapter_names
-                elif "adapter_names" not in supported_merge_kwargs and adapter_names is not None:
-                    raise ValueError(
-                        "The `adapter_names` argument is not supported with your PEFT version. Please upgrade"
-                        " to the latest version of PEFT. `pip install -U peft`"
-                    )
-
-                module.merge(**merge_kwargs)
+            module.merge(**merge_kwargs)
 
     def unfuse_lora(self):
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for `unfuse_lora()`.")
         self.apply(self._unfuse_lora_apply)
 
     def _unfuse_lora_apply(self, module):
-        if not USE_PEFT_BACKEND:
-            if hasattr(module, "_unfuse_lora"):
-                module._unfuse_lora()
-        else:
-            from peft.tuners.tuners_utils import BaseTunerLayer
+        from peft.tuners.tuners_utils import BaseTunerLayer
 
-            if isinstance(module, BaseTunerLayer):
-                module.unmerge()
+        if isinstance(module, BaseTunerLayer):
+            module.unmerge()
+
+    def unload_lora(self):
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for `unload_lora()`.")
+
+        from ..utils import recurse_remove_peft_layers
+
+        recurse_remove_peft_layers(self)
+        if hasattr(self, "peft_config"):
+            del self.peft_config
 
     def set_adapters(
         self,
@@ -847,7 +850,12 @@ class UNet2DConditionLoadersMixin:
             embed_dims = state_dict["proj_in.weight"].shape[1]
             output_dims = state_dict["proj_out.weight"].shape[0]
             hidden_dims = state_dict["latents"].shape[2]
-            heads = state_dict["layers.0.0.to_q.weight"].shape[0] // 64
+            attn_key_present = any("attn" in k for k in state_dict)
+            heads = (
+                state_dict["layers.0.attn.to_q.weight"].shape[0] // 64
+                if attn_key_present
+                else state_dict["layers.0.0.to_q.weight"].shape[0] // 64
+            )
 
             with init_context():
                 image_projection = IPAdapterPlusImageProjection(
@@ -860,26 +868,53 @@ class UNet2DConditionLoadersMixin:
 
             for key, value in state_dict.items():
                 diffusers_name = key.replace("0.to", "2.to")
-                diffusers_name = diffusers_name.replace("1.0.weight", "3.0.weight")
-                diffusers_name = diffusers_name.replace("1.0.bias", "3.0.bias")
-                diffusers_name = diffusers_name.replace("1.1.weight", "3.1.net.0.proj.weight")
-                diffusers_name = diffusers_name.replace("1.3.weight", "3.1.net.2.weight")
 
-                if "norm1" in diffusers_name:
-                    updated_state_dict[diffusers_name.replace("0.norm1", "0")] = value
-                elif "norm2" in diffusers_name:
-                    updated_state_dict[diffusers_name.replace("0.norm2", "1")] = value
-                elif "to_kv" in diffusers_name:
+                diffusers_name = diffusers_name.replace("0.0.norm1", "0.ln0")
+                diffusers_name = diffusers_name.replace("0.0.norm2", "0.ln1")
+                diffusers_name = diffusers_name.replace("1.0.norm1", "1.ln0")
+                diffusers_name = diffusers_name.replace("1.0.norm2", "1.ln1")
+                diffusers_name = diffusers_name.replace("2.0.norm1", "2.ln0")
+                diffusers_name = diffusers_name.replace("2.0.norm2", "2.ln1")
+                diffusers_name = diffusers_name.replace("3.0.norm1", "3.ln0")
+                diffusers_name = diffusers_name.replace("3.0.norm2", "3.ln1")
+
+                if "to_kv" in diffusers_name:
+                    parts = diffusers_name.split(".")
+                    parts[2] = "attn"
+                    diffusers_name = ".".join(parts)
                     v_chunk = value.chunk(2, dim=0)
                     updated_state_dict[diffusers_name.replace("to_kv", "to_k")] = v_chunk[0]
                     updated_state_dict[diffusers_name.replace("to_kv", "to_v")] = v_chunk[1]
+                elif "to_q" in diffusers_name:
+                    parts = diffusers_name.split(".")
+                    parts[2] = "attn"
+                    diffusers_name = ".".join(parts)
+                    updated_state_dict[diffusers_name] = value
                 elif "to_out" in diffusers_name:
+                    parts = diffusers_name.split(".")
+                    parts[2] = "attn"
+                    diffusers_name = ".".join(parts)
                     updated_state_dict[diffusers_name.replace("to_out", "to_out.0")] = value
                 else:
+                    diffusers_name = diffusers_name.replace("0.1.0", "0.ff.0")
+                    diffusers_name = diffusers_name.replace("0.1.1", "0.ff.1.net.0.proj")
+                    diffusers_name = diffusers_name.replace("0.1.3", "0.ff.1.net.2")
+
+                    diffusers_name = diffusers_name.replace("1.1.0", "1.ff.0")
+                    diffusers_name = diffusers_name.replace("1.1.1", "1.ff.1.net.0.proj")
+                    diffusers_name = diffusers_name.replace("1.1.3", "1.ff.1.net.2")
+
+                    diffusers_name = diffusers_name.replace("2.1.0", "2.ff.0")
+                    diffusers_name = diffusers_name.replace("2.1.1", "2.ff.1.net.0.proj")
+                    diffusers_name = diffusers_name.replace("2.1.3", "2.ff.1.net.2")
+
+                    diffusers_name = diffusers_name.replace("3.1.0", "3.ff.0")
+                    diffusers_name = diffusers_name.replace("3.1.1", "3.ff.1.net.0.proj")
+                    diffusers_name = diffusers_name.replace("3.1.3", "3.ff.1.net.2")
                     updated_state_dict[diffusers_name] = value
 
