update

* EulerAncestral add `rescale_betas_zero_snr`

Uses same infinite sigma fix from EulerDiscrete. Interestingly the
ancestral version had the opposite problem: too much contrast instead of
too little.

* UT for EulerAncestral `rescale_betas_zero_snr`

* EulerAncestral upcast samples during step()

It helps this scheduler too, particularly when the model is using bf16.

While the noise dtype is still the model's it's automatically upcasted
for the add so all it affects is determinism.

---------

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

.github/workflows/benchmark.yml

@@ -0,0 +1,52 @@
+name: Benchmarking tests
+
+on:
+  schedule:
+    - cron: "30 1 1,15 * *" # every 2 weeks on the 1st and the 15th of every month at 1:30 AM
+
+env:
+  DIFFUSERS_IS_CI: yes
+  HF_HOME: /mnt/cache
+  OMP_NUM_THREADS: 8
+  MKL_NUM_THREADS: 8
+
+jobs:
+  torch_pipelines_cuda_benchmark_tests:
+    name: Torch Core Pipelines CUDA Benchmarking Tests
+    strategy:
+      fail-fast: false
+      max-parallel: 1
+    runs-on: [single-gpu, nvidia-gpu, a10, ci]
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --gpus 0
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+      - name: NVIDIA-SMI
+        run: |
+          nvidia-smi
+      - name: Install dependencies
+        run: |
+          apt-get update && apt-get install libsndfile1-dev libgl1 -y
+          python -m pip install -e .[quality,test]
+          python -m pip install pandas
+      - name: Environment
+        run: |
+          python utils/print_env.py
+      - name: Diffusers Benchmarking
+        env:
+            HUGGING_FACE_HUB_TOKEN: ${{ secrets.DIFFUSERS_BOT_TOKEN }}
+            BASE_PATH: benchmark_outputs
+        run: |
+          export TOTAL_GPU_MEMORY=$(python -c "import torch; print(torch.cuda.get_device_properties(0).total_memory / (1024**3))")
+          cd benchmarks && mkdir ${BASE_PATH} && python run_all.py && python push_results.py
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: benchmark_test_reports
+          path: benchmarks/benchmark_outputs

.github/workflows/pr_test_fetcher.yml

@@ -1,12 +1,6 @@
 name: Fast tests for PRs - Test Fetcher
 
-on:
-  pull_request:
-    branches:
-      - main
-  push:
-    branches:
-      - ci-*
+on: workflow_dispatch
 
 env:
   DIFFUSERS_IS_CI: yes

.github/workflows/pr_tests.yml

@@ -113,6 +113,7 @@ jobs:
     - name: Run example PyTorch CPU tests
       if: ${{ matrix.config.framework == 'pytorch_examples' }}
       run: |
+        python -m pip install peft
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           --make-reports=tests_${{ matrix.config.report }} \
           examples

.github/workflows/push_tests_fast.yml

@@ -98,6 +98,7 @@ jobs:
     - name: Run example PyTorch CPU tests
       if: ${{ matrix.config.framework == 'pytorch_examples' }}
       run: |
+        python -m pip install peft
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           --make-reports=tests_${{ matrix.config.report }} \
           examples

Makefile

@@ -3,7 +3,7 @@
 # make sure to test the local checkout in scripts and not the pre-installed one (don't use quotes!)
 export PYTHONPATH = src
 
-check_dirs := examples scripts src tests utils
+check_dirs := examples scripts src tests utils benchmarks
 
 modified_only_fixup:
 	$(eval modified_py_files := $(shell python utils/get_modified_files.py $(check_dirs)))

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 15000+ 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 16000+ 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
-- +6000 other amazing GitHub repositories 💪
+- +7000 other amazing GitHub repositories 💪
 
 Thank you for using us ❤️.
 

benchmarks/base_classes.py

@@ -0,0 +1,316 @@
+import os
+import sys
+
+import torch
+
+from diffusers import (
+    AutoPipelineForImage2Image,
+    AutoPipelineForInpainting,
+    AutoPipelineForText2Image,
+    ControlNetModel,
+    LCMScheduler,
+    StableDiffusionAdapterPipeline,
+    StableDiffusionControlNetPipeline,
+    StableDiffusionXLAdapterPipeline,
+    StableDiffusionXLControlNetPipeline,
+    T2IAdapter,
+    WuerstchenCombinedPipeline,
+)
+from diffusers.utils import load_image
+
+
+sys.path.append(".")
+
+from utils import (  # noqa: E402
+    BASE_PATH,
+    PROMPT,
+    BenchmarkInfo,
+    benchmark_fn,
+    bytes_to_giga_bytes,
+    flush,
+    generate_csv_dict,
+    write_to_csv,
+)
+
+
+RESOLUTION_MAPPING = {
+    "runwayml/stable-diffusion-v1-5": (512, 512),
+    "lllyasviel/sd-controlnet-canny": (512, 512),
+    "diffusers/controlnet-canny-sdxl-1.0": (1024, 1024),
+    "TencentARC/t2iadapter_canny_sd14v1": (512, 512),
+    "TencentARC/t2i-adapter-canny-sdxl-1.0": (1024, 1024),
+    "stabilityai/stable-diffusion-2-1": (768, 768),
+    "stabilityai/stable-diffusion-xl-base-1.0": (1024, 1024),
+    "stabilityai/stable-diffusion-xl-refiner-1.0": (1024, 1024),
+    "stabilityai/sdxl-turbo": (512, 512),
+}
+
+
+class BaseBenchmak:
+    pipeline_class = None
+
+    def __init__(self, args):
+        super().__init__()
+
+    def run_inference(self, args):
+        raise NotImplementedError
+
+    def benchmark(self, args):
+        raise NotImplementedError
+
+    def get_result_filepath(self, args):
+        pipeline_class_name = str(self.pipe.__class__.__name__)
+        name = (
+            args.ckpt.replace("/", "_")
+            + "_"
+            + pipeline_class_name
+            + f"-bs@{args.batch_size}-steps@{args.num_inference_steps}-mco@{args.model_cpu_offload}-compile@{args.run_compile}.csv"
+        )
+        filepath = os.path.join(BASE_PATH, name)
+        return filepath
+
+
+class TextToImageBenchmark(BaseBenchmak):
+    pipeline_class = AutoPipelineForText2Image
+
+    def __init__(self, args):
+        pipe = self.pipeline_class.from_pretrained(args.ckpt, torch_dtype=torch.float16)
+        pipe = pipe.to("cuda")
+
+        if args.run_compile:
+            if not isinstance(pipe, WuerstchenCombinedPipeline):
+                pipe.unet.to(memory_format=torch.channels_last)
+                print("Run torch compile")
+                pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+
+                if hasattr(pipe, "movq") and getattr(pipe, "movq", None) is not None:
+                    pipe.movq.to(memory_format=torch.channels_last)
+                    pipe.movq = torch.compile(pipe.movq, mode="reduce-overhead", fullgraph=True)
+            else:
+                print("Run torch compile")
+                pipe.decoder = torch.compile(pipe.decoder, mode="reduce-overhead", fullgraph=True)
+                pipe.vqgan = torch.compile(pipe.vqgan, mode="reduce-overhead", fullgraph=True)
+
+        pipe.set_progress_bar_config(disable=True)
+        self.pipe = pipe
+
+    def run_inference(self, pipe, args):
+        _ = pipe(
+            prompt=PROMPT,
+            num_inference_steps=args.num_inference_steps,
+            num_images_per_prompt=args.batch_size,
+        )
+
+    def benchmark(self, args):
+        flush()
+
+        print(f"[INFO] {self.pipe.__class__.__name__}: Running benchmark with: {vars(args)}\n")
+
+        time = benchmark_fn(self.run_inference, self.pipe, args)  # in seconds.
+        memory = bytes_to_giga_bytes(torch.cuda.max_memory_allocated())  # in GBs.
+        benchmark_info = BenchmarkInfo(time=time, memory=memory)
+
+        pipeline_class_name = str(self.pipe.__class__.__name__)
+        flush()
+        csv_dict = generate_csv_dict(
+            pipeline_cls=pipeline_class_name, ckpt=args.ckpt, args=args, benchmark_info=benchmark_info
+        )
+        filepath = self.get_result_filepath(args)
+        write_to_csv(filepath, csv_dict)
+        print(f"Logs written to: {filepath}")
+        flush()
+
+
+class TurboTextToImageBenchmark(TextToImageBenchmark):
+    def __init__(self, args):
+        super().__init__(args)
+
+    def run_inference(self, pipe, args):
+        _ = pipe(
+            prompt=PROMPT,
+            num_inference_steps=args.num_inference_steps,
+            num_images_per_prompt=args.batch_size,
+            guidance_scale=0.0,
+        )
+
+
+class LCMLoRATextToImageBenchmark(TextToImageBenchmark):
+    lora_id = "latent-consistency/lcm-lora-sdxl"
+
+    def __init__(self, args):
+        super().__init__(args)
+        self.pipe.load_lora_weights(self.lora_id)
+        self.pipe.fuse_lora()
+        self.pipe.scheduler = LCMScheduler.from_config(self.pipe.scheduler.config)
+
+    def get_result_filepath(self, args):
+        pipeline_class_name = str(self.pipe.__class__.__name__)
+        name = (
+            self.lora_id.replace("/", "_")
+            + "_"
+            + pipeline_class_name
+            + f"-bs@{args.batch_size}-steps@{args.num_inference_steps}-mco@{args.model_cpu_offload}-compile@{args.run_compile}.csv"
+        )
+        filepath = os.path.join(BASE_PATH, name)
+        return filepath
+
+    def run_inference(self, pipe, args):
+        _ = pipe(
+            prompt=PROMPT,
+            num_inference_steps=args.num_inference_steps,
+            num_images_per_prompt=args.batch_size,
+            guidance_scale=1.0,
+        )
+
+    def benchmark(self, args):
+        flush()
+
+        print(f"[INFO] {self.pipe.__class__.__name__}: Running benchmark with: {vars(args)}\n")
+
+        time = benchmark_fn(self.run_inference, self.pipe, args)  # in seconds.
+        memory = bytes_to_giga_bytes(torch.cuda.max_memory_allocated())  # in GBs.
+        benchmark_info = BenchmarkInfo(time=time, memory=memory)
+
+        pipeline_class_name = str(self.pipe.__class__.__name__)
+        flush()
+        csv_dict = generate_csv_dict(
+            pipeline_cls=pipeline_class_name, ckpt=self.lora_id, args=args, benchmark_info=benchmark_info
+        )
+        filepath = self.get_result_filepath(args)
+        write_to_csv(filepath, csv_dict)
+        print(f"Logs written to: {filepath}")
+        flush()
+
+
+class ImageToImageBenchmark(TextToImageBenchmark):
+    pipeline_class = AutoPipelineForImage2Image
+    url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/benchmarking/1665_Girl_with_a_Pearl_Earring.jpg"
+    image = load_image(url).convert("RGB")
+
+    def __init__(self, args):
+        super().__init__(args)
+        self.image = self.image.resize(RESOLUTION_MAPPING[args.ckpt])
+
+    def run_inference(self, pipe, args):
+        _ = pipe(
+            prompt=PROMPT,
+            image=self.image,
+            num_inference_steps=args.num_inference_steps,
+            num_images_per_prompt=args.batch_size,
+        )
+
+
+class TurboImageToImageBenchmark(ImageToImageBenchmark):
+    def __init__(self, args):
+        super().__init__(args)
+
+    def run_inference(self, pipe, args):
+        _ = pipe(
+            prompt=PROMPT,
+            image=self.image,
+            num_inference_steps=args.num_inference_steps,
+            num_images_per_prompt=args.batch_size,
+            guidance_scale=0.0,
+            strength=0.5,
+        )
+
+
+class InpaintingBenchmark(ImageToImageBenchmark):
+    pipeline_class = AutoPipelineForInpainting
+    mask_url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/benchmarking/overture-creations-5sI6fQgYIuo_mask.png"
+    mask = load_image(mask_url).convert("RGB")
+
+    def __init__(self, args):
+        super().__init__(args)
+        self.image = self.image.resize(RESOLUTION_MAPPING[args.ckpt])
+        self.mask = self.mask.resize(RESOLUTION_MAPPING[args.ckpt])
+
+    def run_inference(self, pipe, args):
+        _ = pipe(
+            prompt=PROMPT,
+            image=self.image,
+            mask_image=self.mask,
+            num_inference_steps=args.num_inference_steps,
+            num_images_per_prompt=args.batch_size,
+        )
+
+
+class ControlNetBenchmark(TextToImageBenchmark):
+    pipeline_class = StableDiffusionControlNetPipeline
+    aux_network_class = ControlNetModel
+    root_ckpt = "runwayml/stable-diffusion-v1-5"
+
+    url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/benchmarking/canny_image_condition.png"
+    image = load_image(url).convert("RGB")
+
+    def __init__(self, args):
+        aux_network = self.aux_network_class.from_pretrained(args.ckpt, torch_dtype=torch.float16)
+        pipe = self.pipeline_class.from_pretrained(self.root_ckpt, controlnet=aux_network, torch_dtype=torch.float16)
+        pipe = pipe.to("cuda")
+
+        pipe.set_progress_bar_config(disable=True)
+        self.pipe = pipe
+
+        if args.run_compile:
+            pipe.unet.to(memory_format=torch.channels_last)
+            pipe.controlnet.to(memory_format=torch.channels_last)
+
+            print("Run torch compile")
+            pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+            pipe.controlnet = torch.compile(pipe.controlnet, mode="reduce-overhead", fullgraph=True)
+
+        self.image = self.image.resize(RESOLUTION_MAPPING[args.ckpt])
+
+    def run_inference(self, pipe, args):
+        _ = pipe(
+            prompt=PROMPT,
+            image=self.image,
+            num_inference_steps=args.num_inference_steps,
+            num_images_per_prompt=args.batch_size,
+        )
+
+
+class ControlNetSDXLBenchmark(ControlNetBenchmark):
+    pipeline_class = StableDiffusionXLControlNetPipeline
+    root_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+
+    def __init__(self, args):
+        super().__init__(args)
+
+
+class T2IAdapterBenchmark(ControlNetBenchmark):
+    pipeline_class = StableDiffusionAdapterPipeline
+    aux_network_class = T2IAdapter
+    root_ckpt = "CompVis/stable-diffusion-v1-4"
+
+    url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/benchmarking/canny_for_adapter.png"
+    image = load_image(url).convert("L")
+
+    def __init__(self, args):
+        aux_network = self.aux_network_class.from_pretrained(args.ckpt, torch_dtype=torch.float16)
+        pipe = self.pipeline_class.from_pretrained(self.root_ckpt, adapter=aux_network, torch_dtype=torch.float16)
+        pipe = pipe.to("cuda")
+
+        pipe.set_progress_bar_config(disable=True)
+        self.pipe = pipe
+
+        if args.run_compile:
+            pipe.unet.to(memory_format=torch.channels_last)
+            pipe.adapter.to(memory_format=torch.channels_last)
+
+            print("Run torch compile")
+            pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+            pipe.adapter = torch.compile(pipe.adapter, mode="reduce-overhead", fullgraph=True)
+
+        self.image = self.image.resize(RESOLUTION_MAPPING[args.ckpt])
+
+
+class T2IAdapterSDXLBenchmark(T2IAdapterBenchmark):
+    pipeline_class = StableDiffusionXLAdapterPipeline
+    root_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+
+    url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/benchmarking/canny_for_adapter_sdxl.png"
+    image = load_image(url)
+
+    def __init__(self, args):
+        super().__init__(args)

benchmarks/benchmark_controlnet.py

@@ -0,0 +1,26 @@
+import argparse
+import sys
+
+
+sys.path.append(".")
+from base_classes import ControlNetBenchmark, ControlNetSDXLBenchmark  # noqa: E402
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--ckpt",
+        type=str,
+        default="lllyasviel/sd-controlnet-canny",
+        choices=["lllyasviel/sd-controlnet-canny", "diffusers/controlnet-canny-sdxl-1.0"],
+    )
+    parser.add_argument("--batch_size", type=int, default=1)
+    parser.add_argument("--num_inference_steps", type=int, default=50)
+    parser.add_argument("--model_cpu_offload", action="store_true")
+    parser.add_argument("--run_compile", action="store_true")
+    args = parser.parse_args()
+
+    benchmark_pipe = (
+        ControlNetBenchmark(args) if args.ckpt == "lllyasviel/sd-controlnet-canny" else ControlNetSDXLBenchmark(args)
+    )
+    benchmark_pipe.benchmark(args)

benchmarks/benchmark_sd_img.py

@@ -0,0 +1,29 @@
+import argparse
+import sys
+
+
+sys.path.append(".")
+from base_classes import ImageToImageBenchmark, TurboImageToImageBenchmark  # noqa: E402
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--ckpt",
+        type=str,
+        default="runwayml/stable-diffusion-v1-5",
+        choices=[
+            "runwayml/stable-diffusion-v1-5",
+            "stabilityai/stable-diffusion-2-1",
+            "stabilityai/stable-diffusion-xl-refiner-1.0",
+            "stabilityai/sdxl-turbo",
+        ],
+    )
+    parser.add_argument("--batch_size", type=int, default=1)
+    parser.add_argument("--num_inference_steps", type=int, default=50)
+    parser.add_argument("--model_cpu_offload", action="store_true")
+    parser.add_argument("--run_compile", action="store_true")
+    args = parser.parse_args()
+
+    benchmark_pipe = ImageToImageBenchmark(args) if "turbo" not in args.ckpt else TurboImageToImageBenchmark(args)
+    benchmark_pipe.benchmark(args)

benchmarks/benchmark_sd_inpainting.py

@@ -0,0 +1,28 @@
+import argparse
+import sys
+
+
+sys.path.append(".")
+from base_classes import InpaintingBenchmark  # noqa: E402
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--ckpt",
+        type=str,
+        default="runwayml/stable-diffusion-v1-5",
+        choices=[
+            "runwayml/stable-diffusion-v1-5",
+            "stabilityai/stable-diffusion-2-1",
+            "stabilityai/stable-diffusion-xl-base-1.0",
+        ],
+    )
+    parser.add_argument("--batch_size", type=int, default=1)
+    parser.add_argument("--num_inference_steps", type=int, default=50)
+    parser.add_argument("--model_cpu_offload", action="store_true")
+    parser.add_argument("--run_compile", action="store_true")
+    args = parser.parse_args()
+
+    benchmark_pipe = InpaintingBenchmark(args)
+    benchmark_pipe.benchmark(args)