         if not low_cpu_mem_usage:
-            image_projection.load_state_dict(updated_state_dict)
+            image_projection.load_state_dict(updated_state_dict, strict=True)
         else:
             load_model_dict_into_meta(image_projection, updated_state_dict, device=self.device, dtype=self.dtype)
 

src/diffusers/models/__init__.py

@@ -31,11 +31,15 @@ if is_torch_available():
     _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
     _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
     _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
+    _import_structure["autoencoders.vq_model"] = ["VQModel"]
     _import_structure["controlnet"] = ["ControlNetModel"]
     _import_structure["controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"]
     _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
     _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
+    _import_structure["transformers.dit_transformer_2d"] = ["DiTTransformer2DModel"]
+    _import_structure["transformers.hunyuan_transformer_2d"] = ["HunyuanDiT2DModel"]
+    _import_structure["transformers.pixart_transformer_2d"] = ["PixArtTransformer2DModel"]
     _import_structure["transformers.prior_transformer"] = ["PriorTransformer"]
     _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
@@ -50,7 +54,6 @@ if is_torch_available():
     _import_structure["unets.unet_spatio_temporal_condition"] = ["UNetSpatioTemporalConditionModel"]
     _import_structure["unets.unet_stable_cascade"] = ["StableCascadeUNet"]
     _import_structure["unets.uvit_2d"] = ["UVit2DModel"]
-    _import_structure["vq_model"] = ["VQModel"]
 
 if is_flax_available():
     _import_structure["controlnet_flax"] = ["FlaxControlNetModel"]
@@ -67,13 +70,17 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKLTemporalDecoder,
             AutoencoderTiny,
             ConsistencyDecoderVAE,
+            VQModel,
         )
         from .controlnet import ControlNetModel
         from .controlnet_xs import ControlNetXSAdapter, UNetControlNetXSModel
         from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin
         from .transformers import (
+            DiTTransformer2DModel,
             DualTransformer2DModel,
+            HunyuanDiT2DModel,
+            PixArtTransformer2DModel,
             PriorTransformer,
             T5FilmDecoder,
             Transformer2DModel,
@@ -92,7 +99,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             UNetSpatioTemporalConditionModel,
             UVit2DModel,
         )
-        from .vq_model import VQModel
 
     if is_flax_available():
         from .controlnet_flax import FlaxControlNetModel

src/diffusers/models/activations.py

@@ -50,6 +50,18 @@ def get_activation(act_fn: str) -> nn.Module:
         raise ValueError(f"Unsupported activation function: {act_fn}")
 
 
+class FP32SiLU(nn.Module):
+    r"""
+    SiLU activation function with input upcasted to torch.float32.
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        return F.silu(inputs.float(), inplace=False).to(inputs.dtype)
+
+
 class GELU(nn.Module):
     r"""
     GELU activation function with tanh approximation support with `approximate="tanh"`.

src/diffusers/models/attention_processor.py

@@ -103,6 +103,7 @@ class Attention(nn.Module):
         upcast_softmax: bool = False,
         cross_attention_norm: Optional[str] = None,
         cross_attention_norm_num_groups: int = 32,
+        qk_norm: Optional[str] = None,
         added_kv_proj_dim: Optional[int] = None,
         norm_num_groups: Optional[int] = None,
         spatial_norm_dim: Optional[int] = None,
@@ -161,6 +162,15 @@ class Attention(nn.Module):
         else:
             self.spatial_norm = None
 
+        if qk_norm is None:
+            self.norm_q = None
+            self.norm_k = None
+        elif qk_norm == "layer_norm":
+            self.norm_q = nn.LayerNorm(dim_head, eps=eps)
+            self.norm_k = nn.LayerNorm(dim_head, eps=eps)
+        else:
+            raise ValueError(f"unknown qk_norm: {qk_norm}. Should be None or 'layer_norm'")
+
         if cross_attention_norm is None:
             self.norm_cross = None
         elif cross_attention_norm == "layer_norm":
@@ -1426,6 +1436,104 @@ class AttnProcessor2_0:
         return hidden_states
 
 
+class HunyuanAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
+    used in the HunyuanDiT model. It applies a s normalization layer and rotary embedding on query and key vector.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        from .embeddings import apply_rotary_emb
+
+        residual = hidden_states
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb)
+            if not attn.is_cross_attention:
+                key = apply_rotary_emb(key, image_rotary_emb)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
 class FusedAttnProcessor2_0:
     r"""
     Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). It uses

src/diffusers/models/autoencoders/__init__.py

@@ -3,3 +3,4 @@ from .autoencoder_kl import AutoencoderKL
 from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
 from .autoencoder_tiny import AutoencoderTiny
 from .consistency_decoder_vae import ConsistencyDecoderVAE
+from .vq_model import VQModel

src/diffusers/models/autoencoders/autoencoder_asym_kl.py

@@ -176,7 +176,7 @@ class AsymmetricAutoencoderKL(ModelMixin, ConfigMixin):
             z = posterior.sample(generator=generator)
         else:
             z = posterior.mode()
-        dec = self.decode(z, sample, mask).sample
+        dec = self.decode(z, generator, sample, mask).sample
 
         if not return_dict:
             return (dec,)

src/diffusers/models/autoencoders/autoencoder_kl.py

@@ -245,11 +245,13 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         Args:
             x (`torch.Tensor`): Input batch of images.
             return_dict (`bool`, *optional*, defaults to `True`):
-                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+                Whether to return a [`~models.autoencoders.autoencoder_kl.AutoencoderKLOutput`] instead of a plain
+                tuple.
 
         Returns:
                 The latent representations of the encoded images. If `return_dict` is True, a
-                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+                [`~models.autoencoders.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is
+                returned.
         """
         if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
             return self.tiled_encode(x, return_dict=return_dict)
@@ -331,12 +333,13 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         Args:
             x (`torch.Tensor`): Input batch of images.
             return_dict (`bool`, *optional*, defaults to `True`):
-                Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+                Whether or not to return a [`~models.autoencoders.autoencoder_kl.AutoencoderKLOutput`] instead of a
+                plain tuple.
 
         Returns:
-            [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
-                If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain
-                `tuple` is returned.
+            [`~models.autoencoders.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
+                If return_dict is True, a [`~models.autoencoders.autoencoder_kl.AutoencoderKLOutput`] is returned,
+                otherwise a plain `tuple` is returned.
         """
         overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
         blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)

src/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py

@@ -323,11 +323,13 @@ class AutoencoderKLTemporalDecoder(ModelMixin, ConfigMixin):
         Args:
             x (`torch.Tensor`): Input batch of images.
             return_dict (`bool`, *optional*, defaults to `True`):
-                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+                Whether to return a [`~models.autoencoders.autoencoder_kl.AutoencoderKLOutput`] instead of a plain
+                tuple.
 
         Returns:
                 The latent representations of the encoded images. If `return_dict` is True, a
-                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+                [`~models.autoencoders.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is
+                returned.
         """
         h = self.encoder(x)
         moments = self.quant_conv(h)

src/diffusers/models/autoencoders/consistency_decoder_vae.py

@@ -284,13 +284,13 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         Args:
             x (`torch.Tensor`): Input batch of images.
             return_dict (`bool`, *optional*, defaults to `True`):
-                Whether to return a [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] instead of a plain
-                tuple.
+                Whether to return a [`~models.autoencoders.consistency_decoder_vae.ConsistencyDecoderVAEOutput`]
+                instead of a plain tuple.
 