benchmarks/benchmark_t2i_adapter.py

@@ -0,0 +1,28 @@
+import argparse
+import sys
+
+
+sys.path.append(".")
+from base_classes import T2IAdapterBenchmark, T2IAdapterSDXLBenchmark  # noqa: E402
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--ckpt",
+        type=str,
+        default="TencentARC/t2iadapter_canny_sd14v1",
+        choices=["TencentARC/t2iadapter_canny_sd14v1", "TencentARC/t2i-adapter-canny-sdxl-1.0"],
+    )
+    parser.add_argument("--batch_size", type=int, default=1)
+    parser.add_argument("--num_inference_steps", type=int, default=50)
+    parser.add_argument("--model_cpu_offload", action="store_true")
+    parser.add_argument("--run_compile", action="store_true")
+    args = parser.parse_args()
+
+    benchmark_pipe = (
+        T2IAdapterBenchmark(args)
+        if args.ckpt == "TencentARC/t2iadapter_canny_sd14v1"
+        else T2IAdapterSDXLBenchmark(args)
+    )
+    benchmark_pipe.benchmark(args)

benchmarks/benchmark_t2i_lcm_lora.py

@@ -0,0 +1,23 @@
+import argparse
+import sys
+
+
+sys.path.append(".")
+from base_classes import LCMLoRATextToImageBenchmark  # noqa: E402
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--ckpt",
+        type=str,
+        default="stabilityai/stable-diffusion-xl-base-1.0",
+    )
+    parser.add_argument("--batch_size", type=int, default=1)
+    parser.add_argument("--num_inference_steps", type=int, default=4)
+    parser.add_argument("--model_cpu_offload", action="store_true")
+    parser.add_argument("--run_compile", action="store_true")
+    args = parser.parse_args()
+
+    benchmark_pipe = LCMLoRATextToImageBenchmark(args)
+    benchmark_pipe.benchmark(args)

benchmarks/benchmark_text_to_image.py

@@ -0,0 +1,40 @@
+import argparse
+import sys
+
+
+sys.path.append(".")
+from base_classes import TextToImageBenchmark, TurboTextToImageBenchmark  # noqa: E402
+
+
+ALL_T2I_CKPTS = [
+    "runwayml/stable-diffusion-v1-5",
+    "segmind/SSD-1B",
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    "kandinsky-community/kandinsky-2-2-decoder",
+    "warp-ai/wuerstchen",
+    "stabilityai/sdxl-turbo",
+]
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--ckpt",
+        type=str,
+        default="runwayml/stable-diffusion-v1-5",
+        choices=ALL_T2I_CKPTS,
+    )
+    parser.add_argument("--batch_size", type=int, default=1)
+    parser.add_argument("--num_inference_steps", type=int, default=50)
+    parser.add_argument("--model_cpu_offload", action="store_true")
+    parser.add_argument("--run_compile", action="store_true")
+    args = parser.parse_args()
+
+    benchmark_cls = None
+    if "turbo" in args.ckpt:
+        benchmark_cls = TurboTextToImageBenchmark
+    else:
+        benchmark_cls = TextToImageBenchmark
+
+    benchmark_pipe = benchmark_cls(args)
+    benchmark_pipe.benchmark(args)

benchmarks/push_results.py

@@ -0,0 +1,72 @@
+import glob
+import sys
+
+import pandas as pd
+from huggingface_hub import hf_hub_download, upload_file
+from huggingface_hub.utils._errors import EntryNotFoundError
+
+
+sys.path.append(".")
+from utils import BASE_PATH, FINAL_CSV_FILE, GITHUB_SHA, REPO_ID, collate_csv  # noqa: E402
+
+
+def has_previous_benchmark() -> str:
+    csv_path = None
+    try:
+        csv_path = hf_hub_download(repo_id=REPO_ID, repo_type="dataset", filename=FINAL_CSV_FILE)
+    except EntryNotFoundError:
+        csv_path = None
+    return csv_path
+
+
+def filter_float(value):
+    if isinstance(value, str):
+        return float(value.split()[0])
+    return value
+
+
+def push_to_hf_dataset():
+    all_csvs = sorted(glob.glob(f"{BASE_PATH}/*.csv"))
+    collate_csv(all_csvs, FINAL_CSV_FILE)
+
+    # If there's an existing benchmark file, we should report the changes.
+    csv_path = has_previous_benchmark()
+    if csv_path is not None:
+        current_results = pd.read_csv(FINAL_CSV_FILE)
+        previous_results = pd.read_csv(csv_path)
+
+        numeric_columns = current_results.select_dtypes(include=["float64", "int64"]).columns
+        numeric_columns = [
+            c for c in numeric_columns if c not in ["batch_size", "num_inference_steps", "actual_gpu_memory (gbs)"]
+        ]
+
+        for column in numeric_columns:
+            previous_results[column] = previous_results[column].map(lambda x: filter_float(x))
+
+            # Calculate the percentage change
+            current_results[column] = current_results[column].astype(float)
+            previous_results[column] = previous_results[column].astype(float)
+            percent_change = ((current_results[column] - previous_results[column]) / previous_results[column]) * 100
+
+            # Format the values with '+' or '-' sign and append to original values
+            current_results[column] = current_results[column].map(str) + percent_change.map(
+                lambda x: f" ({'+' if x > 0 else ''}{x:.2f}%)"
+            )
+            # There might be newly added rows. So, filter out the NaNs.
+            current_results[column] = current_results[column].map(lambda x: x.replace(" (nan%)", ""))
+
+        # Overwrite the current result file.
+        current_results.to_csv(FINAL_CSV_FILE, index=False)
+
+    commit_message = f"upload from sha: {GITHUB_SHA}" if GITHUB_SHA is not None else "upload benchmark results"
+    upload_file(
+        repo_id=REPO_ID,
+        path_in_repo=FINAL_CSV_FILE,
+        path_or_fileobj=FINAL_CSV_FILE,
+        repo_type="dataset",
+        commit_message=commit_message,
+    )
+
+
+if __name__ == "__main__":
+    push_to_hf_dataset()

benchmarks/run_all.py

@@ -0,0 +1,97 @@
+import glob
+import subprocess
+import sys
+from typing import List
+
+
+sys.path.append(".")
+from benchmark_text_to_image import ALL_T2I_CKPTS  # noqa: E402
+
+
+PATTERN = "benchmark_*.py"
+
+
+class SubprocessCallException(Exception):
+    pass
+
+
+# Taken from `test_examples_utils.py`
+def run_command(command: List[str], return_stdout=False):
+    """
+    Runs `command` with `subprocess.check_output` and will potentially return the `stdout`. Will also properly capture
+    if an error occurred while running `command`
+    """
+    try:
+        output = subprocess.check_output(command, stderr=subprocess.STDOUT)
+        if return_stdout:
+            if hasattr(output, "decode"):
+                output = output.decode("utf-8")
+            return output
+    except subprocess.CalledProcessError as e:
+        raise SubprocessCallException(
+            f"Command `{' '.join(command)}` failed with the following error:\n\n{e.output.decode()}"
+        ) from e
+
+
+def main():
+    python_files = glob.glob(PATTERN)
+
+    for file in python_files:
+        print(f"****** Running file: {file} ******")
+
+        # Run with canonical settings.
+        if file != "benchmark_text_to_image.py":
+            command = f"python {file}"
+            run_command(command.split())
+
+            command += " --run_compile"
+            run_command(command.split())
+
+    # Run variants.
+    for file in python_files:
+        if file == "benchmark_text_to_image.py":
+            for ckpt in ALL_T2I_CKPTS:
+                command = f"python {file} --ckpt {ckpt}"
+
+                if "turbo" in ckpt:
+                    command += " --num_inference_steps 1"
+
+                run_command(command.split())
+
+                command += " --run_compile"
+                run_command(command.split())
+
+        elif file == "benchmark_sd_img.py":
+            for ckpt in ["stabilityai/stable-diffusion-xl-refiner-1.0", "stabilityai/sdxl-turbo"]:
+                command = f"python {file} --ckpt {ckpt}"
+
+                if ckpt == "stabilityai/sdxl-turbo":
+                    command += " --num_inference_steps 2"
+
+                run_command(command.split())
+                command += " --run_compile"
+                run_command(command.split())
+
+        elif file == "benchmark_sd_inpainting.py":
+            sdxl_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+            command = f"python {file} --ckpt {sdxl_ckpt}"
+            run_command(command.split())
+
+            command += " --run_compile"
+            run_command(command.split())
+
+        elif file in ["benchmark_controlnet.py", "benchmark_t2i_adapter.py"]:
+            sdxl_ckpt = (
+                "diffusers/controlnet-canny-sdxl-1.0"
+                if "controlnet" in file
+                else "TencentARC/t2i-adapter-canny-sdxl-1.0"
+            )
+            command = f"python {file} --ckpt {sdxl_ckpt}"
+            run_command(command.split())
+
+            command += " --run_compile"
+            run_command(command.split())
+
+
+if __name__ == "__main__":
+    main()

benchmarks/utils.py

@@ -0,0 +1,98 @@
+import argparse
+import csv
+import gc
+import os
+from dataclasses import dataclass
+from typing import Dict, List, Union
+
+import torch
+import torch.utils.benchmark as benchmark
+
+
+GITHUB_SHA = os.getenv("GITHUB_SHA", None)
+BENCHMARK_FIELDS = [
+    "pipeline_cls",
+    "ckpt_id",
+    "batch_size",
+    "num_inference_steps",
+    "model_cpu_offload",
+    "run_compile",
+    "time (secs)",
+    "memory (gbs)",
+    "actual_gpu_memory (gbs)",
+    "github_sha",
+]
+
+PROMPT = "ghibli style, a fantasy landscape with castles"
+BASE_PATH = os.getenv("BASE_PATH", ".")
+TOTAL_GPU_MEMORY = float(os.getenv("TOTAL_GPU_MEMORY", torch.cuda.get_device_properties(0).total_memory / (1024**3)))
+
+REPO_ID = "diffusers/benchmarks"
+FINAL_CSV_FILE = "collated_results.csv"
+
+
+@dataclass
+class BenchmarkInfo:
+    time: float
+    memory: float
+
+
+def flush():
+    """Wipes off memory."""
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+
+def bytes_to_giga_bytes(bytes):
+    return f"{(bytes / 1024 / 1024 / 1024):.3f}"
+
+
+def benchmark_fn(f, *args, **kwargs):
+    t0 = benchmark.Timer(
+        stmt="f(*args, **kwargs)",
+        globals={"args": args, "kwargs": kwargs, "f": f},
+        num_threads=torch.get_num_threads(),
+    )
+    return f"{(t0.blocked_autorange().mean):.3f}"
+
+
+def generate_csv_dict(
+    pipeline_cls: str, ckpt: str, args: argparse.Namespace, benchmark_info: BenchmarkInfo
+) -> Dict[str, Union[str, bool, float]]:
+    """Packs benchmarking data into a dictionary for latter serialization."""
+    data_dict = {
+        "pipeline_cls": pipeline_cls,
+        "ckpt_id": ckpt,
+        "batch_size": args.batch_size,
+        "num_inference_steps": args.num_inference_steps,
+        "model_cpu_offload": args.model_cpu_offload,
+        "run_compile": args.run_compile,
+        "time (secs)": benchmark_info.time,
+        "memory (gbs)": benchmark_info.memory,
+        "actual_gpu_memory (gbs)": f"{(TOTAL_GPU_MEMORY):.3f}",
+        "github_sha": GITHUB_SHA,
+    }
+    return data_dict
+
+
+def write_to_csv(file_name: str, data_dict: Dict[str, Union[str, bool, float]]):
+    """Serializes a dictionary into a CSV file."""
+    with open(file_name, mode="w", newline="") as csvfile:
+        writer = csv.DictWriter(csvfile, fieldnames=BENCHMARK_FIELDS)
+        writer.writeheader()
+        writer.writerow(data_dict)
+
+
+def collate_csv(input_files: List[str], output_file: str):
+    """Collates multiple identically structured CSVs into a single CSV file."""
+    with open(output_file, mode="w", newline="") as outfile:
+        writer = csv.DictWriter(outfile, fieldnames=BENCHMARK_FIELDS)
+        writer.writeheader()
+
+        for file in input_files:
+            with open(file, mode="r") as infile:
+                reader = csv.DictReader(infile)
+                for row in reader:
+                    writer.writerow(row)

docs/source/en/_toctree.yml

@@ -198,6 +198,8 @@
       title: Outputs
     title: Main Classes
   - sections:
+    - local: api/loaders/ip_adapter
+      title: IP-Adapter
     - local: api/loaders/lora
       title: LoRA
     - local: api/loaders/single_file
@@ -242,14 +244,10 @@
   - sections:
     - local: api/pipelines/overview
       title: Overview
-    - local: api/pipelines/alt_diffusion
-      title: AltDiffusion
     - local: api/pipelines/animatediff
       title: AnimateDiff
     - local: api/pipelines/attend_and_excite
       title: Attend-and-Excite
-    - local: api/pipelines/audio_diffusion
-      title: Audio Diffusion
     - local: api/pipelines/audioldm
       title: AudioLDM
     - local: api/pipelines/audioldm2
@@ -264,8 +262,10 @@
       title: ControlNet
     - local: api/pipelines/controlnet_sdxl
       title: ControlNet with Stable Diffusion XL
-    - local: api/pipelines/cycle_diffusion
-      title: Cycle Diffusion
+    - local: api/pipelines/controlnetxs
+      title: ControlNet-XS
+    - local: api/pipelines/controlnetxs_sdxl
+      title: ControlNet-XS with Stable Diffusion XL
     - local: api/pipelines/dance_diffusion
       title: Dance Diffusion
     - local: api/pipelines/ddim
@@ -296,26 +296,14 @@
       title: MusicLDM
     - local: api/pipelines/paint_by_example
       title: Paint by Example
-    - local: api/pipelines/paradigms
-      title: Parallel Sampling of Diffusion Models
-    - local: api/pipelines/pix2pix_zero
-      title: Pix2Pix Zero
     - local: api/pipelines/pixart
       title: PixArt-α
-    - local: api/pipelines/pndm
-      title: PNDM
-    - local: api/pipelines/repaint
-      title: RePaint
-    - local: api/pipelines/score_sde_ve
-      title: Score SDE VE
     - local: api/pipelines/self_attention_guidance
       title: Self-Attention Guidance
     - local: api/pipelines/semantic_stable_diffusion
       title: Semantic Guidance
     - local: api/pipelines/shap_e
       title: Shap-E
-    - local: api/pipelines/spectrogram_diffusion
-      title: Spectrogram Diffusion
     - sections:
       - local: api/pipelines/stable_diffusion/overview
         title: Overview
@@ -350,26 +338,16 @@
       title: Stable Diffusion
     - local: api/pipelines/stable_unclip
       title: Stable unCLIP
-    - local: api/pipelines/stochastic_karras_ve
-      title: Stochastic Karras VE
-    - local: api/pipelines/model_editing
-      title: Text-to-image model editing
     - local: api/pipelines/text_to_video
       title: Text-to-video
     - local: api/pipelines/text_to_video_zero
       title: Text2Video-Zero
     - local: api/pipelines/unclip
       title: unCLIP
-    - local: api/pipelines/latent_diffusion_uncond
-      title: Unconditional Latent Diffusion
     - local: api/pipelines/unidiffuser
       title: UniDiffuser
     - local: api/pipelines/value_guided_sampling
       title: Value-guided sampling
-    - local: api/pipelines/versatile_diffusion
-      title: Versatile Diffusion
-    - local: api/pipelines/vq_diffusion
-      title: VQ Diffusion
     - local: api/pipelines/wuerstchen
       title: Wuerstchen
     title: Pipelines

docs/source/en/api/attnprocessor.md

@@ -20,6 +20,9 @@ An attention processor is a class for applying different types of attention mech
 ## AttnProcessor2_0
 [[autodoc]] models.attention_processor.AttnProcessor2_0
 
+## FusedAttnProcessor2_0
+[[autodoc]] models.attention_processor.FusedAttnProcessor2_0
+
 ## LoRAAttnProcessor
 [[autodoc]] models.attention_processor.LoRAAttnProcessor
 

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

@@ -0,0 +1,25 @@
+<!--Copyright 2023 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.
+-->
+
+# IP-Adapter
+
+[IP-Adapter](https://hf.co/papers/2308.06721) is a lightweight adapter that enables prompting a diffusion model with an image. This method decouples the cross-attention layers of the image and text features. The image features are generated from an image encoder. Files generated from IP-Adapter are only ~100MBs.
+
+<Tip>
+
+Learn how to load an IP-Adapter checkpoint and image in the [IP-Adapter](../../using-diffusers/loading_adapters#ip-adapter) loading guide.
+
+</Tip>
+
+## IPAdapterMixin
+
+[[autodoc]] loaders.ip_adapter.IPAdapterMixin

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

@@ -49,12 +49,12 @@ make_image_grid([original_image, mask_image, image], rows=1, cols=3)
 
 ## AsymmetricAutoencoderKL
 
-[[autodoc]] models.autoencoder_asym_kl.AsymmetricAutoencoderKL
+[[autodoc]] models.autoencoders.autoencoder_asym_kl.AsymmetricAutoencoderKL
 
 ## AutoencoderKLOutput
 
-[[autodoc]] models.autoencoder_kl.AutoencoderKLOutput
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
 
 ## DecoderOutput
 
-[[autodoc]] models.vae.DecoderOutput
+[[autodoc]] models.autoencoders.vae.DecoderOutput

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

@@ -54,4 +54,4 @@ image
 
 ## AutoencoderTinyOutput
 
-[[autodoc]] models.autoencoder_tiny.AutoencoderTinyOutput
+[[autodoc]] models.autoencoders.autoencoder_tiny.AutoencoderTinyOutput

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

@@ -36,11 +36,11 @@ model = AutoencoderKL.from_single_file(url)
 
 ## AutoencoderKLOutput
 
-[[autodoc]] models.autoencoder_kl.AutoencoderKLOutput
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
 
 ## DecoderOutput
 
-[[autodoc]] models.vae.DecoderOutput
+[[autodoc]] models.autoencoders.vae.DecoderOutput
 
 ## FlaxAutoencoderKL
 

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

@@ -1,47 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# AltDiffusion
-
-AltDiffusion was proposed in [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://huggingface.co/papers/2211.06679) by Zhongzhi Chen, Guang Liu, Bo-Wen Zhang, Fulong Ye, Qinghong Yang, Ledell Wu.
-
-The abstract from the paper is:
-
-*In this work, we present a conceptually simple and effective method to train a strong bilingual/multilingual multimodal representation model. Starting from the pre-trained multimodal representation model CLIP released by OpenAI, we altered its text encoder with a pre-trained multilingual text encoder XLM-R, and aligned both languages and image representations by a two-stage training schema consisting of teacher learning and contrastive learning. We validate our method through evaluations of a wide range of tasks. We set new state-of-the-art performances on a bunch of tasks including ImageNet-CN, Flicker30k-CN, COCO-CN and XTD. Further, we obtain very close performances with CLIP on almost all tasks, suggesting that one can simply alter the text encoder in CLIP for extended capabilities such as multilingual understanding. Our models and code are available at [this https URL](https://github.com/FlagAI-Open/FlagAI).*
-
-## Tips
-
-`AltDiffusion` is conceptually the same as [Stable Diffusion](./stable_diffusion/overview).
-
-<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.
-
-</Tip>
-
-## AltDiffusionPipeline
-
-[[autodoc]] AltDiffusionPipeline
-	- all
-	- __call__
-
-## AltDiffusionImg2ImgPipeline
-
-[[autodoc]] AltDiffusionImg2ImgPipeline
-	- all
-	- __call__
-
-## AltDiffusionPipelineOutput
-
-[[autodoc]] pipelines.alt_diffusion.AltDiffusionPipelineOutput
-	- all
-	- __call__