         Returns:
                 The latent representations of the encoded images. If `return_dict` is True, a
-                [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] is returned, otherwise a plain `tuple`
-                is returned.
+                [`~models.autoencoders.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] is returned, otherwise a
+                plain `tuple` is returned.
         """
         if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
             return self.tiled_encode(x, return_dict=return_dict)
@@ -382,13 +382,13 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         Args:
             x (`torch.Tensor`): Input batch of images.
             return_dict (`bool`, *optional*, defaults to `True`):
-                Whether or not to return a [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] instead of a
-                plain tuple.
+                Whether or not to return a [`~models.autoencoders.consistency_decoder_vae.ConsistencyDecoderVAEOutput`]
+                instead of a plain tuple.
 
         Returns:
-            [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] or `tuple`:
-                If return_dict is True, a [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] is returned,
-                otherwise a plain `tuple` is returned.
+            [`~models.autoencoders.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] or `tuple`:
+                If return_dict is True, a [`~models.autoencoders.consistency_decoder_vae.ConsistencyDecoderVAEOutput`]
+                is returned, otherwise a plain `tuple` is returned.
         """
         overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
         blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)

src/diffusers/models/autoencoders/vq_model.py

@@ -0,0 +1,182 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput
+from ...utils.accelerate_utils import apply_forward_hook
+from ..autoencoders.vae import Decoder, DecoderOutput, Encoder, VectorQuantizer
+from ..modeling_utils import ModelMixin
+
+
+@dataclass
+class VQEncoderOutput(BaseOutput):
+    """
+    Output of VQModel encoding method.
+
+    Args:
+        latents (`torch.Tensor` of shape `(batch_size, num_channels, height, width)`):
+            The encoded output sample from the last layer of the model.
+    """
+
+    latents: torch.Tensor
+
+
+class VQModel(ModelMixin, ConfigMixin):
+    r"""
+    A VQ-VAE model for decoding latent representations.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            Tuple of downsample block types.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            Tuple of upsample block types.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+            Tuple of block output channels.
+        layers_per_block (`int`, *optional*, defaults to `1`): Number of layers per block.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        latent_channels (`int`, *optional*, defaults to `3`): Number of channels in the latent space.
+        sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+        num_vq_embeddings (`int`, *optional*, defaults to `256`): Number of codebook vectors in the VQ-VAE.
+        norm_num_groups (`int`, *optional*, defaults to `32`): Number of groups for normalization layers.
+        vq_embed_dim (`int`, *optional*): Hidden dim of codebook vectors in the VQ-VAE.
+        scaling_factor (`float`, *optional*, defaults to `0.18215`):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+        norm_type (`str`, *optional*, defaults to `"group"`):
+            Type of normalization layer to use. Can be one of `"group"` or `"spatial"`.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",),
+        up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
+        block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_block: int = 1,
+        act_fn: str = "silu",
+        latent_channels: int = 3,
+        sample_size: int = 32,
+        num_vq_embeddings: int = 256,
+        norm_num_groups: int = 32,
+        vq_embed_dim: Optional[int] = None,
+        scaling_factor: float = 0.18215,
+        norm_type: str = "group",  # group, spatial
+        mid_block_add_attention=True,
+        lookup_from_codebook=False,
+        force_upcast=False,
+    ):
+        super().__init__()
+
+        # pass init params to Encoder
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_num_groups=norm_num_groups,
+            double_z=False,
+            mid_block_add_attention=mid_block_add_attention,
+        )
+
+        vq_embed_dim = vq_embed_dim if vq_embed_dim is not None else latent_channels
+
+        self.quant_conv = nn.Conv2d(latent_channels, vq_embed_dim, 1)
+        self.quantize = VectorQuantizer(num_vq_embeddings, vq_embed_dim, beta=0.25, remap=None, sane_index_shape=False)
+        self.post_quant_conv = nn.Conv2d(vq_embed_dim, latent_channels, 1)
+
+        # pass init params to Decoder
+        self.decoder = Decoder(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            up_block_types=up_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_num_groups=norm_num_groups,
+            norm_type=norm_type,
+            mid_block_add_attention=mid_block_add_attention,
+        )
+
+    @apply_forward_hook
+    def encode(self, x: torch.Tensor, return_dict: bool = True) -> VQEncoderOutput:
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+
+        if not return_dict:
+            return (h,)
+
+        return VQEncoderOutput(latents=h)
+
+    @apply_forward_hook
+    def decode(
+        self, h: torch.Tensor, force_not_quantize: bool = False, return_dict: bool = True, shape=None
+    ) -> Union[DecoderOutput, torch.Tensor]:
+        # also go through quantization layer
+        if not force_not_quantize:
+            quant, commit_loss, _ = self.quantize(h)
+        elif self.config.lookup_from_codebook:
+            quant = self.quantize.get_codebook_entry(h, shape)
+            commit_loss = torch.zeros((h.shape[0])).to(h.device, dtype=h.dtype)
+        else:
+            quant = h
+            commit_loss = torch.zeros((h.shape[0])).to(h.device, dtype=h.dtype)
+        quant2 = self.post_quant_conv(quant)
+        dec = self.decoder(quant2, quant if self.config.norm_type == "spatial" else None)
+
+        if not return_dict:
+            return dec, commit_loss
+
+        return DecoderOutput(sample=dec, commit_loss=commit_loss)
+
+    def forward(
+        self, sample: torch.Tensor, return_dict: bool = True
+    ) -> Union[DecoderOutput, Tuple[torch.Tensor, ...]]:
+        r"""
+        The [`VQModel`] forward method.
+
+        Args:
+            sample (`torch.Tensor`): Input sample.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.vq_model.VQEncoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vq_model.VQEncoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vq_model.VQEncoderOutput`] is returned, otherwise a plain `tuple`
+                is returned.
+        """
+
+        h = self.encode(sample).latents
+        dec = self.decode(h)
+
+        if not return_dict:
+            return dec.sample, dec.commit_loss
+        return dec

src/diffusers/models/controlnet_xs.py

@@ -851,8 +851,8 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
         if hasattr(module, "gradient_checkpointing"):
             module.gradient_checkpointing = value
 
-    # copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel
     @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
     def attn_processors(self) -> Dict[str, AttentionProcessor]:
         r"""
         Returns:
@@ -911,7 +911,7 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
         for name, module in self.named_children():
             fn_recursive_attn_processor(name, module, processor)
 
-    # copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """
         Disables custom attention processors and sets the default attention implementation.
@@ -927,7 +927,7 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
 
         self.set_attn_processor(processor)
 
-    # copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.enable_freeu
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.enable_freeu
     def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
         r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
 
@@ -952,7 +952,7 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
             setattr(upsample_block, "b1", b1)
             setattr(upsample_block, "b2", b2)
 
-    # copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.disable_freeu
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.disable_freeu
     def disable_freeu(self):
         """Disables the FreeU mechanism."""
         freeu_keys = {"s1", "s2", "b1", "b2"}
@@ -961,7 +961,7 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
                 if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
                     setattr(upsample_block, k, None)
 
-    # copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
     def fuse_qkv_projections(self):
         """
         Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
@@ -985,7 +985,7 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
             if isinstance(module, Attention):
                 module.fuse_projections(fuse=True)
 
-    # copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
     def unfuse_qkv_projections(self):
         """Disables the fused QKV projection if enabled.
 

src/diffusers/models/dual_transformer_2d.py

@@ -1,20 +0,0 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-from ..utils import deprecate
-from .transformers.dual_transformer_2d import DualTransformer2DModel
-
-
-class DualTransformer2DModel(DualTransformer2DModel):
-    deprecation_message = "Importing `DualTransformer2DModel` from `diffusers.models.dual_transformer_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.dual_transformer_2d import DualTransformer2DModel`, instead."
-    deprecate("DualTransformer2DModel", "0.29", deprecation_message)

src/diffusers/models/embeddings.py

@@ -16,10 +16,11 @@ from typing import List, Optional, Tuple, Union
 
 import numpy as np
 import torch
+import torch.nn.functional as F
 from torch import nn
 
 from ..utils import deprecate
-from .activations import get_activation
+from .activations import FP32SiLU, get_activation
 from .attention_processor import Attention
 
 
@@ -135,6 +136,7 @@ class PatchEmbed(nn.Module):
         flatten=True,
         bias=True,
         interpolation_scale=1,
+        pos_embed_type="sincos",
     ):
         super().__init__()
 
@@ -156,10 +158,18 @@ class PatchEmbed(nn.Module):
         self.height, self.width = height // patch_size, width // patch_size
         self.base_size = height // patch_size
         self.interpolation_scale = interpolation_scale
-        pos_embed = get_2d_sincos_pos_embed(
-            embed_dim, int(num_patches**0.5), base_size=self.base_size, interpolation_scale=self.interpolation_scale
-        )
-        self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
+        if pos_embed_type is None:
+            self.pos_embed = None
+        elif pos_embed_type == "sincos":
+            pos_embed = get_2d_sincos_pos_embed(
+                embed_dim,
+                int(num_patches**0.5),
+                base_size=self.base_size,
+                interpolation_scale=self.interpolation_scale,
+            )
+            self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
+        else:
+            raise ValueError(f"Unsupported pos_embed_type: {pos_embed_type}")
 
     def forward(self, latent):
         height, width = latent.shape[-2] // self.patch_size, latent.shape[-1] // self.patch_size
@@ -169,6 +179,8 @@ class PatchEmbed(nn.Module):
             latent = latent.flatten(2).transpose(1, 2)  # BCHW -> BNC
         if self.layer_norm:
             latent = self.norm(latent)
+        if self.pos_embed is None:
+            return latent.to(latent.dtype)
 