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

@@ -1,35 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# Audio Diffusion
-
-[Audio Diffusion](https://github.com/teticio/audio-diffusion) is by Robert Dargavel Smith, and it leverages the recent advances in image generation from diffusion models by converting audio samples to and from Mel spectrogram images.
-
-<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.
-
-</Tip>
-
-## AudioDiffusionPipeline
-[[autodoc]] AudioDiffusionPipeline
-	- all
-	- __call__
-
-## AudioPipelineOutput
-[[autodoc]] pipelines.AudioPipelineOutput
-
-## ImagePipelineOutput
-[[autodoc]] pipelines.ImagePipelineOutput
-
-## Mel
-[[autodoc]] Mel

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

@@ -0,0 +1,39 @@
+<!--Copyright 2023 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.
+-->
+
+# ControlNet-XS
+
+ControlNet-XS was introduced in [ControlNet-XS](https://vislearn.github.io/ControlNet-XS/) by Denis Zavadski and Carsten Rother. It is based on the observation that the control model in the [original ControlNet](https://huggingface.co/papers/2302.05543) can be made much smaller and still produce good results.
+
+Like the original ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
+
+ControlNet-XS generates images with comparable quality to a regular ControlNet, but it is 20-25% faster ([see benchmark](https://github.com/UmerHA/controlnet-xs-benchmark/blob/main/Speed%20Benchmark.ipynb) with StableDiffusion-XL) and uses ~45% less memory.
+
+Here's the overview from the [project page](https://vislearn.github.io/ControlNet-XS/):
+
+*With increasing computing capabilities, current model architectures appear to follow the trend of simply upscaling all components without validating the necessity for doing so. In this project we investigate the size and architectural design of ControlNet [Zhang et al., 2023] for controlling the image generation process with stable diffusion-based models. We show that a new architecture with as little as 1% of the parameters of the base model achieves state-of-the art results, considerably better than ControlNet in terms of FID score. Hence we call it ControlNet-XS. We provide the code for controlling StableDiffusion-XL [Podell et al., 2023] (Model B, 48M Parameters) and StableDiffusion 2.1 [Rombach et al. 2022] (Model B, 14M Parameters), all under openrail license.*
+
+This model was contributed by [UmerHA](https://twitter.com/UmerHAdil). ❤️
+
+<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.
+
+</Tip>
+
+## StableDiffusionControlNetXSPipeline
+[[autodoc]] StableDiffusionControlNetXSPipeline
+	- all
+	- __call__
+
+## StableDiffusionPipelineOutput
+[[autodoc]] pipelines.stable_diffusion.StableDiffusionPipelineOutput

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

@@ -0,0 +1,45 @@
+<!--Copyright 2023 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.
+-->
+
+# ControlNet-XS with Stable Diffusion XL
+
+ControlNet-XS was introduced in [ControlNet-XS](https://vislearn.github.io/ControlNet-XS/) by Denis Zavadski and Carsten Rother. It is based on the observation that the control model in the [original ControlNet](https://huggingface.co/papers/2302.05543) can be made much smaller and still produce good results.
+
+Like the original ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
+
+ControlNet-XS generates images with comparable quality to a regular ControlNet, but it is 20-25% faster ([see benchmark](https://github.com/UmerHA/controlnet-xs-benchmark/blob/main/Speed%20Benchmark.ipynb)) and uses ~45% less memory.
+
+Here's the overview from the [project page](https://vislearn.github.io/ControlNet-XS/):
+
+*With increasing computing capabilities, current model architectures appear to follow the trend of simply upscaling all components without validating the necessity for doing so. In this project we investigate the size and architectural design of ControlNet [Zhang et al., 2023] for controlling the image generation process with stable diffusion-based models. We show that a new architecture with as little as 1% of the parameters of the base model achieves state-of-the art results, considerably better than ControlNet in terms of FID score. Hence we call it ControlNet-XS. We provide the code for controlling StableDiffusion-XL [Podell et al., 2023] (Model B, 48M Parameters) and StableDiffusion 2.1 [Rombach et al. 2022] (Model B, 14M Parameters), all under openrail license.*
+
+This model was contributed by [UmerHA](https://twitter.com/UmerHAdil). ❤️
+
+<Tip warning={true}>
+
+🧪 Many of the SDXL ControlNet checkpoints are experimental, and there is a lot of room for improvement. Feel free to open an [Issue](https://github.com/huggingface/diffusers/issues/new/choose) and leave us feedback on how we can improve!
+
+</Tip>
+
+<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.
+
+</Tip>
+
+## StableDiffusionXLControlNetXSPipeline
+[[autodoc]] StableDiffusionXLControlNetXSPipeline
+	- all
+	- __call__
+
+## StableDiffusionPipelineOutput
+[[autodoc]] pipelines.stable_diffusion.StableDiffusionPipelineOutput

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

@@ -1,33 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# Cycle Diffusion
-
-Cycle Diffusion is a text guided image-to-image generation model proposed in [Unifying Diffusion Models' Latent Space, with Applications to CycleDiffusion and Guidance](https://huggingface.co/papers/2210.05559) by Chen Henry Wu, Fernando De la Torre.
-
-The abstract from the paper is:
-
-*Diffusion models have achieved unprecedented performance in generative modeling. The commonly-adopted formulation of the latent code of diffusion models is a sequence of gradually denoised samples, as opposed to the simpler (e.g., Gaussian) latent space of GANs, VAEs, and normalizing flows. This paper provides an alternative, Gaussian formulation of the latent space of various diffusion models, as well as an invertible DPM-Encoder that maps images into the latent space. While our formulation is purely based on the definition of diffusion models, we demonstrate several intriguing consequences. (1) Empirically, we observe that a common latent space emerges from two diffusion models trained independently on related domains. In light of this finding, we propose CycleDiffusion, which uses DPM-Encoder for unpaired image-to-image translation. Furthermore, applying CycleDiffusion to text-to-image diffusion models, we show that large-scale text-to-image diffusion models can be used as zero-shot image-to-image editors. (2) One can guide pre-trained diffusion models and GANs by controlling the latent codes in a unified, plug-and-play formulation based on energy-based models. Using the CLIP model and a face recognition model as guidance, we demonstrate that diffusion models have better coverage of low-density sub-populations and individuals than GANs. The code is publicly available at [this https URL](https://github.com/ChenWu98/cycle-diffusion).*
-
-<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.
-
-</Tip>
-
-## CycleDiffusionPipeline
-[[autodoc]] CycleDiffusionPipeline
-	- all
-	- __call__
-
-## StableDiffusionPiplineOutput
-[[autodoc]] pipelines.stable_diffusion.StableDiffusionPipelineOutput

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

@@ -1,35 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# Unconditional Latent Diffusion
-
-Unconditional Latent Diffusion was proposed in [High-Resolution Image Synthesis with Latent Diffusion Models](https://huggingface.co/papers/2112.10752) by Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Björn Ommer.
-
-The abstract from the paper is:
-
-*By decomposing the image formation process into a sequential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results on image data and beyond. Additionally, their formulation allows for a guiding mechanism to control the image generation process without retraining. However, since these models typically operate directly in pixel space, optimization of powerful DMs often consumes hundreds of GPU days and inference is expensive due to sequential evaluations. To enable DM training on limited computational resources while retaining their quality and flexibility, we apply them in the latent space of powerful pretrained autoencoders. In contrast to previous work, training diffusion models on such a representation allows for the first time to reach a near-optimal point between complexity reduction and detail preservation, greatly boosting visual fidelity. By introducing cross-attention layers into the model architecture, we turn diffusion models into powerful and flexible generators for general conditioning inputs such as text or bounding boxes and high-resolution synthesis becomes possible in a convolutional manner. Our latent diffusion models (LDMs) achieve a new state of the art for image inpainting and highly competitive performance on various tasks, including unconditional image generation, semantic scene synthesis, and super-resolution, while significantly reducing computational requirements compared to pixel-based DMs.*
-
-The original codebase can be found at [CompVis/latent-diffusion](https://github.com/CompVis/latent-diffusion).
-
-<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.
-
-</Tip>
-
-## LDMPipeline
-[[autodoc]] LDMPipeline
-	- all
-	- __call__
-
-## ImagePipelineOutput
-[[autodoc]] pipelines.ImagePipelineOutput

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

@@ -1,35 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# Text-to-image model editing
-
-[Editing Implicit Assumptions in Text-to-Image Diffusion Models](https://huggingface.co/papers/2303.08084) is by Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. This pipeline enables editing diffusion model weights, such that its assumptions of a given concept are changed. The resulting change is expected to take effect in all prompt generations related to the edited concept.
-
-The abstract from the paper is:
-
-*Text-to-image diffusion models often make implicit assumptions about the world when generating images. While some assumptions are useful (e.g., the sky is blue), they can also be outdated, incorrect, or reflective of social biases present in the training data. Thus, there is a need to control these assumptions without requiring explicit user input or costly re-training. In this work, we aim to edit a given implicit assumption in a pre-trained diffusion model. Our Text-to-Image Model Editing method, TIME for short, receives a pair of inputs: a "source" under-specified prompt for which the model makes an implicit assumption (e.g., "a pack of roses"), and a "destination" prompt that describes the same setting, but with a specified desired attribute (e.g., "a pack of blue roses"). TIME then updates the model's cross-attention layers, as these layers assign visual meaning to textual tokens. We edit the projection matrices in these layers such that the source prompt is projected close to the destination prompt. Our method is highly efficient, as it modifies a mere 2.2% of the model's parameters in under one second. To evaluate model editing approaches, we introduce TIMED (TIME Dataset), containing 147 source and destination prompt pairs from various domains. Our experiments (using Stable Diffusion) show that TIME is successful in model editing, generalizes well for related prompts unseen during editing, and imposes minimal effect on unrelated generations.*
-
-You can find additional information about model editing on the [project page](https://time-diffusion.github.io/), [original codebase](https://github.com/bahjat-kawar/time-diffusion), and try it out in a [demo](https://huggingface.co/spaces/bahjat-kawar/time-diffusion).
-
-<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.
-
-</Tip>
-
-## StableDiffusionModelEditingPipeline
-[[autodoc]] StableDiffusionModelEditingPipeline
-	- __call__
-	- all
-
-## StableDiffusionPipelineOutput
-[[autodoc]] pipelines.stable_diffusion.StableDiffusionPipelineOutput

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

@@ -40,6 +40,8 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [Consistency Models](consistency_models) | unconditional image generation |
 | [ControlNet](controlnet) | text2image, image2image, inpainting |
 | [ControlNet with Stable Diffusion XL](controlnet_sdxl) | text2image |
+| [ControlNet-XS](controlnetxs) | text2image |
+| [ControlNet-XS with Stable Diffusion XL](controlnetxs_sdxl) | text2image |
 | [Cycle Diffusion](cycle_diffusion) | image2image |
 | [Dance Diffusion](dance_diffusion) | unconditional audio generation |
 | [DDIM](ddim) | unconditional image generation |
@@ -71,6 +73,7 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [Stable Diffusion](stable_diffusion/overview) | text2image, image2image, depth2image, inpainting, image variation, latent upscaler, super-resolution |
 | [Stable Diffusion Model Editing](model_editing) | model editing |
 | [Stable Diffusion XL](stable_diffusion/stable_diffusion_xl) | text2image, image2image, inpainting |
+| [Stable Diffusion XL Turbo](stable_diffusion/sdxl_turbo) | text2image, image2image, inpainting |
 | [Stable unCLIP](stable_unclip) | text2image, image variation |
 | [Stochastic Karras VE](stochastic_karras_ve) | unconditional image generation |
 | [T2I-Adapter](stable_diffusion/adapter) | text2image |

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

@@ -1,51 +0,0 @@
-<!--Copyright 2023 ParaDiGMS 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.
--->
-
-# Parallel Sampling of Diffusion Models
-
-[Parallel Sampling of Diffusion Models](https://huggingface.co/papers/2305.16317) is by Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, Nima Anari.
-
-The abstract from the paper is:
-
-*Diffusion models are powerful generative models but suffer from slow sampling, often taking 1000 sequential denoising steps for one sample. As a result, considerable efforts have been directed toward reducing the number of denoising steps, but these methods hurt sample quality. Instead of reducing the number of denoising steps (trading quality for speed), in this paper we explore an orthogonal approach: can we run the denoising steps in parallel (trading compute for speed)? In spite of the sequential nature of the denoising steps, we show that surprisingly it is possible to parallelize sampling via Picard iterations, by guessing the solution of future denoising steps and iteratively refining until convergence. With this insight, we present ParaDiGMS, a novel method to accelerate the sampling of pretrained diffusion models by denoising multiple steps in parallel. ParaDiGMS is the first diffusion sampling method that enables trading compute for speed and is even compatible with existing fast sampling techniques such as DDIM and DPMSolver. Using ParaDiGMS, we improve sampling speed by 2-4x across a range of robotics and image generation models, giving state-of-the-art sampling speeds of 0.2s on 100-step DiffusionPolicy and 14.6s on 1000-step StableDiffusion-v2 with no measurable degradation of task reward, FID score, or CLIP score.*
-
-The original codebase can be found at [AndyShih12/paradigms](https://github.com/AndyShih12/paradigms), and the pipeline was contributed by [AndyShih12](https://github.com/AndyShih12). ❤️
-
-## Tips
-
-This pipeline improves sampling speed by running denoising steps in parallel, at the cost of increased total FLOPs.
-Therefore, it is better to call this pipeline when running on multiple GPUs. Otherwise, without enough GPU bandwidth
-sampling may be even slower than sequential sampling.
-
-The two parameters to play with are `parallel` (batch size) and `tolerance`.
-- If it fits in memory, for a 1000-step DDPM you can aim for a batch size of around 100 (for example, 8 GPUs and `batch_per_device=12` to get `parallel=96`). A higher batch size may not fit in memory, and lower batch size gives less parallelism.
-- For tolerance, using a higher tolerance may get better speedups but can risk sample quality degradation. If there is quality degradation with the default tolerance, then use a lower tolerance like `0.001`.
-
-For a 1000-step DDPM on 8 A100 GPUs, you can expect around a 3x speedup from [`StableDiffusionParadigmsPipeline`] compared to the [`StableDiffusionPipeline`]
-by setting `parallel=80` and `tolerance=0.1`.
-
-🤗 Diffusers offers [distributed inference support](../../training/distributed_inference) for generating multiple prompts
-in parallel on multiple GPUs. But [`StableDiffusionParadigmsPipeline`] is designed for speeding up sampling of a single prompt by using multiple GPUs.
-
-<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.
-
-</Tip>
-
-## StableDiffusionParadigmsPipeline
-[[autodoc]] StableDiffusionParadigmsPipeline
-	- __call__
-	- all
-
-## StableDiffusionPipelineOutput
-[[autodoc]] pipelines.stable_diffusion.StableDiffusionPipelineOutput