         # Interpolate positional embeddings if needed.
         # (For PixArt-Alpha: https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L162C151-L162C160)
@@ -187,6 +199,113 @@ class PatchEmbed(nn.Module):
         return (latent + pos_embed).to(latent.dtype)
 
 
+def get_2d_rotary_pos_embed(embed_dim, crops_coords, grid_size, use_real=True):
+    """
+    RoPE for image tokens with 2d structure.
+
+    Args:
+    embed_dim: (`int`):
+        The embedding dimension size
+    crops_coords (`Tuple[int]`)
+        The top-left and bottom-right coordinates of the crop.
+    grid_size (`Tuple[int]`):
+        The grid size of the positional embedding.
+    use_real (`bool`):
+        If True, return real part and imaginary part separately. Otherwise, return complex numbers.
+
+    Returns:
+        `torch.Tensor`: positional embdding with shape `( grid_size * grid_size, embed_dim/2)`.
+    """
+    start, stop = crops_coords
+    grid_h = np.linspace(start[0], stop[0], grid_size[0], endpoint=False, dtype=np.float32)
+    grid_w = np.linspace(start[1], stop[1], grid_size[1], endpoint=False, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)  # [2, W, H]
+
+    grid = grid.reshape([2, 1, *grid.shape[1:]])
+    pos_embed = get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=use_real)
+    return pos_embed
+
+
+def get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=False):
+    assert embed_dim % 4 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_rotary_pos_embed(embed_dim // 2, grid[0].reshape(-1), use_real=use_real)  # (H*W, D/4)
+    emb_w = get_1d_rotary_pos_embed(embed_dim // 2, grid[1].reshape(-1), use_real=use_real)  # (H*W, D/4)
+
+    if use_real:
+        cos = torch.cat([emb_h[0], emb_w[0]], dim=1)  # (H*W, D/2)
+        sin = torch.cat([emb_h[1], emb_w[1]], dim=1)  # (H*W, D/2)
+        return cos, sin
+    else:
+        emb = torch.cat([emb_h, emb_w], dim=1)  # (H*W, D/2)
+        return emb
+
+
+def get_1d_rotary_pos_embed(dim: int, pos: Union[np.ndarray, int], theta: float = 10000.0, use_real=False):
+    """
+    Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
+
+    This function calculates a frequency tensor with complex exponentials using the given dimension 'dim' and the end
+    index 'end'. The 'theta' parameter scales the frequencies. The returned tensor contains complex values in complex64
+    data type.
+
+    Args:
+        dim (`int`): Dimension of the frequency tensor.
+        pos (`np.ndarray` or `int`): Position indices for the frequency tensor. [S] or scalar
+        theta (`float`, *optional*, defaults to 10000.0):
+            Scaling factor for frequency computation. Defaults to 10000.0.
+        use_real (`bool`, *optional*):
+            If True, return real part and imaginary part separately. Otherwise, return complex numbers.
+
+    Returns:
+        `torch.Tensor`: Precomputed frequency tensor with complex exponentials. [S, D/2]
+    """
+    if isinstance(pos, int):
+        pos = np.arange(pos)
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))  # [D/2]
+    t = torch.from_numpy(pos).to(freqs.device)  # type: ignore  # [S]
+    freqs = torch.outer(t, freqs).float()  # type: ignore   # [S, D/2]
+    if use_real:
+        freqs_cos = freqs.cos().repeat_interleave(2, dim=1)  # [S, D]
+        freqs_sin = freqs.sin().repeat_interleave(2, dim=1)  # [S, D]
+        return freqs_cos, freqs_sin
+    else:
+        freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64     # [S, D/2]
+        return freqs_cis
+
+
+def apply_rotary_emb(
+    x: torch.Tensor,
+    freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
+    to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
+    reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
+    tensors contain rotary embeddings and are returned as real tensors.
+
+    Args:
+        x (`torch.Tensor`):
+            Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
+        freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
+    """
+    cos, sin = freqs_cis  # [S, D]
+    cos = cos[None, None]
+    sin = sin[None, None]
+    cos, sin = cos.to(x.device), sin.to(x.device)
+
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+    x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+    out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+    return out
+
+
 class TimestepEmbedding(nn.Module):
     def __init__(
         self,
@@ -507,6 +626,88 @@ class CombinedTimestepLabelEmbeddings(nn.Module):
         return conditioning
 