docs/source/en/api/pipelines/pix2pix_zero.md

@@ -1,289 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# Pix2Pix Zero
-
-[Zero-shot Image-to-Image Translation](https://huggingface.co/papers/2302.03027) is by Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu.
-
-The abstract from the paper is:
-
-*Large-scale text-to-image generative models have shown their remarkable ability to synthesize diverse and high-quality images. However, it is still challenging to directly apply these models for editing real images for two reasons. First, it is hard for users to come up with a perfect text prompt that accurately describes every visual detail in the input image. Second, while existing models can introduce desirable changes in certain regions, they often dramatically alter the input content and introduce unexpected changes in unwanted regions. In this work, we propose pix2pix-zero, an image-to-image translation method that can preserve the content of the original image without manual prompting. We first automatically discover editing directions that reflect desired edits in the text embedding space. To preserve the general content structure after editing, we further propose cross-attention guidance, which aims to retain the cross-attention maps of the input image throughout the diffusion process. In addition, our method does not need additional training for these edits and can directly use the existing pre-trained text-to-image diffusion model. We conduct extensive experiments and show that our method outperforms existing and concurrent works for both real and synthetic image editing.*
-
-You can find additional information about Pix2Pix Zero on the [project page](https://pix2pixzero.github.io/),  [original codebase](https://github.com/pix2pixzero/pix2pix-zero), and try it out in a [demo](https://huggingface.co/spaces/pix2pix-zero-library/pix2pix-zero-demo).
-
-## Tips
-
-* The pipeline can be conditioned on real input images. Check out the code examples below to know more.
-* The pipeline exposes two arguments namely `source_embeds` and `target_embeds`
-that let you control the direction of the semantic edits in the final image to be generated. Let's say,
-you wanted to translate from "cat" to "dog". In this case, the edit direction will be "cat -> dog". To reflect
-this in the pipeline, you simply have to set the embeddings related to the phrases including "cat" to
-`source_embeds` and "dog" to `target_embeds`. Refer to the code example below for more details.
-* When you're using this pipeline from a prompt, specify the _source_ concept in the prompt. Taking
-the above example, a valid input prompt would be: "a high resolution painting of a **cat** in the style of van gogh".
-* If you wanted to reverse the direction in the example above, i.e., "dog -> cat", then it's recommended to:
-    * Swap the `source_embeds` and `target_embeds`.
-    * Change the input prompt to include "dog".
-* To learn more about how the source and target embeddings are generated, refer to the [original paper](https://arxiv.org/abs/2302.03027). Below, we also provide some directions on how to generate the embeddings.
-* Note that the quality of the outputs generated with this pipeline is dependent on how good the `source_embeds` and `target_embeds` are. Please, refer to [this discussion](#generating-source-and-target-embeddings) for some suggestions on the topic.
-
-## Available Pipelines:
-
-| Pipeline | Tasks | Demo
-|---|---|:---:|
-| [StableDiffusionPix2PixZeroPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_pix2pix_zero.py) | *Text-Based Image Editing* | [🤗 Space](https://huggingface.co/spaces/pix2pix-zero-library/pix2pix-zero-demo) |
-
-<!-- TODO: add Colab -->
-
-## Usage example
-
-### Based on an image generated with the input prompt
-
-```python
-import requests
-import torch
-
-from diffusers import DDIMScheduler, StableDiffusionPix2PixZeroPipeline
-
-
-def download(embedding_url, local_filepath):
-    r = requests.get(embedding_url)
-    with open(local_filepath, "wb") as f:
-        f.write(r.content)
-
-
-model_ckpt = "CompVis/stable-diffusion-v1-4"
-pipeline = StableDiffusionPix2PixZeroPipeline.from_pretrained(
-    model_ckpt, conditions_input_image=False, torch_dtype=torch.float16
-)
-pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
-pipeline.to("cuda")
-
-prompt = "a high resolution painting of a cat in the style of van gogh"
-src_embs_url = "https://github.com/pix2pixzero/pix2pix-zero/raw/main/assets/embeddings_sd_1.4/cat.pt"
-target_embs_url = "https://github.com/pix2pixzero/pix2pix-zero/raw/main/assets/embeddings_sd_1.4/dog.pt"
-
-for url in [src_embs_url, target_embs_url]:
-    download(url, url.split("/")[-1])
-
-src_embeds = torch.load(src_embs_url.split("/")[-1])
-target_embeds = torch.load(target_embs_url.split("/")[-1])
-
-image = pipeline(
-    prompt,
-    source_embeds=src_embeds,
-    target_embeds=target_embeds,
-    num_inference_steps=50,
-    cross_attention_guidance_amount=0.15,
-).images[0]
-image
-```
-
-### Based on an input image
-
-When the pipeline is conditioned on an input image, we first obtain an inverted
-noise from it using a `DDIMInverseScheduler` with the help of a generated caption. Then the inverted noise is used to start the generation process.
-
-First, let's load our pipeline:
-
-```py
-import torch
-from transformers import BlipForConditionalGeneration, BlipProcessor
-from diffusers import DDIMScheduler, DDIMInverseScheduler, StableDiffusionPix2PixZeroPipeline
-
-captioner_id = "Salesforce/blip-image-captioning-base"
-processor = BlipProcessor.from_pretrained(captioner_id)
-model = BlipForConditionalGeneration.from_pretrained(captioner_id, torch_dtype=torch.float16, low_cpu_mem_usage=True)
-
-sd_model_ckpt = "CompVis/stable-diffusion-v1-4"
-pipeline = StableDiffusionPix2PixZeroPipeline.from_pretrained(
-    sd_model_ckpt,
-    caption_generator=model,
-    caption_processor=processor,
-    torch_dtype=torch.float16,
-    safety_checker=None,
-)
-pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
-pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
-pipeline.enable_model_cpu_offload()
-```
-
-Then, we load an input image for conditioning and obtain a suitable caption for it:
-
-```py
-from diffusers.utils import load_image
-
-img_url = "https://github.com/pix2pixzero/pix2pix-zero/raw/main/assets/test_images/cats/cat_6.png"
-raw_image = load_image(url).resize((512, 512))
-caption = pipeline.generate_caption(raw_image)
-caption
-```
-
-Then we employ the generated caption and the input image to get the inverted noise:
-
-```py
-generator = torch.manual_seed(0)
-inv_latents = pipeline.invert(caption, image=raw_image, generator=generator).latents
-```
-
-Now, generate the image with edit directions:
-
-```py
-# See the "Generating source and target embeddings" section below to
-# automate the generation of these captions with a pre-trained model like Flan-T5 as explained below.
-source_prompts = ["a cat sitting on the street", "a cat playing in the field", "a face of a cat"]
-target_prompts = ["a dog sitting on the street", "a dog playing in the field", "a face of a dog"]
-
-source_embeds = pipeline.get_embeds(source_prompts, batch_size=2)
-target_embeds = pipeline.get_embeds(target_prompts, batch_size=2)
-
-
-image = pipeline(
-    caption,
-    source_embeds=source_embeds,
-    target_embeds=target_embeds,
-    num_inference_steps=50,
-    cross_attention_guidance_amount=0.15,
-    generator=generator,
-    latents=inv_latents,
-    negative_prompt=caption,
-).images[0]
-image
-```
-
-## Generating source and target embeddings
-
-The authors originally used the [GPT-3 API](https://openai.com/api/) to generate the source and target captions for discovering
-edit directions. However, we can also leverage open source and public models for the same purpose.
-Below, we provide an end-to-end example with the [Flan-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5) model
-for generating captions and [CLIP](https://huggingface.co/docs/transformers/model_doc/clip) for
-computing embeddings on the generated captions.
-
-**1. Load the generation model**:
-
-```py
-import torch
-from transformers import AutoTokenizer, T5ForConditionalGeneration
-
-tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-xl")
-model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-xl", device_map="auto", torch_dtype=torch.float16)
-```
-
-**2. Construct a starting prompt**:
-
-```py
-source_concept = "cat"
-target_concept = "dog"
-
-source_text = f"Provide a caption for images containing a {source_concept}. "
-"The captions should be in English and should be no longer than 150 characters."
-
-target_text = f"Provide a caption for images containing a {target_concept}. "
-"The captions should be in English and should be no longer than 150 characters."
-```
-
-Here, we're interested in the "cat -> dog" direction.
-
-**3. Generate captions**:
-
-We can use a utility like so for this purpose.
-
-```py
-def generate_captions(input_prompt):
-    input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids.to("cuda")
-
-    outputs = model.generate(
-        input_ids, temperature=0.8, num_return_sequences=16, do_sample=True, max_new_tokens=128, top_k=10
-    )
-    return tokenizer.batch_decode(outputs, skip_special_tokens=True)
-```
-
-And then we just call it to generate our captions:
-
-```py
-source_captions = generate_captions(source_text)
-target_captions = generate_captions(target_concept)
-print(source_captions, target_captions, sep='\n')
-```
-
-We encourage you to play around with the different parameters supported by the
-`generate()` method ([documentation](https://huggingface.co/docs/transformers/main/en/main_classes/text_generation#transformers.generation_tf_utils.TFGenerationMixin.generate)) for the generation quality you are looking for.
-
-**4. Load the embedding model**:
-
-Here, we need to use the same text encoder model used by the subsequent Stable Diffusion model.
-
-```py
-from diffusers import StableDiffusionPix2PixZeroPipeline
-
-pipeline = StableDiffusionPix2PixZeroPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16
-)
-pipeline = pipeline.to("cuda")
-tokenizer = pipeline.tokenizer
-text_encoder = pipeline.text_encoder
-```
-
-**5. Compute embeddings**:
-
-```py
-import torch
-
-def embed_captions(sentences, tokenizer, text_encoder, device="cuda"):
-    with torch.no_grad():
-        embeddings = []
-        for sent in sentences:
-            text_inputs = tokenizer(
-                sent,
-                padding="max_length",
-                max_length=tokenizer.model_max_length,
-                truncation=True,
-                return_tensors="pt",
-            )
-            text_input_ids = text_inputs.input_ids
-            prompt_embeds = text_encoder(text_input_ids.to(device), attention_mask=None)[0]
-            embeddings.append(prompt_embeds)
-    return torch.concatenate(embeddings, dim=0).mean(dim=0).unsqueeze(0)
-
-source_embeddings = embed_captions(source_captions, tokenizer, text_encoder)
-target_embeddings = embed_captions(target_captions, tokenizer, text_encoder)
-```
-
-And you're done! [Here](https://colab.research.google.com/drive/1tz2C1EdfZYAPlzXXbTnf-5PRBiR8_R1F?usp=sharing) is a Colab Notebook that you can use to interact with the entire process.
-
-Now, you can use these embeddings directly while calling the pipeline:
-
-```py
-from diffusers import DDIMScheduler
-
-pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
-
-image = pipeline(
-    prompt,
-    source_embeds=source_embeddings,
-    target_embeds=target_embeddings,
-    num_inference_steps=50,
-    cross_attention_guidance_amount=0.15,
-).images[0]
-image
-```
-
-<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.
-
-</Tip>
-
-## StableDiffusionPix2PixZeroPipeline
-[[autodoc]] StableDiffusionPix2PixZeroPipeline
-	- __call__
-	- all

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

@@ -1,35 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# PNDM
-
-[Pseudo Numerical Methods for Diffusion Models on Manifolds](https://huggingface.co/papers/2202.09778) (PNDM) is by Luping Liu, Yi Ren, Zhijie Lin and Zhou Zhao.
-
-The abstract from the paper is:
-
-*Denoising Diffusion Probabilistic Models (DDPMs) can generate high-quality samples such as image and audio samples. However, DDPMs require hundreds to thousands of iterations to produce final samples. Several prior works have successfully accelerated DDPMs through adjusting the variance schedule (e.g., Improved Denoising Diffusion Probabilistic Models) or the denoising equation (e.g., Denoising Diffusion Implicit Models (DDIMs)). However, these acceleration methods cannot maintain the quality of samples and even introduce new noise at a high speedup rate, which limit their practicability. To accelerate the inference process while keeping the sample quality, we provide a fresh perspective that DDPMs should be treated as solving differential equations on manifolds. Under such a perspective, we propose pseudo numerical methods for diffusion models (PNDMs). Specifically, we figure out how to solve differential equations on manifolds and show that DDIMs are simple cases of pseudo numerical methods. We change several classical numerical methods to corresponding pseudo numerical methods and find that the pseudo linear multi-step method is the best in most situations. According to our experiments, by directly using pre-trained models on Cifar10, CelebA and LSUN, PNDMs can generate higher quality synthetic images with only 50 steps compared with 1000-step DDIMs (20x speedup), significantly outperform DDIMs with 250 steps (by around 0.4 in FID) and have good generalization on different variance schedules.*
-
-The original codebase can be found at [luping-liu/PNDM](https://github.com/luping-liu/PNDM).
-
-<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.
-
-</Tip>
-
-## PNDMPipeline
-[[autodoc]] PNDMPipeline
-	- all
-	- __call__
-
-## ImagePipelineOutput
-[[autodoc]] pipelines.ImagePipelineOutput

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

@@ -1,37 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# RePaint
-
-[RePaint: Inpainting using Denoising Diffusion Probabilistic Models](https://huggingface.co/papers/2201.09865) is by Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, Luc Van Gool.
-
-The abstract from the paper is:
-
-*Free-form inpainting is the task of adding new content to an image in the regions specified by an arbitrary binary mask. Most existing approaches train for a certain distribution of masks, which limits their generalization capabilities to unseen mask types. Furthermore, training with pixel-wise and perceptual losses often leads to simple textural extensions towards the missing areas instead of semantically meaningful generation. In this work, we propose RePaint: A Denoising Diffusion Probabilistic Model (DDPM) based inpainting approach that is applicable to even extreme masks. We employ a pretrained unconditional DDPM as the generative prior. To condition the generation process, we only alter the reverse diffusion iterations by sampling the unmasked regions using the given image information. Since this technique does not modify or condition the original DDPM network itself, the model produces high-quality and diverse output images for any inpainting form. We validate our method for both faces and general-purpose image inpainting using standard and extreme masks.
-RePaint outperforms state-of-the-art Autoregressive, and GAN approaches for at least five out of six mask distributions.*
-
-The original codebase can be found at [andreas128/RePaint](https://github.com/andreas128/RePaint).
-
-<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.
-
-</Tip>
-
-
-## RePaintPipeline
-[[autodoc]] RePaintPipeline
-	- all
-	- __call__
-
-## ImagePipelineOutput
-[[autodoc]] pipelines.ImagePipelineOutput

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

@@ -1,35 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# Score SDE VE
-
-[Score-Based Generative Modeling through Stochastic Differential Equations](https://huggingface.co/papers/2011.13456) (Score SDE) is by Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon and Ben Poole. This pipeline implements the variance expanding (VE) variant of the stochastic differential equation method.
-
-The abstract from the paper is:
-
-*Creating noise from data is easy; creating data from noise is generative modeling. We present a stochastic differential equation (SDE) that smoothly transforms a complex data distribution to a known prior distribution by slowly injecting noise, and a corresponding reverse-time SDE that transforms the prior distribution back into the data distribution by slowly removing the noise. Crucially, the reverse-time SDE depends only on the time-dependent gradient field (\aka, score) of the perturbed data distribution. By leveraging advances in score-based generative modeling, we can accurately estimate these scores with neural networks, and use numerical SDE solvers to generate samples. We show that this framework encapsulates previous approaches in score-based generative modeling and diffusion probabilistic modeling, allowing for new sampling procedures and new modeling capabilities. In particular, we introduce a predictor-corrector framework to correct errors in the evolution of the discretized reverse-time SDE. We also derive an equivalent neural ODE that samples from the same distribution as the SDE, but additionally enables exact likelihood computation, and improved sampling efficiency. In addition, we provide a new way to solve inverse problems with score-based models, as demonstrated with experiments on class-conditional generation, image inpainting, and colorization. Combined with multiple architectural improvements, we achieve record-breaking performance for unconditional image generation on CIFAR-10 with an Inception score of 9.89 and FID of 2.20, a competitive likelihood of 2.99 bits/dim, and demonstrate high fidelity generation of 1024 x 1024 images for the first time from a score-based generative model.*
-
-The original codebase can be found at [yang-song/score_sde_pytorch](https://github.com/yang-song/score_sde_pytorch).
-
-<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.
-
-</Tip>
-
-## ScoreSdeVePipeline
-[[autodoc]] ScoreSdeVePipeline
-	- all
-	- __call__
-
-## ImagePipelineOutput
-[[autodoc]] pipelines.ImagePipelineOutput

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

@@ -1,37 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# Spectrogram Diffusion
-
-[Spectrogram Diffusion](https://huggingface.co/papers/2206.05408) is by Curtis Hawthorne, Ian Simon, Adam Roberts, Neil Zeghidour, Josh Gardner, Ethan Manilow, and Jesse Engel.
-
-*An ideal music synthesizer should be both interactive and expressive, generating high-fidelity audio in realtime for arbitrary combinations of instruments and notes. Recent neural synthesizers have exhibited a tradeoff between domain-specific models that offer detailed control of only specific instruments, or raw waveform models that can train on any music but with minimal control and slow generation. In this work, we focus on a middle ground of neural synthesizers that can generate audio from MIDI sequences with arbitrary combinations of instruments in realtime. This enables training on a wide range of transcription datasets with a single model, which in turn offers note-level control of composition and instrumentation across a wide range of instruments. We use a simple two-stage process: MIDI to spectrograms with an encoder-decoder Transformer, then spectrograms to audio with a generative adversarial network (GAN) spectrogram inverter. We compare training the decoder as an autoregressive model and as a Denoising Diffusion Probabilistic Model (DDPM) and find that the DDPM approach is superior both qualitatively and as measured by audio reconstruction and Fréchet distance metrics. Given the interactivity and generality of this approach, we find this to be a promising first step towards interactive and expressive neural synthesis for arbitrary combinations of instruments and notes.*
-
-The original codebase can be found at [magenta/music-spectrogram-diffusion](https://github.com/magenta/music-spectrogram-diffusion).
-
-![img](https://storage.googleapis.com/music-synthesis-with-spectrogram-diffusion/architecture.png)
-
-As depicted above the model takes as input a MIDI file and tokenizes it into a sequence of 5 second intervals. Each tokenized interval then together with positional encodings is passed through the Note Encoder and its representation is concatenated with the previous window's generated spectrogram representation obtained via the Context Encoder. For the initial 5 second window this is set to zero. The resulting context is then used as conditioning to sample the denoised Spectrogram from the MIDI window and we concatenate this spectrogram to the final output as well as use it for the context of the next MIDI window. The process repeats till we have gone over all the MIDI inputs. Finally a MelGAN decoder converts the potentially long spectrogram to audio which is the final result of this pipeline.
-
-<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.
-
-</Tip>
-
-## SpectrogramDiffusionPipeline
-[[autodoc]] SpectrogramDiffusionPipeline
-	- all
-	- __call__
-
-## AudioPipelineOutput
-[[autodoc]] pipelines.AudioPipelineOutput

docs/source/en/api/pipelines/stable_diffusion/sdxl_turbo.md

@@ -20,7 +20,7 @@ The abstract from the paper is:
 
 ## Tips
 
-- SDXL Turbo uses the exact same architecture as [SDXL](./stable_diffusion_xl).
+- SDXL Turbo uses the exact same architecture as [SDXL](./stable_diffusion_xl), which means it also has the same API. Please refer to the [SDXL](./stable_diffusion_xl) API reference for more details.
 - SDXL Turbo should disable guidance scale by setting `guidance_scale=0.0`
 - SDXL Turbo should use `timestep_spacing='trailing'` for the scheduler and use between 1 and 4 steps.
 - SDXL Turbo has been trained to generate images of size 512x512.
@@ -28,26 +28,8 @@ The abstract from the paper is:
 
 <Tip>
 
-To learn how to use SDXL Turbo for various tasks, how to optimize performance, and other usage examples, take a look at the [Stable Diffusion XL](../../../using-diffusers/sdxl_turbo) guide.
+To learn how to use SDXL Turbo for various tasks, how to optimize performance, and other usage examples, take a look at the [SDXL Turbo](../../../using-diffusers/sdxl_turbo) guide.
 
 Check out the [Stability AI](https://huggingface.co/stabilityai) Hub organization for the official base and refiner model checkpoints!
 
 </Tip>
-
-## StableDiffusionXLPipeline
-
-[[autodoc]] StableDiffusionXLPipeline
-	- all
-	- __call__
-
-## StableDiffusionXLImg2ImgPipeline
-
-[[autodoc]] StableDiffusionXLImg2ImgPipeline
-	- all
-	- __call__
-
-## StableDiffusionXLInpaintPipeline
-
-[[autodoc]] StableDiffusionXLInpaintPipeline
-	- all
-	- __call__

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

@@ -1,33 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# Stochastic Karras VE
-
-[Elucidating the Design Space of Diffusion-Based Generative Models](https://huggingface.co/papers/2206.00364) is by Tero Karras, Miika Aittala, Timo Aila and Samuli Laine. This pipeline implements the stochastic sampling tailored to variance expanding (VE) models.
-
-The abstract from the paper:
-
-*We argue that the theory and practice of diffusion-based generative models are currently unnecessarily convoluted and seek to remedy the situation by presenting a design space that clearly separates the concrete design choices. This lets us identify several changes to both the sampling and training processes, as well as preconditioning of the score networks. Together, our improvements yield new state-of-the-art FID of 1.79 for CIFAR-10 in a class-conditional setting and 1.97 in an unconditional setting, with much faster sampling (35 network evaluations per image) than prior designs. To further demonstrate their modular nature, we show that our design changes dramatically improve both the efficiency and quality obtainable with pre-trained score networks from previous work, including improving the FID of a previously trained ImageNet-64 model from 2.07 to near-SOTA 1.55, and after re-training with our proposed improvements to a new SOTA of 1.36.*
-
-<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.
-
-</Tip>
-
-## KarrasVePipeline
-[[autodoc]] KarrasVePipeline
-	- all
-	- __call__
-
-## ImagePipelineOutput
-[[autodoc]] pipelines.ImagePipelineOutput

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

@@ -1,54 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# Versatile Diffusion
-
-Versatile Diffusion was proposed in [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://huggingface.co/papers/2211.08332) by Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, Humphrey Shi.
-
-The abstract from the paper is:
-
-*Recent advances in diffusion models have set an impressive milestone in many generation tasks, and trending works such as DALL-E2, Imagen, and Stable Diffusion have attracted great interest. Despite the rapid landscape changes, recent new approaches focus on extensions and performance rather than capacity, thus requiring separate models for separate tasks. In this work, we expand the existing single-flow diffusion pipeline into a multi-task multimodal network, dubbed Versatile Diffusion (VD), that handles multiple flows of text-to-image, image-to-text, and variations in one unified model. The pipeline design of VD instantiates a unified multi-flow diffusion framework, consisting of sharable and swappable layer modules that enable the crossmodal generality beyond images and text. Through extensive experiments, we demonstrate that VD successfully achieves the following: a) VD outperforms the baseline approaches and handles all its base tasks with competitive quality; b) VD enables novel extensions such as disentanglement of style and semantics, dual- and multi-context blending, etc.; c) The success of our multi-flow multimodal framework over images and text may inspire further diffusion-based universal AI research.*
-
-## Tips
-
-You can load the more memory intensive "all-in-one" [`VersatileDiffusionPipeline`] that supports all the tasks or use the individual pipelines which are more memory efficient.
-
-| **Pipeline**                                         | **Supported tasks**               |
-|------------------------------------------------------|-----------------------------------|
-| [`VersatileDiffusionPipeline`]                       | all of the below                  |
-| [`VersatileDiffusionTextToImagePipeline`]            | text-to-image                     |
-| [`VersatileDiffusionImageVariationPipeline`]         | image variation                   |
-| [`VersatileDiffusionDualGuidedPipeline`]             | image-text dual guided generation |
-
-<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.
-
-</Tip>
-
-## VersatileDiffusionPipeline
-[[autodoc]] VersatileDiffusionPipeline
-
-## VersatileDiffusionTextToImagePipeline
-[[autodoc]] VersatileDiffusionTextToImagePipeline
-	- all
-	- __call__
-
-## VersatileDiffusionImageVariationPipeline
-[[autodoc]] VersatileDiffusionImageVariationPipeline
-	- all
-	- __call__
-
-## VersatileDiffusionDualGuidedPipeline
-[[autodoc]] VersatileDiffusionDualGuidedPipeline
-	- all
-	- __call__

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

@@ -1,35 +0,0 @@
-<!--Copyright 2023 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.
--->
-
-# VQ Diffusion
-
-[Vector Quantized Diffusion Model for Text-to-Image Synthesis](https://huggingface.co/papers/2111.14822) is by Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, Baining Guo.
-
-The abstract from the paper is:
-
-*We present the vector quantized diffusion (VQ-Diffusion) model for text-to-image generation. This method is based on a vector quantized variational autoencoder (VQ-VAE) whose latent space is modeled by a conditional variant of the recently developed Denoising Diffusion Probabilistic Model (DDPM). We find that this latent-space method is well-suited for text-to-image generation tasks because it not only eliminates the unidirectional bias with existing methods but also allows us to incorporate a mask-and-replace diffusion strategy to avoid the accumulation of errors, which is a serious problem with existing methods. Our experiments show that the VQ-Diffusion produces significantly better text-to-image generation results when compared with conventional autoregressive (AR) models with similar numbers of parameters. Compared with previous GAN-based text-to-image methods, our VQ-Diffusion can handle more complex scenes and improve the synthesized image quality by a large margin. Finally, we show that the image generation computation in our method can be made highly efficient by reparameterization. With traditional AR methods, the text-to-image generation time increases linearly with the output image resolution and hence is quite time consuming even for normal size images. The VQ-Diffusion allows us to achieve a better trade-off between quality and speed. Our experiments indicate that the VQ-Diffusion model with the reparameterization is fifteen times faster than traditional AR methods while achieving a better image quality.*
-
-The original codebase can be found at [microsoft/VQ-Diffusion](https://github.com/microsoft/VQ-Diffusion).
-
-<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.
-
-</Tip>
-
-## VQDiffusionPipeline
-[[autodoc]] VQDiffusionPipeline
-	- all
-	- __call__
-
-## ImagePipelineOutput
-[[autodoc]] pipelines.ImagePipelineOutput

docs/source/en/training/lora.md

@@ -179,7 +179,7 @@ accelerate launch --mixed_precision="fp16"  train_text_to_image_lora.py \
   --pretrained_model_name_or_path=$MODEL_NAME \
   --dataset_name=$DATASET_NAME \
   --dataloader_num_workers=8 \
-  --resolution=512 
+  --resolution=512 \
   --center_crop \
   --random_flip \
   --train_batch_size=1 \
@@ -214,4 +214,4 @@ image = pipeline("A pokemon with blue eyes").images[0]
 Congratulations on training a new model with LoRA! To learn more about how to use your new model, the following guides may be helpful:
 
 - Learn how to [load different LoRA formats](../using-diffusers/loading_adapters#LoRA) trained using community trainers like Kohya and TheLastBen.
-- Learn how to use and [combine multiple LoRA's](../tutorials/using_peft_for_inference) with PEFT for inference.
+- Learn how to use and [combine multiple LoRA's](../tutorials/using_peft_for_inference) with PEFT for inference.

docs/source/en/training/t2i_adapters.md

@@ -224,4 +224,4 @@ image.save("./output.png")
 
 Congratulations on training a T2I-Adapter model! 🎉 To learn more:
 
-- Read the [Efficient Controllable Generation for SDXL with T2I-Adapters](https://www.cs.cmu.edu/~custom-diffusion/) blog post to learn more details about the experimental results from the T2I-Adapter team.
+- Read the [Efficient Controllable Generation for SDXL with T2I-Adapters](https://huggingface.co/blog/t2i-sdxl-adapters) blog post to learn more details about the experimental results from the T2I-Adapter team.

docs/source/en/training/unconditional_training.md

@@ -186,7 +186,7 @@ accelerate launch train_unconditional.py \
 If you're training with more than one GPU, add the `--multi_gpu` parameter to the training command:
 
 ```bash
-accelerate launch --mixed_precision="fp16" --multi_gpu train_unconditional.py \
+accelerate launch --multi_gpu train_unconditional.py \
   --dataset_name="huggan/flowers-102-categories" \
   --output_dir="ddpm-ema-flowers-64" \
   --mixed_precision="fp16" \

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

@@ -203,7 +203,7 @@ def make_inpaint_condition(image, image_mask):
     image_mask = np.array(image_mask.convert("L")).astype(np.float32) / 255.0
 
     assert image.shape[0:1] == image_mask.shape[0:1]
-    image[image_mask > 0.5] = 1.0  # set as masked pixel
+    image[image_mask > 0.5] = -1.0  # set as masked pixel
     image = np.expand_dims(image, 0).transpose(0, 3, 1, 2)
     image = torch.from_numpy(image)
     return image

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

@@ -485,6 +485,69 @@ image.save("sdxl_t2i.png")
   </div>
 </div>
 
+You can use the IP-Adapter face model to apply specific faces to your images.  It is an effective way to maintain consistent characters in your image generations.
+Weights are loaded with the same method used for the other IP-Adapters.  
+
+```python
+# Load ip-adapter-full-face_sd15.bin
+pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter-full-face_sd15.bin")
+```
+
+<Tip>
+
+It is recommended to use `DDIMScheduler` and `EulerDiscreteScheduler` for face model. 
+
+
+</Tip>
+
+```python
+import torch
+from diffusers import StableDiffusionPipeline, DDIMScheduler
+from diffusers.utils import load_image
+
+noise_scheduler = DDIMScheduler(
+    num_train_timesteps=1000,
+    beta_start=0.00085,
+    beta_end=0.012,
+    beta_schedule="scaled_linear",
+    clip_sample=False,
+    set_alpha_to_one=False,
+    steps_offset=1
+)
+
+pipeline = StableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    scheduler=noise_scheduler,
+).to("cuda")
+
+pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter-full-face_sd15.bin")
+
+pipeline.set_ip_adapter_scale(0.7)
+
+image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ai_face2.png")
+
+generator = torch.Generator(device="cpu").manual_seed(33)
+
+image = pipeline(
+    prompt="A photo of a girl wearing a black dress, holding red roses in hand, upper body, behind is the Eiffel Tower",
+    ip_adapter_image=image,
+    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
+    num_inference_steps=50, num_images_per_prompt=1, width=512, height=704,
+    generator=generator,
+).images[0]
+```
+
+<div class="flex flex-row gap-4">
+  <div class="flex-1">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ai_face2.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">input image</figcaption>
+  </div>
+  <div class="flex-1">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/ipadapter_full_face_output.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">output image</figcaption>
+  </div>
+</div>
 
 ### LCM-Lora
 

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

@@ -41,6 +41,20 @@ Now, define four different `Generator`s and assign each `Generator` a seed (`0`
 generator = [torch.Generator(device="cuda").manual_seed(i) for i in range(4)]
 ```
 