 
+class HunyuanDiTAttentionPool(nn.Module):
+    # Copied from https://github.com/Tencent/HunyuanDiT/blob/cb709308d92e6c7e8d59d0dff41b74d35088db6a/hydit/modules/poolers.py#L6
+
+    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim + 1, embed_dim) / embed_dim**0.5)
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.permute(1, 0, 2)  # NLC -> LNC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (L+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (L+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x[:1],
+            key=x,
+            value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False,
+        )
+        return x.squeeze(0)
+
+
+class HunyuanCombinedTimestepTextSizeStyleEmbedding(nn.Module):
+    def __init__(self, embedding_dim, pooled_projection_dim=1024, seq_len=256, cross_attention_dim=2048):
+        super().__init__()
+
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.pooler = HunyuanDiTAttentionPool(
+            seq_len, cross_attention_dim, num_heads=8, output_dim=pooled_projection_dim
+        )
+        # Here we use a default learned embedder layer for future extension.
+        self.style_embedder = nn.Embedding(1, embedding_dim)
+        extra_in_dim = 256 * 6 + embedding_dim + pooled_projection_dim
+        self.extra_embedder = PixArtAlphaTextProjection(
+            in_features=extra_in_dim,
+            hidden_size=embedding_dim * 4,
+            out_features=embedding_dim,
+            act_fn="silu_fp32",
+        )
+
+    def forward(self, timestep, encoder_hidden_states, image_meta_size, style, hidden_dtype=None):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, 256)
+
+        # extra condition1: text
+        pooled_projections = self.pooler(encoder_hidden_states)  # (N, 1024)
+
+        # extra condition2: image meta size embdding
+        image_meta_size = get_timestep_embedding(image_meta_size.view(-1), 256, True, 0)
+        image_meta_size = image_meta_size.to(dtype=hidden_dtype)
+        image_meta_size = image_meta_size.view(-1, 6 * 256)  # (N, 1536)
+
+        # extra condition3: style embedding
+        style_embedding = self.style_embedder(style)  # (N, embedding_dim)
+
+        # Concatenate all extra vectors
+        extra_cond = torch.cat([pooled_projections, image_meta_size, style_embedding], dim=1)
+        conditioning = timesteps_emb + self.extra_embedder(extra_cond)  # [B, D]
+
+        return conditioning
+
+
 class TextTimeEmbedding(nn.Module):
     def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64):
         super().__init__()
@@ -793,11 +994,18 @@ class PixArtAlphaTextProjection(nn.Module):
     Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
     """
 
-    def __init__(self, in_features, hidden_size, num_tokens=120):
+    def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh"):
         super().__init__()
+        if out_features is None:
+            out_features = hidden_size
         self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True)
-        self.act_1 = nn.GELU(approximate="tanh")
-        self.linear_2 = nn.Linear(in_features=hidden_size, out_features=hidden_size, bias=True)
+        if act_fn == "gelu_tanh":
+            self.act_1 = nn.GELU(approximate="tanh")
+        elif act_fn == "silu_fp32":
+            self.act_1 = FP32SiLU()
+        else:
+            raise ValueError(f"Unknown activation function: {act_fn}")
+        self.linear_2 = nn.Linear(in_features=hidden_size, out_features=out_features, bias=True)
 
     def forward(self, caption):
         hidden_states = self.linear_1(caption)
@@ -806,89 +1014,6 @@ class PixArtAlphaTextProjection(nn.Module):
         return hidden_states
 