+<Tip warning={true}>
+
+To create a batched seed, you should use a list comprehension that iterates over the length specified in `range()`. This creates a unique `Generator` object for each image in the batch. If you only multiply the `Generator` by the batch size, this only creates one `Generator` object that is used sequentially for each image in the batch.
+
+For example, if you want to use the same seed to create 4 identical images:
+
+```py
+❌ [torch.Generator().manual_seed(seed)] * 4
+
+✅ [torch.Generator().manual_seed(seed) for _ in range(4)]
+```
+
+</Tip>
+
 Generate the images and have a look:
 
 ```python

examples/README.md

@@ -18,8 +18,7 @@ limitations under the License.
 Diffusers examples are a collection of scripts to demonstrate how to effectively use the `diffusers` library
 for a variety of use cases involving training or fine-tuning.
 
-**Note**: If you are looking for **official** examples on how to use `diffusers` for inference, 
-please have a look at [src/diffusers/pipelines](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines).
+**Note**: If you are looking for **official** examples on how to use `diffusers` for inference, please have a look at [src/diffusers/pipelines](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines).
 
 Our examples aspire to be **self-contained**, **easy-to-tweak**, **beginner-friendly** and for **one-purpose-only**.
 More specifically, this means:
@@ -27,11 +26,10 @@ More specifically, this means:
 - **Self-contained**: An example script shall only depend on "pip-install-able" Python packages that can be found in a `requirements.txt` file. Example scripts shall **not** depend on any local files. This means that one can simply download an example script, *e.g.* [train_unconditional.py](https://github.com/huggingface/diffusers/blob/main/examples/unconditional_image_generation/train_unconditional.py), install the required dependencies, *e.g.* [requirements.txt](https://github.com/huggingface/diffusers/blob/main/examples/unconditional_image_generation/requirements.txt) and execute the example script.
 - **Easy-to-tweak**: While we strive to present as many use cases as possible, the example scripts are just that - examples. It is expected that they won't work out-of-the box on your specific problem and that you will be required to change a few lines of code to adapt them to your needs. To help you with that, most of the examples fully expose the preprocessing of the data and the training loop to allow you to tweak and edit them as required.
 - **Beginner-friendly**: We do not aim for providing state-of-the-art training scripts for the newest models, but rather examples that can be used as a way to better understand diffusion models and how to use them with the `diffusers` library. We often purposefully leave out certain state-of-the-art methods if we consider them too complex for beginners.
-- **One-purpose-only**: Examples should show one task and one task only. Even if a task is from a modeling 
-point of view very similar, *e.g.* image super-resolution and image modification tend to use the same model and training method, we want examples to showcase only one task to keep them as readable and easy-to-understand as possible.
+- **One-purpose-only**: Examples should show one task and one task only. Even if a task is from a modeling point of view very similar, *e.g.* image super-resolution and image modification tend to use the same model and training method, we want examples to showcase only one task to keep them as readable and easy-to-understand as possible.
 
 We provide **official** examples that cover the most popular tasks of diffusion models.
-*Official* examples are **actively** maintained by the `diffusers` maintainers and we try to rigorously follow our example philosophy as defined above. 
+*Official* examples are **actively** maintained by the `diffusers` maintainers and we try to rigorously follow our example philosophy as defined above.
 If you feel like another important example should exist, we are more than happy to welcome a [Feature Request](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=&template=feature_request.md&title=) or directly a [Pull Request](https://github.com/huggingface/diffusers/compare) from you!
 
 Training examples show how to pretrain or fine-tune diffusion models for a variety of tasks. Currently we support:
@@ -39,7 +37,7 @@ Training examples show how to pretrain or fine-tune diffusion models for a varie
 | Task | 🤗 Accelerate | 🤗 Datasets | Colab
 |---|---|:---:|:---:|
 | [**Unconditional Image Generation**](./unconditional_image_generation) | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
-| [**Text-to-Image fine-tuning**](./text_to_image) | ✅ | ✅ | 
+| [**Text-to-Image fine-tuning**](./text_to_image) | ✅ | ✅ |
 | [**Textual Inversion**](./textual_inversion) | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb)
 | [**Dreambooth**](./dreambooth) | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_dreambooth_training.ipynb)
 | [**ControlNet**](./controlnet) | ✅ | ✅ | -

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -112,7 +112,7 @@ def save_model_card(
     repo_folder=None,
     vae_path=None,
 ):
-    img_str = "widget:\n" if images else ""
+    img_str = "widget:\n"
     for i, image in enumerate(images):
         image.save(os.path.join(repo_folder, f"image_{i}.png"))
         img_str += f"""
@@ -121,6 +121,10 @@ def save_model_card(
             url:
                 "image_{i}.png"
         """
+    if not images:
+        img_str += f"""
+        - text: '{instance_prompt}'
+        """
 
     trigger_str = f"You should use {instance_prompt} to trigger the image generation."
     diffusers_imports_pivotal = ""
@@ -133,10 +137,10 @@ def save_model_card(
         diffusers_imports_pivotal = """from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file
         """
-        diffusers_example_pivotal = f"""embedding_path = hf_hub_download(repo_id="{repo_id}", filename="embeddings.safetensors", repo_type="model")
+        diffusers_example_pivotal = f"""embedding_path = hf_hub_download(repo_id='{repo_id}', filename="embeddings.safetensors", repo_type="model")
 state_dict = load_file(embedding_path)
-pipeline.load_textual_inversion(state_dict["clip_l"], token=["<s0>", "<s1>"], text_encoder=pipe.text_encoder, tokenizer=pipe.tokenizer)
-pipeline.load_textual_inversion(state_dict["clip_g"], token=["<s0>", "<s1>"], text_encoder=pipe.text_encoder_2, tokenizer=pipe.tokenizer_2)
+pipeline.load_textual_inversion(state_dict["clip_l"], token=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
+pipeline.load_textual_inversion(state_dict["clip_g"], token=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
         """
         if token_abstraction_dict:
             for key, value in token_abstraction_dict.items():
@@ -145,8 +149,7 @@ pipeline.load_textual_inversion(state_dict["clip_g"], token=["<s0>", "<s1>"], te
 to trigger concept `{key}` → use `{tokens}` in your prompt \n
 """
 
-    yaml = f"""
----
+    yaml = f"""---
 tags:
 - stable-diffusion-xl
 - stable-diffusion-xl-diffusers
@@ -159,7 +162,7 @@ base_model: {base_model}
 instance_prompt: {instance_prompt}
 license: openrail++
 ---
-    """
+"""
 
     model_card = f"""
 # SDXL LoRA DreamBooth - {repo_id}
@@ -170,14 +173,6 @@ license: openrail++
 
 ### These are {repo_id} LoRA adaption weights for {base_model}.
 
-The weights were trained  using [DreamBooth](https://dreambooth.github.io/).
-
-LoRA for the text encoder was enabled: {train_text_encoder}.
-
-Pivotal tuning was enabled: {train_text_encoder_ti}.
-
-Special VAE used for training: {vae_path}.
-
 ## Trigger words
 
 {trigger_str}
@@ -196,11 +191,24 @@ image = pipeline('{validation_prompt if validation_prompt else instance_prompt}'
 
 For more details, including weighting, merging and fusing LoRAs, check the [documentation on loading LoRAs in diffusers](https://huggingface.co/docs/diffusers/main/en/using-diffusers/loading_adapters)
 
-## Download model (use it with UIs such as AUTO1111, Comfy, SD.Next, Invoke)
+## Download model
 
-Weights for this model are available in Safetensors format.
+### Use it with UIs such as AUTOMATIC1111, Comfy UI, SD.Next, Invoke
 
-[Download]({repo_id}/tree/main) them in the Files & versions tab.
+- Download the LoRA *.safetensors [here](/{repo_id}/blob/main/pytorch_lora_weights.safetensors). Rename it and place it on your Lora folder.
+- Download the text embeddings *.safetensors [here](/{repo_id}/blob/main/embeddings.safetensors). Rename it and place it on it on your embeddings folder.
+
+All [Files & versions](/{repo_id}/tree/main).
+
+## Details
+
+The weights were trained using [🧨 diffusers Advanced Dreambooth Training Script](https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py).
+
+LoRA for the text encoder was enabled. {train_text_encoder}.
+
+Pivotal tuning was enabled: {train_text_encoder_ti}.
+
+Special VAE used for training: {vae_path}.
 
 """
     with open(os.path.join(repo_folder, "README.md"), "w") as f:
@@ -667,6 +675,12 @@ def parse_args(input_args=None):
         default=4,
         help=("The dimension of the LoRA update matrices."),
     )
+    parser.add_argument(
+        "--cache_latents",
+        action="store_true",
+        default=False,
+        help="Cache the VAE latents",
+    )
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -1170,6 +1184,7 @@ def main(args):
         revision=args.revision,
         variant=args.variant,
     )
+    vae_scaling_factor = vae.config.scaling_factor
     unet = UNet2DConditionModel.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
     )
@@ -1600,6 +1615,20 @@ def main(args):
             args.validation_prompt = args.validation_prompt.replace(token_abs, "".join(token_replacement))
     print("validation prompt:", args.validation_prompt)
 
+    if args.cache_latents:
+        latents_cache = []
+        for batch in tqdm(train_dataloader, desc="Caching latents"):
+            with torch.no_grad():
+                batch["pixel_values"] = batch["pixel_values"].to(
+                    accelerator.device, non_blocking=True, dtype=torch.float32
+                )
+                latents_cache.append(vae.encode(batch["pixel_values"]).latent_dist)
+
+        if args.validation_prompt is None:
+            del vae
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
     # 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)
@@ -1715,9 +1744,7 @@ def main(args):
         unet.train()
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
-                pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
                 prompts = batch["prompts"]
-                # print(prompts)
                 # encode batch prompts when custom prompts are provided for each image -
                 if train_dataset.custom_instance_prompts:
                     if freeze_text_encoder:
@@ -1729,9 +1756,13 @@ def main(args):
                         tokens_one = tokenize_prompt(tokenizer_one, prompts, add_special_tokens)
                         tokens_two = tokenize_prompt(tokenizer_two, prompts, add_special_tokens)
 
-                # Convert images to latent space
-                model_input = vae.encode(pixel_values).latent_dist.sample()
-                model_input = model_input * vae.config.scaling_factor
+                if args.cache_latents:
+                    model_input = latents_cache[step].sample()
+                else:
+                    pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
+                    model_input = vae.encode(pixel_values).latent_dist.sample()
+
+                model_input = model_input * vae_scaling_factor
                 if args.pretrained_vae_model_name_or_path is None:
                     model_input = model_input.to(weight_dtype)
 
@@ -1981,43 +2012,42 @@ def main(args):
             text_encoder_lora_layers=text_encoder_lora_layers,
             text_encoder_2_lora_layers=text_encoder_2_lora_layers,
         )
-
-        # Final inference
-        # Load previous pipeline
-        vae = AutoencoderKL.from_pretrained(
-            vae_path,
-            subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
-            revision=args.revision,
-            variant=args.variant,
-            torch_dtype=weight_dtype,
-        )
-        pipeline = StableDiffusionXLPipeline.from_pretrained(
-            args.pretrained_model_name_or_path,
-            vae=vae,
-            revision=args.revision,
-            variant=args.variant,
-            torch_dtype=weight_dtype,
-        )
-
-        # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
-        scheduler_args = {}
-
-        if "variance_type" in pipeline.scheduler.config:
-            variance_type = pipeline.scheduler.config.variance_type
-
-            if variance_type in ["learned", "learned_range"]:
-                variance_type = "fixed_small"
-
-            scheduler_args["variance_type"] = variance_type
-
-        pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
-
-        # load attention processors
-        pipeline.load_lora_weights(args.output_dir)
-
-        # run inference
         images = []
         if args.validation_prompt and args.num_validation_images > 0:
+            # Final inference
+            # Load previous pipeline
+            vae = AutoencoderKL.from_pretrained(
+                vae_path,
+                subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
+                revision=args.revision,
+                variant=args.variant,
+                torch_dtype=weight_dtype,
+            )
+            pipeline = StableDiffusionXLPipeline.from_pretrained(
+                args.pretrained_model_name_or_path,
+                vae=vae,
+                revision=args.revision,
+                variant=args.variant,
+                torch_dtype=weight_dtype,
+            )
+
+            # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
+            scheduler_args = {}
+
+            if "variance_type" in pipeline.scheduler.config:
+                variance_type = pipeline.scheduler.config.variance_type
+
+                if variance_type in ["learned", "learned_range"]:
+                    variance_type = "fixed_small"
+
+                scheduler_args["variance_type"] = variance_type
+
+            pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
+
+            # load attention processors
+            pipeline.load_lora_weights(args.output_dir)
+
+            # run inference
             pipeline = pipeline.to(accelerator.device)
             generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
             images = [

examples/community/README.md

@@ -8,13 +8,13 @@ If a community doesn't work as expected, please open an issue and ping the autho
 
 | Example                                                                                                                               | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                                              | Colab                                                                                                                                                                                                              |                                                        Author |
 |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
-| LLM-grounded Diffusion (LMD+)                                                                                                         | LMD greatly improves the prompt following ability of text-to-image generation models by introducing an LLM as a front-end prompt parser and layout planner. [Project page.](https://llm-grounded-diffusion.github.io/) [See our full codebase (also with diffusers).](https://github.com/TonyLianLong/LLM-groundedDiffusion)                                                                                                                                                                                                                                                                                                                                                                                                                                   | [LLM-grounded Diffusion (LMD+)](#llm-grounded-diffusion)                             | [Huggingface Demo](https://huggingface.co/spaces/longlian/llm-grounded-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1SXzMSeAB-LJYISb2yrUOdypLz4OYWUKj) |                [Long (Tony) Lian](https://tonylian.com/) | 
-| CLIP Guided Stable Diffusion                                                                                                          | Doing CLIP guidance for text to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                   | [CLIP Guided Stable Diffusion](#clip-guided-stable-diffusion)                             | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb) |                [Suraj Patil](https://github.com/patil-suraj/) | 
+| LLM-grounded Diffusion (LMD+)                                                                                                         | LMD greatly improves the prompt following ability of text-to-image generation models by introducing an LLM as a front-end prompt parser and layout planner. [Project page.](https://llm-grounded-diffusion.github.io/) [See our full codebase (also with diffusers).](https://github.com/TonyLianLong/LLM-groundedDiffusion)                                                                                                                                                                                                                                                                                                                                                                                                                                   | [LLM-grounded Diffusion (LMD+)](#llm-grounded-diffusion)                             | [Huggingface Demo](https://huggingface.co/spaces/longlian/llm-grounded-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1SXzMSeAB-LJYISb2yrUOdypLz4OYWUKj) |                [Long (Tony) Lian](https://tonylian.com/) |
+| CLIP Guided Stable Diffusion                                                                                                          | Doing CLIP guidance for text to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                   | [CLIP Guided Stable Diffusion](#clip-guided-stable-diffusion)                             | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb) |                [Suraj Patil](https://github.com/patil-suraj/) |
 | One Step U-Net (Dummy)                                                                                                                | Example showcasing of how to use Community Pipelines (see https://github.com/huggingface/diffusers/issues/841)                                                                                                                                                                                                                                                                                                                                                                                           | [One Step U-Net](#one-step-unet)                                                          | -                                                                                                                                                                                                                  |    [Patrick von Platen](https://github.com/patrickvonplaten/) |
 | Stable Diffusion Interpolation                                                                                                        | Interpolate the latent space of Stable Diffusion between different prompts/seeds                                                                                                                                                                                                                                                                                                                                                                                                                         | [Stable Diffusion Interpolation](#stable-diffusion-interpolation)                         | -                                                                                                                                                                                                                  |                       [Nate Raw](https://github.com/nateraw/) |
 | Stable Diffusion Mega                                                                                                                 | **One** Stable Diffusion Pipeline with all functionalities of [Text2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py), [Image2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) and [Inpainting](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | [Stable Diffusion Mega](#stable-diffusion-mega)                                           | -                                                                                                                                                                                                                  |    [Patrick von Platen](https://github.com/patrickvonplaten/) |
 | Long Prompt Weighting Stable Diffusion                                                                                                | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt.                                                                                                                                                                                                                                                                                                                                                                                                  | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion)         | -                                                                                                                                                                                                                  |                           [SkyTNT](https://github.com/SkyTNT) |
-| Speech to Image                                                                                                                       | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images                                                                                                                                                                                                                                                                                                                                                                                                            | [Speech to Image](#speech-to-image)                                                       | -                                                                                                                                                                                                                  |             [Mikail Duzenli](https://github.com/MikailINTech) 
+| Speech to Image                                                                                                                       | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images                                                                                                                                                                                                                                                                                                                                                                                                            | [Speech to Image](#speech-to-image)                                                       | -                                                                                                                                                                                                                  |             [Mikail Duzenli](https://github.com/MikailINTech)
 | Wild Card Stable Diffusion                                                                                                            | Stable Diffusion Pipeline that supports prompts that contain wildcard terms (indicated by surrounding double underscores), with values instantiated randomly from a corresponding txt file or a dictionary of possible values                                                                                                                                                                                                                                                                            | [Wildcard Stable Diffusion](#wildcard-stable-diffusion)                                   | -                                                                                                                                                                                                                  |              [Shyam Sudhakaran](https://github.com/shyamsn97) |
 | [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "|" in prompts (as an AND condition) and weights (separated by "|" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                            | [Composable Stable Diffusion](#composable-stable-diffusion)                               | -                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
 | Seed Resizing Stable Diffusion                                                                                                        | Stable Diffusion Pipeline that supports resizing an image and retaining the concepts of the 512 by 512 generation.                                                                                                                                                                                                                                                                                                                                                                                       | [Seed Resizing](#seed-resizing)                                                           | -                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
@@ -24,32 +24,34 @@ If a community doesn't work as expected, please open an issue and ping the autho
 | Text Based Inpainting Stable Diffusion                                                                                                | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting                                                                                                                                                                                                                                                                                                                                                                                              | [Text Based Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion)     | -                                                                                                                                                                                                                  |                   [Dhruv Karan](https://github.com/unography) |
 | Bit Diffusion                                                                                                                         | Diffusion on discrete data                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | [Bit Diffusion](#bit-diffusion)                                                           | -  |                       [Stuti R.](https://github.com/kingstut) |
 | K-Diffusion Stable Diffusion                                                                                                          | Run Stable Diffusion with any of [K-Diffusion's samplers](https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py)                                                                                                                                                                                                                                                                                                                                                                  | [Stable Diffusion with K Diffusion](#stable-diffusion-with-k-diffusion)                   | -  |    [Patrick von Platen](https://github.com/patrickvonplaten/) |
-| Checkpoint Merger Pipeline                                                                                                            | Diffusion Pipeline that enables merging of saved model checkpoints                                                                                                                                                                                                                                                                                                                                                                                                                                       | [Checkpoint Merger Pipeline](#checkpoint-merger-pipeline)                                 | -                                                                                                                                                                                                                  | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) | 
+| Checkpoint Merger Pipeline                                                                                                            | Diffusion Pipeline that enables merging of saved model checkpoints                                                                                                                                                                                                                                                                                                                                                                                                                                       | [Checkpoint Merger Pipeline](#checkpoint-merger-pipeline)                                 | -                                                                                                                                                                                                                  | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
  Stable Diffusion v1.1-1.4 Comparison                                                                                                  | Run all 4 model checkpoints for Stable Diffusion and compare their results together                                                                                                                                                                                                                                                                                                                                                                                                                      | [Stable Diffusion Comparison](#stable-diffusion-comparisons)                              | - |        [Suvaditya Mukherjee](https://github.com/suvadityamuk) |
  MagicMix                                                                                                                              | Diffusion Pipeline for semantic mixing of an image and a text prompt                                                                                                                                                                                                                                                                                                                                                                                                                                     | [MagicMix](#magic-mix)                                                                    | - |                    [Partho Das](https://github.com/daspartho) |
 | Stable UnCLIP                                                                                                                         | Diffusion Pipeline for combining prior model (generate clip image embedding from text, UnCLIPPipeline `"kakaobrain/karlo-v1-alpha"`) and decoder pipeline (decode clip image embedding to image, StableDiffusionImageVariationPipeline `"lambdalabs/sd-image-variations-diffusers"` ).                                                                                                                                                                                                                   | [Stable UnCLIP](#stable-unclip)                                                           | -  |                                [Ray Wang](https://wrong.wang) |
-| UnCLIP Text Interpolation Pipeline                                                                                                    | Diffusion Pipeline that allows passing two prompts and produces images while interpolating between the text-embeddings of the two prompts                                                                                                                                                                                                                                                                                                                                                                | [UnCLIP Text Interpolation Pipeline](#unclip-text-interpolation-pipeline)                 | -                                                                                                                                                                                                                  | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) | 
-| UnCLIP Image Interpolation Pipeline                                                                                                   | Diffusion Pipeline that allows passing two images/image_embeddings and produces images while interpolating between their image-embeddings                                                                                                                                                                                                                                                                                                                                                                | [UnCLIP Image Interpolation Pipeline](#unclip-image-interpolation-pipeline)               | -                                                                                                                                                                                                                  | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) | 
+| UnCLIP Text Interpolation Pipeline                                                                                                    | Diffusion Pipeline that allows passing two prompts and produces images while interpolating between the text-embeddings of the two prompts                                                                                                                                                                                                                                                                                                                                                                | [UnCLIP Text Interpolation Pipeline](#unclip-text-interpolation-pipeline)                 | -                                                                                                                                                                                                                  | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
+| UnCLIP Image Interpolation Pipeline                                                                                                   | Diffusion Pipeline that allows passing two images/image_embeddings and produces images while interpolating between their image-embeddings                                                                                                                                                                                                                                                                                                                                                                | [UnCLIP Image Interpolation Pipeline](#unclip-image-interpolation-pipeline)               | -                                                                                                                                                                                                                  | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
 | DDIM Noise Comparative Analysis Pipeline                                                                                              | Investigating how the diffusion models learn visual concepts from each noise level (which is a contribution of [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227))                                                                                                                                                                                                                                                                                                                             | [DDIM Noise Comparative Analysis Pipeline](#ddim-noise-comparative-analysis-pipeline)     | - |              [Aengus (Duc-Anh)](https://github.com/aengusng8) |
-| CLIP Guided Img2Img Stable Diffusion Pipeline                                                                                         | Doing CLIP guidance for image to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                  | [CLIP Guided Img2Img Stable Diffusion](#clip-guided-img2img-stable-diffusion)             | - |               [Nipun Jindal](https://github.com/nipunjindal/) | 
+| CLIP Guided Img2Img Stable Diffusion Pipeline                                                                                         | Doing CLIP guidance for image to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                  | [CLIP Guided Img2Img Stable Diffusion](#clip-guided-img2img-stable-diffusion)             | - |               [Nipun Jindal](https://github.com/nipunjindal/) |
 | TensorRT Stable Diffusion Text to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
-| EDICT Image Editing Pipeline                                                                                                          | Diffusion pipeline for text-guided image editing                                                                                                                                                                                                                                                                                                                                                                                                                                                         | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline)                             | - |                    [Joqsan Azocar](https://github.com/Joqsan) | 
-| Stable Diffusion RePaint                                                                                                              | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.0986) for inpainting.                                                                                                                                                                                                                                                                                                                                                                                                               | [Stable Diffusion RePaint](#stable-diffusion-repaint )                                    | - |                  [Markus Pobitzer](https://github.com/Markus-Pobitzer) | 
+| EDICT Image Editing Pipeline                                                                                                          | Diffusion pipeline for text-guided image editing                                                                                                                                                                                                                                                                                                                                                                                                                                                         | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline)                             | - |                    [Joqsan Azocar](https://github.com/Joqsan) |
+| Stable Diffusion RePaint                                                                                                              | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.0986) for inpainting.                                                                                                                                                                                                                                                                                                                                                                                                               | [Stable Diffusion RePaint](#stable-diffusion-repaint )                                    | - |                  [Markus Pobitzer](https://github.com/Markus-Pobitzer) |
 | TensorRT Stable Diffusion Image to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Image2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Image to Image Pipeline](#tensorrt-image2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
-| Stable Diffusion IPEX Pipeline | Accelerate Stable Diffusion inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [Stable Diffusion on IPEX](#stable-diffusion-on-ipex) | - | [Yingjie Han](https://github.com/yingjie-han/) | 
-| CLIP Guided Images Mixing Stable Diffusion Pipeline | Сombine images using usual diffusion models. | [CLIP Guided Images Mixing Using Stable Diffusion](#clip-guided-images-mixing-with-stable-diffusion) | - | [Karachev Denis](https://github.com/TheDenk) |  
+| Stable Diffusion IPEX Pipeline | Accelerate Stable Diffusion inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [Stable Diffusion on IPEX](#stable-diffusion-on-ipex) | - | [Yingjie Han](https://github.com/yingjie-han/) |
+| CLIP Guided Images Mixing Stable Diffusion Pipeline | Сombine images using usual diffusion models. | [CLIP Guided Images Mixing Using Stable Diffusion](#clip-guided-images-mixing-with-stable-diffusion) | - | [Karachev Denis](https://github.com/TheDenk) |
 | TensorRT Stable Diffusion Inpainting Pipeline                                                                                                    | Accelerates the Stable Diffusion Inpainting Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Inpainting Pipeline](#tensorrt-inpainting-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
-|   IADB Pipeline                                                                                                    | Implementation of [Iterative α-(de)Blending: a Minimalist Deterministic Diffusion Model](https://arxiv.org/abs/2305.03486)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [IADB Pipeline](#iadb-pipeline)      | - |              [Thomas Chambon](https://github.com/tchambon) 
+|   IADB Pipeline                                                                                                    | Implementation of [Iterative α-(de)Blending: a Minimalist Deterministic Diffusion Model](https://arxiv.org/abs/2305.03486)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [IADB Pipeline](#iadb-pipeline)      | - |              [Thomas Chambon](https://github.com/tchambon)
 |   Zero1to3 Pipeline                                                                                                    | Implementation of [Zero-1-to-3: Zero-shot One Image to 3D Object](https://arxiv.org/abs/2303.11328)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Zero1to3 Pipeline](#Zero1to3-pipeline)      | - |              [Xin Kong](https://github.com/kxhit) |
-Stable Diffusion XL Long Weighted Prompt Pipeline | A pipeline support unlimited length of prompt and negative prompt, use A1111 style of prompt weighting | [Stable Diffusion XL Long Weighted Prompt Pipeline](#stable-diffusion-xl-long-weighted-prompt-pipeline) | - | [Andrew Zhu](https://xhinker.medium.com/) | 
-FABRIC - Stable Diffusion with feedback Pipeline | pipeline supports feedback from liked and disliked images | [Stable Diffusion Fabric Pipeline](#stable-diffusion-fabric-pipeline) | - | [Shauray Singh](https://shauray8.github.io/about_shauray/) | 
-sketch inpaint - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion Pipeline](#stable-diffusion-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) | 
-prompt-to-prompt | change parts of a prompt and retain image structure (see [paper page](https://prompt-to-prompt.github.io/)) | [Prompt2Prompt Pipeline](#prompt2prompt-pipeline) | - | [Umer H. Adil](https://twitter.com/UmerHAdil) | 
+| Stable Diffusion XL Long Weighted Prompt Pipeline | A pipeline support unlimited length of prompt and negative prompt, use A1111 style of prompt weighting | [Stable Diffusion XL Long Weighted Prompt Pipeline](#stable-diffusion-xl-long-weighted-prompt-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1LsqilswLR40XLLcp6XFOl5nKb_wOe26W?usp=sharing) | [Andrew Zhu](https://xhinker.medium.com/) |
+FABRIC - Stable Diffusion with feedback Pipeline | pipeline supports feedback from liked and disliked images | [Stable Diffusion Fabric Pipeline](#stable-diffusion-fabric-pipeline) | - | [Shauray Singh](https://shauray8.github.io/about_shauray/) |
+sketch inpaint - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion Pipeline](#stable-diffusion-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
+prompt-to-prompt | change parts of a prompt and retain image structure (see [paper page](https://prompt-to-prompt.github.io/)) | [Prompt2Prompt Pipeline](#prompt2prompt-pipeline) | - | [Umer H. Adil](https://twitter.com/UmerHAdil) |
 |   Latent Consistency Pipeline                                                                                                    | Implementation of [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://arxiv.org/abs/2310.04378)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Latent Consistency Pipeline](#latent-consistency-pipeline)      | - |              [Simian Luo](https://github.com/luosiallen) |
 |   Latent Consistency Img2img Pipeline                                                                                                    | Img2img pipeline for Latent Consistency Models                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Img2Img Pipeline](#latent-consistency-img2img-pipeline)      | - |              [Logan Zoellner](https://github.com/nagolinc) |
 |   Latent Consistency Interpolation Pipeline                                                                                                    | Interpolate the latent space of Latent Consistency Models with multiple prompts                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Interpolation Pipeline](#latent-consistency-interpolation-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1pK3NrLWJSiJsBynLns1K1-IDTW9zbPvl?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
+| SDE Drag Pipeline                                                                                                                         | The pipeline supports drag editing of images using stochastic differential equations                                                                                                                                                                                                                                                                                                                                                                                                                | [SDE Drag Pipeline](#sde-drag-pipeline)                                                     | - | [NieShen](https://github.com/NieShenRuc) [Fengqi Zhu](https://github.com/Monohydroxides) |
 |   Regional Prompting Pipeline                                                                                               | Assign multiple prompts for different regions                                                                                                                                                                                                                                                                                                                                                    |  [Regional Prompting Pipeline](#regional-prompting-pipeline) | - | [hako-mikan](https://github.com/hako-mikan) |
 | LDM3D-sr (LDM3D upscaler)                                                                                                             | Upscale low resolution RGB and depth inputs to high resolution                                                                                                                                                                                                                                                                                                                                                                                                                              | [StableDiffusionUpscaleLDM3D Pipeline](https://github.com/estelleafl/diffusers/tree/ldm3d_upscaler_community/examples/community#stablediffusionupscaleldm3d-pipeline)                                                                             | -                                                                                                                                                                                                             |                                                        [Estelle Aflalo](https://github.com/estelleafl) |
+| AnimateDiff ControlNet Pipeline                                                                                                    | Combines AnimateDiff with precise motion control using ControlNets                                                                                                                                                                                                                                                                                                                                                                                                                                    | [AnimateDiff ControlNet Pipeline](#animatediff-controlnet-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1SKboYeGjEQmQPWoFC0aLYpBlYdHXkvAu?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) and [Edoardo Botta](https://github.com/EdoardoBotta) |
 |   DemoFusion Pipeline                                                                                                    | Implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [DemoFusion Pipeline](#DemoFusion)      | - |              [Ruoyi Du](https://github.com/RuoyiDu) |
 
 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.
@@ -76,7 +78,7 @@ import torch
 from diffusers import DiffusionPipeline
 
 pipe = DiffusionPipeline.from_pretrained(
-    "longlian/lmd_plus", 
+    "longlian/lmd_plus",
     custom_pipeline="llm_grounded_diffusion",
     custom_revision="main",
     variant="fp16", torch_dtype=torch.float16
@@ -111,7 +113,7 @@ import torch
 from diffusers import DiffusionPipeline
 
 pipe = DiffusionPipeline.from_pretrained(
-    "longlian/lmd_plus", 
+    "longlian/lmd_plus",
     custom_pipeline="llm_grounded_diffusion",
     variant="fp16", torch_dtype=torch.float16
 )
@@ -138,7 +140,7 @@ images[0].save("./lmd_plus_generation.jpg")
 
 ### CLIP Guided Stable Diffusion
 
-CLIP guided stable diffusion can help to generate more realistic images 
+CLIP guided stable diffusion can help to generate more realistic images
 by guiding stable diffusion at every denoising step with an additional CLIP model.
 
 The following code requires roughly 12GB of GPU RAM.
@@ -158,7 +160,7 @@ guided_pipeline = DiffusionPipeline.from_pretrained(
     custom_pipeline="clip_guided_stable_diffusion",
     clip_model=clip_model,
     feature_extractor=feature_extractor,
-    
+
     torch_dtype=torch.float16,
 )
 guided_pipeline.enable_attention_slicing()
@@ -179,7 +181,7 @@ for i in range(4):
         generator=generator,
     ).images[0]
     images.append(image)
-    
+
 # save images locally
 for i, img in enumerate(images):
     img.save(f"./clip_guided_sd/image_{i}.png")
@@ -233,7 +235,7 @@ frame_filepaths = pipe.walk(
 )
 ```
 