 
-class IPAdapterPlusImageProjection(nn.Module):
-    """Resampler of IP-Adapter Plus.
-
-    Args:
-        embed_dims (int): The feature dimension. Defaults to 768. output_dims (int): The number of output channels,
-        that is the same
-            number of the channels in the `unet.config.cross_attention_dim`. Defaults to 1024.
-        hidden_dims (int):
-            The number of hidden channels. Defaults to 1280. depth (int): The number of blocks. Defaults
-        to 8. dim_head (int): The number of head channels. Defaults to 64. heads (int): Parallel attention heads.
-        Defaults to 16. num_queries (int):
-            The number of queries. Defaults to 8. ffn_ratio (float): The expansion ratio
-        of feedforward network hidden
-            layer channels. Defaults to 4.
-    """
-
-    def __init__(
-        self,
-        embed_dims: int = 768,
-        output_dims: int = 1024,
-        hidden_dims: int = 1280,
-        depth: int = 4,
-        dim_head: int = 64,
-        heads: int = 16,
-        num_queries: int = 8,
-        ffn_ratio: float = 4,
-    ) -> None:
-        super().__init__()
-        from .attention import FeedForward  # Lazy import to avoid circular import
-
-        self.latents = nn.Parameter(torch.randn(1, num_queries, hidden_dims) / hidden_dims**0.5)
-
-        self.proj_in = nn.Linear(embed_dims, hidden_dims)
-
-        self.proj_out = nn.Linear(hidden_dims, output_dims)
-        self.norm_out = nn.LayerNorm(output_dims)
-
-        self.layers = nn.ModuleList([])
-        for _ in range(depth):
-            self.layers.append(
-                nn.ModuleList(
-                    [
-                        nn.LayerNorm(hidden_dims),
-                        nn.LayerNorm(hidden_dims),
-                        Attention(
-                            query_dim=hidden_dims,
-                            dim_head=dim_head,
-                            heads=heads,
-                            out_bias=False,
-                        ),
-                        nn.Sequential(
-                            nn.LayerNorm(hidden_dims),
-                            FeedForward(hidden_dims, hidden_dims, activation_fn="gelu", mult=ffn_ratio, bias=False),
-                        ),
-                    ]
-                )
-            )
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        """Forward pass.
-
-        Args:
-            x (torch.Tensor): Input Tensor.
-        Returns:
-            torch.Tensor: Output Tensor.
-        """
-        latents = self.latents.repeat(x.size(0), 1, 1)
-
-        x = self.proj_in(x)
-
-        for ln0, ln1, attn, ff in self.layers:
-            residual = latents
-
-            encoder_hidden_states = ln0(x)
-            latents = ln1(latents)
-            encoder_hidden_states = torch.cat([encoder_hidden_states, latents], dim=-2)
-            latents = attn(latents, encoder_hidden_states) + residual
-            latents = ff(latents) + latents
-
-        latents = self.proj_out(latents)
-        return self.norm_out(latents)
-
-
 class IPAdapterPlusImageProjectionBlock(nn.Module):
     def __init__(
         self,
@@ -922,6 +1047,65 @@ class IPAdapterPlusImageProjectionBlock(nn.Module):
         return latents
 
 
+class IPAdapterPlusImageProjection(nn.Module):
+    """Resampler of IP-Adapter Plus.
+
+    Args:
+        embed_dims (int): The feature dimension. Defaults to 768. output_dims (int): The number of output channels,
+        that is the same
+            number of the channels in the `unet.config.cross_attention_dim`. Defaults to 1024.
+        hidden_dims (int):
+            The number of hidden channels. Defaults to 1280. depth (int): The number of blocks. Defaults
+        to 8. dim_head (int): The number of head channels. Defaults to 64. heads (int): Parallel attention heads.
+        Defaults to 16. num_queries (int):
+            The number of queries. Defaults to 8. ffn_ratio (float): The expansion ratio
+        of feedforward network hidden
+            layer channels. Defaults to 4.
+    """
+
+    def __init__(
+        self,
+        embed_dims: int = 768,
+        output_dims: int = 1024,
+        hidden_dims: int = 1280,
+        depth: int = 4,
+        dim_head: int = 64,
+        heads: int = 16,
+        num_queries: int = 8,
+        ffn_ratio: float = 4,
+    ) -> None:
+        super().__init__()
+        self.latents = nn.Parameter(torch.randn(1, num_queries, hidden_dims) / hidden_dims**0.5)
+
+        self.proj_in = nn.Linear(embed_dims, hidden_dims)
+
+        self.proj_out = nn.Linear(hidden_dims, output_dims)
+        self.norm_out = nn.LayerNorm(output_dims)
+
+        self.layers = nn.ModuleList(
+            [IPAdapterPlusImageProjectionBlock(hidden_dims, dim_head, heads, ffn_ratio) for _ in range(depth)]
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Forward pass.
+
+        Args:
+            x (torch.Tensor): Input Tensor.
+        Returns:
+            torch.Tensor: Output Tensor.
+        """
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        for block in self.layers:
+            residual = latents
+            latents = block(x, latents, residual)
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
 class IPAdapterFaceIDPlusImageProjection(nn.Module):
     """FacePerceiverResampler of IP-Adapter Plus.