-The output of the `walk(...)` function returns a list of images saved under the folder as defined in `output_dir`. You can use these images to create videos of stable diffusion. 
+The output of the `walk(...)` function returns a list of images saved under the folder as defined in `output_dir`. You can use these images to create videos of stable diffusion.
 
 > **Please have a look at https://github.com/nateraw/stable-diffusion-videos for more in-detail information on how to create videos using stable diffusion as well as more feature-complete functionality.**
 
@@ -309,7 +311,7 @@ import torch
 pipe = DiffusionPipeline.from_pretrained(
     'hakurei/waifu-diffusion',
     custom_pipeline="lpw_stable_diffusion",
-    
+
     torch_dtype=torch.float16
 )
 pipe=pipe.to("cuda")
@@ -376,7 +378,7 @@ diffuser_pipeline = DiffusionPipeline.from_pretrained(
     custom_pipeline="speech_to_image_diffusion",
     speech_model=model,
     speech_processor=processor,
-    
+
     torch_dtype=torch.float16,
 )
 
@@ -434,7 +436,7 @@ import torch
 pipe = DiffusionPipeline.from_pretrained(
     "CompVis/stable-diffusion-v1-4",
     custom_pipeline="wildcard_stable_diffusion",
-    
+
     torch_dtype=torch.float16,
 )
 prompt = "__animal__ sitting on a __object__ wearing a __clothing__"
@@ -448,7 +450,7 @@ out = pipe(
 )
 ```
 
-### Composable Stable diffusion 
+### Composable Stable diffusion
 
 [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) proposes conjunction and negation (negative prompts) operators for compositional generation with conditional diffusion models.
 
@@ -498,7 +500,7 @@ tvu.save_image(grid, f'{prompt}_{args.weights}' + '.png')
 ```
 
 ### Imagic Stable Diffusion
-Allows you to edit an image using stable diffusion. 
+Allows you to edit an image using stable diffusion.
 
 ```python
 import requests
@@ -538,7 +540,7 @@ image = res.images[0]
 image.save('./imagic/imagic_image_alpha_2.png')
 ```
 
-### Seed Resizing 
+### Seed Resizing
 Test seed resizing. Originally generate an image in 512 by 512, then generate image with same seed at 512 by 592 using seed resizing. Finally, generate 512 by 592 using original stable diffusion pipeline.
 
 ```python
@@ -666,14 +668,14 @@ diffuser_pipeline = DiffusionPipeline.from_pretrained(
     detection_pipeline=language_detection_pipeline,
     translation_model=trans_model,
     translation_tokenizer=trans_tokenizer,
-    
+
     torch_dtype=torch.float16,
 )
 
 diffuser_pipeline.enable_attention_slicing()
 diffuser_pipeline = diffuser_pipeline.to(device)
 
-prompt = ["a photograph of an astronaut riding a horse", 
+prompt = ["a photograph of an astronaut riding a horse",
           "Una casa en la playa",
           "Ein Hund, der Orange isst",
           "Un restaurant parisien"]
@@ -714,7 +716,7 @@ mask_image = PIL.Image.open(mask_path).convert("RGB").resize((512, 512))
 pipe = DiffusionPipeline.from_pretrained(
     "runwayml/stable-diffusion-inpainting",
     custom_pipeline="img2img_inpainting",
-    
+
     torch_dtype=torch.float16
 )
 pipe = pipe.to("cuda")
@@ -757,8 +759,8 @@ prompt = "a cup"  # the masked out region will be replaced with this
 image = pipe(image=image, text=text, prompt=prompt).images[0]
 ```
 
-### Bit Diffusion 
-Based https://arxiv.org/abs/2208.04202, this is used for diffusion on discrete data - eg, discreate image data, DNA sequence data. An unconditional discreate image can be generated like this: 
+### Bit Diffusion
+Based https://arxiv.org/abs/2208.04202, this is used for diffusion on discrete data - eg, discreate image data, DNA sequence data. An unconditional discreate image can be generated like this:
 
 ```python
 from diffusers import DiffusionPipeline
@@ -836,8 +838,8 @@ Usage:-
 ```python
 from diffusers import DiffusionPipeline
 
-#Return a CheckpointMergerPipeline class that allows you to merge checkpoints. 
-#The checkpoint passed here is ignored. But still pass one of the checkpoints you plan to 
+#Return a CheckpointMergerPipeline class that allows you to merge checkpoints.
+#The checkpoint passed here is ignored. But still pass one of the checkpoints you plan to
 #merge for convenience
 pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="checkpoint_merger")
 
@@ -860,16 +862,16 @@ image = merged_pipe(prompt).images[0]
 ```
 Some examples along with the merge details:
 
-1. "CompVis/stable-diffusion-v1-4" + "hakurei/waifu-diffusion" ; Sigmoid interpolation; alpha = 0.8 
+1. "CompVis/stable-diffusion-v1-4" + "hakurei/waifu-diffusion" ; Sigmoid interpolation; alpha = 0.8
 
 ![Stable plus Waifu Sigmoid 0.8](https://huggingface.co/datasets/NagaSaiAbhinay/CheckpointMergerSamples/resolve/main/stability_v1_4_waifu_sig_0.8.png)
 
-2. "hakurei/waifu-diffusion" + "prompthero/openjourney" ; Inverse Sigmoid interpolation; alpha = 0.8 
+2. "hakurei/waifu-diffusion" + "prompthero/openjourney" ; Inverse Sigmoid interpolation; alpha = 0.8
 
 ![Stable plus Waifu Sigmoid 0.8](https://huggingface.co/datasets/NagaSaiAbhinay/CheckpointMergerSamples/resolve/main/waifu_openjourney_inv_sig_0.8.png)
 
 
-3. "CompVis/stable-diffusion-v1-4" + "hakurei/waifu-diffusion" + "prompthero/openjourney"; Add Difference interpolation; alpha = 0.5 
+3. "CompVis/stable-diffusion-v1-4" + "hakurei/waifu-diffusion" + "prompthero/openjourney"; Add Difference interpolation; alpha = 0.5
 
 ![Stable plus Waifu plus openjourney add_diff 0.5](https://huggingface.co/datasets/NagaSaiAbhinay/CheckpointMergerSamples/resolve/main/stable_waifu_openjourney_add_diff_0.5.png)
 
@@ -936,8 +938,8 @@ pipe = DiffusionPipeline.from_pretrained(
 
 img = Image.open('phone.jpg')
 mix_img = pipe(
-    img, 
-    prompt = 'bed', 
+    img,
+    prompt = 'bed',
     kmin = 0.3,
     kmax = 0.5,
     mix_factor = 0.5,
@@ -1048,7 +1050,7 @@ print(pipeline.prior_scheduler)
 
 ### UnCLIP Text Interpolation Pipeline
 
-This Diffusion Pipeline takes two prompts and interpolates between the two input prompts using spherical interpolation ( slerp ). The input prompts are converted to text embeddings by the pipeline's text_encoder and the interpolation is done on the resulting text_embeddings over the number of steps specified. Defaults to 5 steps. 
+This Diffusion Pipeline takes two prompts and interpolates between the two input prompts using spherical interpolation ( slerp ). The input prompts are converted to text embeddings by the pipeline's text_encoder and the interpolation is done on the resulting text_embeddings over the number of steps specified. Defaults to 5 steps.
 
 ```python
 import torch
@@ -1085,7 +1087,7 @@ The resulting images in order:-
 
 ### UnCLIP Image Interpolation Pipeline
 
-This Diffusion Pipeline takes two images or an image_embeddings tensor of size 2 and interpolates between their embeddings using spherical interpolation ( slerp ). The input images/image_embeddings are converted to image embeddings by the pipeline's image_encoder and the interpolation is done on the resulting image_embeddings over the number of steps specified. Defaults to 5 steps. 
+This Diffusion Pipeline takes two images or an image_embeddings tensor of size 2 and interpolates between their embeddings using spherical interpolation ( slerp ). The input images/image_embeddings are converted to image embeddings by the pipeline's image_encoder and the interpolation is done on the resulting image_embeddings over the number of steps specified. Defaults to 5 steps.
 
 ```python
 import torch
@@ -1126,8 +1128,8 @@ The resulting images in order:-
 ![result5](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_to_flowers_5.png)
 
 ### DDIM Noise Comparative Analysis Pipeline
-#### **Research question: What visual concepts do the diffusion models learn from each noise level during training?**  
-The [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227) paper proposed an approach to answer the above question, which is their second contribution.  
+#### **Research question: What visual concepts do the diffusion models learn from each noise level during training?**
+The [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227) paper proposed an approach to answer the above question, which is their second contribution.
 The approach consists of the following steps:
 
 1. The input is an image x0.
@@ -1169,7 +1171,7 @@ Here is the result of this pipeline (which is DDIM) on CelebA-HQ dataset.
 
 ### CLIP Guided Img2Img Stable Diffusion
 
-CLIP guided Img2Img stable diffusion can help to generate more realistic images with an initial image 
+CLIP guided Img2Img stable diffusion can help to generate more realistic images with an initial image
 by guiding stable diffusion at every denoising step with an additional CLIP model.
 
 The following code requires roughly 12GB of GPU RAM.
@@ -1321,8 +1323,8 @@ target_prompt = "A golden retriever"
 
 # run the pipeline
 result_image = pipeline(
-      base_prompt=base_prompt, 
-      target_prompt=target_prompt, 
+      base_prompt=base_prompt,
+      target_prompt=target_prompt,
       image=cropped_image,
 )
 
@@ -1536,7 +1538,7 @@ python -m pip install intel_extension_for_pytorch==<version_name> -f https://dev
 pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_ipex")
 # For Float32
 pipe.prepare_for_ipex(prompt, dtype=torch.float32, height=512, width=512) #value of image height/width should be consistent with the pipeline inference
-# For BFloat16 
+# For BFloat16
 pipe.prepare_for_ipex(prompt, dtype=torch.bfloat16, height=512, width=512) #value of image height/width should be consistent with the pipeline inference
 ```
 
@@ -1544,7 +1546,7 @@ Then you can use the ipex pipeline in a similar way to the default stable diffus
 ```python
 # For Float32
 image = pipe(prompt, num_inference_steps=20, height=512, width=512).images[0] #value of image height/width should be consistent with 'prepare_for_ipex()'
-# For BFloat16 
+# For BFloat16
 with torch.cpu.amp.autocast(enabled=True, dtype=torch.bfloat16):
     image = pipe(prompt, num_inference_steps=20, height=512, width=512).images[0] #value of image height/width should be consistent with 'prepare_for_ipex()'
 ```
@@ -1603,24 +1605,25 @@ latency = elapsed_time(pipe4)
 print("Latency of StableDiffusionPipeline--fp32",latency)
 
 ```
-  
+
 ### CLIP Guided Images Mixing With Stable Diffusion
 
 ![clip_guided_images_mixing_examples](https://huggingface.co/datasets/TheDenk/images_mixing/resolve/main/main.png)
 
-CLIP guided stable diffusion images mixing pipeline allows to combine two images using standard diffusion models.  
-This approach is using (optional) CoCa model to avoid writing image description.  
+CLIP guided stable diffusion images mixing pipeline allows to combine two images using standard diffusion models.
+This approach is using (optional) CoCa model to avoid writing image description.
 [More code examples](https://github.com/TheDenk/images_mixing)
 
 
 ### Stable Diffusion XL Long Weighted Prompt Pipeline
 
-This SDXL pipeline support unlimited length prompt and negative prompt, compatible with A1111 prompt weighted style. 
+This SDXL pipeline support unlimited length prompt and negative prompt, compatible with A1111 prompt weighted style.
 
-You can provide both `prompt` and `prompt_2`. if only one prompt is provided, `prompt_2` will be a copy of the provided `prompt`. Here is a sample code to use this pipeline. 
+You can provide both `prompt` and `prompt_2`. If only one prompt is provided, `prompt_2` will be a copy of the provided `prompt`. Here is a sample code to use this pipeline. 
 
 ```python
 from diffusers import DiffusionPipeline
+from diffusers.utils import load_image
 import torch
 
 pipe = DiffusionPipeline.from_pretrained(
@@ -1631,25 +1634,52 @@ pipe = DiffusionPipeline.from_pretrained(
     , custom_pipeline   = "lpw_stable_diffusion_xl",
 )
 
-prompt = "photo of a cute (white) cat running on the grass"*20
-prompt2 = "chasing (birds:1.5)"*20
+prompt = "photo of a cute (white) cat running on the grass" * 20
+prompt2 = "chasing (birds:1.5)" * 20
 prompt = f"{prompt},{prompt2}"
 neg_prompt = "blur, low quality, carton, animate"
 
 pipe.to("cuda")
-images = pipe(
-    prompt                  = prompt 
-    , negative_prompt       = neg_prompt 
-).images[0]
+
+# text2img
+t2i_images = pipe(
+    prompt=prompt,
+    negative_prompt=neg_prompt,
+).images # alternatively, you can call the .text2img() function
+
+# img2img
+input_image = load_image("/path/to/local/image.png") # or URL to your input image
+i2i_images = pipe.img2img(
+  prompt=prompt,
+  negative_prompt=neg_prompt,
+  image=input_image,
+  strength=0.8, # higher strength will result in more variation compared to original image
+).images
+
+# inpaint
+input_mask = load_image("/path/to/local/mask.png") # or URL to your input inpainting mask
+inpaint_images = pipe.inpaint(
+  prompt="photo of a cute (black) cat running on the grass" * 20,
+  negative_prompt=neg_prompt,
+  image=input_image,
+  mask=input_mask,
+  strength=0.6, # higher strength will result in more variation compared to original image
+).images
 
 pipe.to("cpu")
 torch.cuda.empty_cache()
-images
+
+from IPython.display import display # assuming you are using this code in a notebook
+display(t2i_images[0])
+display(i2i_images[0])
+display(inpaint_images[0])
 ```
 
-In the above code, the `prompt2` is appended to the `prompt`, which is more than 77 tokens. "birds" are showing up in the result. 
+In the above code, the `prompt2` is appended to the `prompt`, which is more than 77 tokens. "birds" are showing up in the result.
 ![Stable Diffusion XL Long Weighted Prompt Pipeline sample](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_long_weighted_prompt.png)
 
+For more results, checkout [PR #6114](https://github.com/huggingface/diffusers/pull/6114).
+
 ## Example Images Mixing (with CoCa)
 ```python
 import requests
@@ -1699,7 +1729,7 @@ mixing_pipeline.enable_attention_slicing()
 mixing_pipeline = mixing_pipeline.to("cuda")
 
 # Pipeline running
-generator = torch.Generator(device="cuda").manual_seed(17) 
+generator = torch.Generator(device="cuda").manual_seed(17)
 
 def download_image(url):
     response = requests.get(url)
@@ -1728,7 +1758,7 @@ pipe_images = mixing_pipeline(
 ### Stable Diffusion Mixture Tiling
 
 This pipeline uses the Mixture. Refer to the [Mixture](https://arxiv.org/abs/2302.02412) paper for more details.
-    
+
 ```python
 from diffusers import LMSDiscreteScheduler, DiffusionPipeline
 
@@ -1801,7 +1831,7 @@ image.save('tensorrt_inpaint_mecha_robot.png')
 ### Stable Diffusion Mixture Canvas
 
 This pipeline uses the Mixture. Refer to the [Mixture](https://arxiv.org/abs/2302.02412) paper for more details.
-    
+
 ```python
 from PIL import Image
 from diffusers import LMSDiscreteScheduler, DiffusionPipeline
@@ -2010,7 +2040,7 @@ Reference Image
 
 ![reference_image](https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png)
 
-Output Image   
+Output Image
 
 `prompt: 1 girl`
 
@@ -2021,7 +2051,7 @@ Reference Image
 ![reference_image](https://github.com/huggingface/diffusers/assets/34944964/449bdab6-e744-4fb2-9620-d4068d9a741b)
 
 
-Output Image 
+Output Image
 
 `prompt: A dog`
 
@@ -2102,7 +2132,7 @@ Let's have a look at the images (*512X512*)
 
 | Without Feedback            | With Feedback  (1st image)          |
 |---------------------|---------------------|
-| ![Image 1](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/fabric_wo_feedback.jpg) | ![Feedback Image 1](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/fabric_w_feedback.png) | 
+| ![Image 1](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/fabric_wo_feedback.jpg) | ![Feedback Image 1](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/fabric_w_feedback.png) |
 
 
 ### Masked Im2Im Stable Diffusion Pipeline
@@ -2255,7 +2285,7 @@ pipe.to(torch_device="cuda", torch_dtype=torch.float32)
 prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
 
 # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
-num_inference_steps = 4 
+num_inference_steps = 4
 
 images = pipe(prompt=prompt, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
 ```
@@ -2291,7 +2321,7 @@ input_image=Image.open("myimg.png")
 strength = 0.5 #strength =0 (no change) strength=1 (completely overwrite image)
 
 # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
-num_inference_steps = 4 
+num_inference_steps = 4
 
 images = pipe(prompt=prompt, image=input_image, strength=strength, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
 ```
@@ -2344,7 +2374,7 @@ assert len(images) == (len(prompts) - 1) * num_interpolation_steps
 ```
 
 ###  StableDiffusionUpscaleLDM3D Pipeline
-[LDM3D-VR](https://arxiv.org/pdf/2311.03226.pdf) is an extended version of LDM3D. 
+[LDM3D-VR](https://arxiv.org/pdf/2311.03226.pdf) is an extended version of LDM3D.
 
 The abstract from the paper is:
 *Latent diffusion models have proven to be state-of-the-art in the creation and manipulation of visual outputs. However, as far as we know, the generation of depth maps jointly with RGB is still limited. We introduce LDM3D-VR, a suite of diffusion models targeting virtual reality development that includes LDM3D-pano and LDM3D-SR. These models enable the generation of panoramic RGBD based on textual prompts and the upscaling of low-resolution inputs to high-resolution RGBD, respectively. Our models are fine-tuned from existing pretrained models on datasets containing panoramic/high-resolution RGB images, depth maps and captions. Both models are evaluated in comparison to existing related methods*
@@ -2385,8 +2415,8 @@ upscaled_depth.save(f"upscaled_lemons_depth.png")
 '''
 
 ### ControlNet + T2I Adapter Pipeline
-This pipelines combines both ControlNet and T2IAdapter into a single pipeline, where the forward pass is executed once. 
-It receives `control_image` and `adapter_image`, as well as `controlnet_conditioning_scale` and `adapter_conditioning_scale`, for the ControlNet and Adapter modules, respectively. Whenever `adapter_conditioning_scale = 0` or `controlnet_conditioning_scale = 0`, it will act as a full ControlNet module or as a full T2IAdapter module, respectively. 
+This pipelines combines both ControlNet and T2IAdapter into a single pipeline, where the forward pass is executed once.
+It receives `control_image` and `adapter_image`, as well as `controlnet_conditioning_scale` and `adapter_conditioning_scale`, for the ControlNet and Adapter modules, respectively. Whenever `adapter_conditioning_scale = 0` or `controlnet_conditioning_scale = 0`, it will act as a full ControlNet module or as a full T2IAdapter module, respectively.
 
 ```py
 import cv2
@@ -2537,7 +2567,7 @@ pipe = RegionalPromptingStableDiffusionPipeline.from_single_file(model_path, vae
 rp_args = {
     "mode":"rows",
     "div": "1;1;1"
-}  
+}
 
 prompt ="""
 green hair twintail BREAK
@@ -2566,7 +2596,7 @@ for image in images:
 ### Cols, Rows mode
 In the Cols, Rows mode, you can split the screen vertically and horizontally and assign prompts to each region. The split ratio can be specified by 'div', and you can set the division ratio like '3;3;2' or '0.1;0.5'. Furthermore, as will be described later, you can also subdivide the split Cols, Rows to specify more complex regions.
 
-In this image, the image is divided into three parts, and a separate prompt is applied to each. The prompts are divided by 'BREAK', and each is applied to the respective region.  
+In this image, the image is divided into three parts, and a separate prompt is applied to each. The prompts are divided by 'BREAK', and each is applied to the respective region.
 ![sample](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/imgs/rp_pipeline2.png)
 ```
 green hair twintail BREAK
@@ -2624,7 +2654,7 @@ prompt ="""
 a girl in street with shirt, tie, skirt BREAK
 red, shirt BREAK
 green, tie BREAK
-blue , skirt 
+blue , skirt
 """
 ```
 ![sample](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/imgs/rp_pipeline3.png)
@@ -2643,7 +2673,7 @@ If only one input is given for multiple regions, they are all assumed to be the
 The difference is that in Prompt, duplicate regions are added, whereas in Prompt-EX, duplicate regions are overwritten sequentially. Since they are processed in order, setting a TARGET with a large regions first makes it easier for the effect of small regions to remain unmuffled.
 
 ### Accuracy
-In the case of a 512 x 512 image, Attention mode reduces the size of the region to about 8 x 8 pixels deep in the U-Net, so that small regions get mixed up; Latent mode calculates 64*64, so that the region is exact.  
+In the case of a 512 x 512 image, Attention mode reduces the size of the region to about 8 x 8 pixels deep in the U-Net, so that small regions get mixed up; Latent mode calculates 64*64, so that the region is exact.
 ```
 girl hair twintail frills,ribbons, dress, face BREAK
 girl, ,face
@@ -2674,13 +2704,13 @@ Negative prompts are equally effective across all regions, but it is possible to
 To activate Regional Prompter, it is necessary to enter settings in rp_args. The items that can be set are as follows. rp_args is a dictionary type.
 
 ### Input Parameters
-Parameters are specified through the `rp_arg`(dictionary type).  
+Parameters are specified through the `rp_arg`(dictionary type).
 
 ```
 rp_args = {
     "mode":"rows",
     "div": "1;1;1"
-}  
+}
 
 pipe(prompt =prompt, rp_args = rp_args)
 ```
@@ -2759,7 +2789,7 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
 
                 def get_kernel(self):
                     return self.k
-                
+
             self.kernel_size = kernel_size
             self.conv = Blurkernel(blur_type='gaussian',
                                 kernel_size=kernel_size,
@@ -2834,11 +2864,75 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
     * ![sample](https://github.com/tongdaxu/Images/assets/22267548/0ceb5575-d42e-4f0b-99c0-50e69c982209)
 * The reconstruction is perceptually similar to the source image, but different in details.
 * In dps_pipeline.py, we also provide a super-resolution example, which should produce:
-    * Downsampled image: 
+    * Downsampled image:
     * ![dps_mea](https://github.com/tongdaxu/Images/assets/22267548/ff6a33d6-26f0-42aa-88ce-f8a76ba45a13)
     * Reconstructed image:
     * ![dps_generated_image](https://github.com/tongdaxu/Images/assets/22267548/b74f084d-93f4-4845-83d8-44c0fa758a5f)
 
+### AnimateDiff ControlNet Pipeline
+
+This pipeline combines AnimateDiff and ControlNet. Enjoy precise motion control for your videos! Refer to [this](https://github.com/huggingface/diffusers/issues/5866) issue for more details.
+
+```py
+import torch
+from diffusers import AutoencoderKL, ControlNetModel, MotionAdapter
+from diffusers.pipelines import DiffusionPipeline
+from diffusers.schedulers import DPMSolverMultistepScheduler
+from PIL import Image
+
+motion_id = "guoyww/animatediff-motion-adapter-v1-5-2"
+adapter = MotionAdapter.from_pretrained(motion_id)
+controlnet = ControlNetModel.from_pretrained("lllyasviel/control_v11p_sd15_openpose", torch_dtype=torch.float16)
+vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16)
+
+model_id = "SG161222/Realistic_Vision_V5.1_noVAE"
+pipe = DiffusionPipeline.from_pretrained(
+    model_id,
+    motion_adapter=adapter,
+    controlnet=controlnet,
+    vae=vae,
+    custom_pipeline="pipeline_animatediff_controlnet",
+).to(device="cuda", dtype=torch.float16)
+pipe.scheduler = DPMSolverMultistepScheduler.from_pretrained(
+    model_id, subfolder="scheduler", clip_sample=False, timestep_spacing="linspace", steps_offset=1
+)
+pipe.enable_vae_slicing()
+
+conditioning_frames = []
+for i in range(1, 16 + 1):
+    conditioning_frames.append(Image.open(f"frame_{i}.png"))
+
+prompt = "astronaut in space, dancing"
+negative_prompt = "bad quality, worst quality, jpeg artifacts, ugly"
+result = pipe(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    width=512,
+    height=768,
+    conditioning_frames=conditioning_frames,
+    num_inference_steps=12,
+).frames[0]
+
+from diffusers.utils import export_to_gif
+export_to_gif(result.frames[0], "result.gif")
+```
+
+<table>
+  <tr><td colspan="2" align=center><b>Conditioning Frames</b></td></tr>
+  <tr align=center>
+    <td align=center><img src="https://user-images.githubusercontent.com/7365912/265043418-23291941-864d-495a-8ba8-d02e05756396.gif" alt="input-frames"></td>
+  </tr>
+  <tr><td colspan="2" align=center><b>AnimateDiff model: SG161222/Realistic_Vision_V5.1_noVAE</b></td></tr>
+  <tr>
+    <td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/baf301e2-d03c-4129-bd84-203a1de2b2be" alt="gif-1"></td>
+    <td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/9f923475-ecaf-452b-92c8-4e42171182d8" alt="gif-2"></td>
+  </tr>
+  <tr><td colspan="2" align=center><b>AnimateDiff model: CardosAnime</b></td></tr>
+  <tr>
+    <td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/b2c41028-38a0-45d6-86ed-fec7446b87f7" alt="gif-1"></td>
+    <td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/eb7d2952-72e4-44fa-b664-077c79b4fc70" alt="gif-2"></td>
+  </tr>
+</table>
 ### DemoFusion
 This pipeline is the official implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973).
 The original repo can be found at [repo](https://github.com/PRIS-CV/DemoFusion).
@@ -2865,7 +2959,7 @@ The original repo can be found at [repo](https://github.com/PRIS-CV/DemoFusion).
 
 - `show_image` (`bool`, defaults to False):
   Determine whether to show intermediate results during generation.
-```
+```py
 from diffusers import DiffusionPipeline
 
 pipe = DiffusionPipeline.from_pretrained(
@@ -2880,24 +2974,24 @@ prompt = "Envision a portrait of an elderly woman, her face a canvas of time, fr
 negative_prompt = "blurry, ugly, duplicate, poorly drawn, deformed, mosaic"
 
 images = pipe(
-    prompt, 
+    prompt,
     negative_prompt=negative_prompt,
-    height=3072, 
-    width=3072, 
-    view_batch_size=16, 
+    height=3072,
+    width=3072,
+    view_batch_size=16,
     stride=64,
-    num_inference_steps=50, 
+    num_inference_steps=50,
     guidance_scale=7.5,
-    cosine_scale_1=3, 
-    cosine_scale_2=1, 
-    cosine_scale_3=1, 
+    cosine_scale_1=3,
+    cosine_scale_2=1,
+    cosine_scale_3=1,
     sigma=0.8,
-    multi_decoder=True, 
+    multi_decoder=True,
     show_image=True
 )
 ```
 You can display and save the generated images as:
-```
+```py
 def image_grid(imgs, save_path=None):
 
     w = 0
@@ -2915,9 +3009,48 @@ def image_grid(imgs, save_path=None):
         if save_path != None:
             img.save(save_path + "/img_{}.jpg".format((i + 1) * 1024))
         w += w_
-        
+
     return grid
 
 image_grid(images, save_path="./outputs/")
 ```
  ![output_example](https://github.com/PRIS-CV/DemoFusion/blob/main/output_example.png)
+
+### SDE Drag pipeline
+
+This pipeline provides drag-and-drop image editing using stochastic differential equations. It enables image editing by inputting prompt, image, mask_image, source_points, and target_points.
+
+![SDE Drag Image](https://github.com/huggingface/diffusers/assets/75928535/bd54f52f-f002-4951-9934-b2a4592771a5)
+
+See [paper](https://arxiv.org/abs/2311.01410), [paper page](https://ml-gsai.github.io/SDE-Drag-demo/), [original repo](https://github.com/ML-GSAI/SDE-Drag) for more infomation.
+
+```py
+import PIL
+import torch
+from diffusers import DDIMScheduler, DiffusionPipeline
+
+# Load the pipeline
+model_path = "runwayml/stable-diffusion-v1-5"
+scheduler = DDIMScheduler.from_pretrained(model_path, subfolder="scheduler")
+pipe = DiffusionPipeline.from_pretrained(model_path, scheduler=scheduler, custom_pipeline="sde_drag")
+pipe.to('cuda')
+
+# To save GPU memory, torch.float16 can be used, but it may compromise image quality.
+# If not training LoRA, please avoid using torch.float16
+# pipe.to(torch.float16)
+
+# Provide prompt, image, mask image, and the starting and target points for drag editing.
+prompt = "prompt of the image"
+image = PIL.Image.open('/path/to/image')
+mask_image = PIL.Image.open('/path/to/mask_image')
+source_points = [[123, 456]]
+target_points = [[234, 567]]
+
+# train_lora is optional, and in most cases, using train_lora can better preserve consistency with the original image.
+pipe.train_lora(prompt, image)
+
+output = pipe(prompt, image, mask_image, source_points, target_points)
+output_image = PIL.Image.fromarray(output)
+output_image.save("./output.png")
+
+```

examples/community/lpw_stable_diffusion_xl.py

@@ -11,10 +11,11 @@ import os
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import torch
+from PIL import Image
 from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
 
 from diffusers import DiffusionPipeline, StableDiffusionXLPipeline
-from diffusers.image_processor import VaeImageProcessor
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
 from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.models.attention_processor import (
@@ -23,7 +24,7 @@ from diffusers.models.attention_processor import (
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
 )
-from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
+from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     is_accelerate_available,
@@ -461,6 +462,65 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     return noise_cfg
 
 
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used,
+            `timesteps` must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+                timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+                must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
 class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion XL.
@@ -526,6 +586,9 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
         self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
         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.mask_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
+        )
         self.default_sample_size = self.unet.config.sample_size
 
         add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
@@ -813,6 +876,7 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
         prompt_2,
         height,
         width,
+        strength,
         callback_steps,
         negative_prompt=None,
         negative_prompt_2=None,
@@ -824,6 +888,9 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
         if height % 8 != 0 or width % 8 != 0:
             raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
 
+        if strength < 0 or strength > 1:
+            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
         if (callback_steps is None) or (
             callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
         ):
@@ -880,23 +947,263 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
                 "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
             )
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
-    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
-        if isinstance(generator, list) and len(generator) != batch_size:
-            raise ValueError(
-                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
-                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+    def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None):
+        # get the original timestep using init_timestep
+        if denoising_start is None:
+            init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+            t_start = max(num_inference_steps - init_timestep, 0)
+        else:
+            t_start = 0
+
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+        # Strength is irrelevant if we directly request a timestep to start at;
+        # that is, strength is determined by the denoising_start instead.
+        if denoising_start is not None:
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (denoising_start * self.scheduler.config.num_train_timesteps)
+                )
             )
 
-        if latents is None:
-            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
-        else:
-            latents = latents.to(device)
+            num_inference_steps = (timesteps < discrete_timestep_cutoff).sum().item()
+            if self.scheduler.order == 2 and num_inference_steps % 2 == 0:
+                # if the scheduler is a 2nd order scheduler we might have to do +1
+                # because `num_inference_steps` might be even given that every timestep
+                # (except the highest one) is duplicated. If `num_inference_steps` is even it would
+                # mean that we cut the timesteps in the middle of the denoising step
+                # (between 1st and 2nd devirative) which leads to incorrect results. By adding 1
+                # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler
+                num_inference_steps = num_inference_steps + 1
 
-        # scale the initial noise by the standard deviation required by the scheduler
-        latents = latents * self.scheduler.init_noise_sigma
-        return latents
+            # because t_n+1 >= t_n, we slice the timesteps starting from the end
+            timesteps = timesteps[-num_inference_steps:]
+            return timesteps, num_inference_steps
+
+        return timesteps, num_inference_steps - t_start
+
+    def prepare_latents(
+        self,
+        image,
+        mask,
+        width,
+        height,
+        num_channels_latents,
+        timestep,
+        batch_size,
+        num_images_per_prompt,
+        dtype,
+        device,
+        generator=None,
+        add_noise=True,
+        latents=None,
+        is_strength_max=True,
+        return_noise=False,
+        return_image_latents=False,
+    ):
+        batch_size *= num_images_per_prompt
+
+        if image is None:
+            shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+            if isinstance(generator, list) and len(generator) != batch_size:
+                raise ValueError(
+                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                )
+
+            if latents is None:
+                latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+            else:
+                latents = latents.to(device)
+
+            # scale the initial noise by the standard deviation required by the scheduler
+            latents = latents * self.scheduler.init_noise_sigma
+            return latents
+
+        elif mask is None:
+            if not isinstance(image, (torch.Tensor, Image.Image, list)):
+                raise ValueError(
+                    f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+                )
+
+            # Offload text encoder if `enable_model_cpu_offload` was enabled
+            if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+                self.text_encoder_2.to("cpu")
+                torch.cuda.empty_cache()
+
+            image = image.to(device=device, dtype=dtype)
+
+            if image.shape[1] == 4:
+                init_latents = image
+
+            else:
+                # make sure the VAE is in float32 mode, as it overflows in float16
+                if self.vae.config.force_upcast:
+                    image = image.float()
+                    self.vae.to(dtype=torch.float32)
+
+                if isinstance(generator, list) and len(generator) != batch_size:
+                    raise ValueError(
+                        f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                        f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                    )
+
+                elif isinstance(generator, list):
+                    init_latents = [
+                        retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+                        for i in range(batch_size)
+                    ]
+                    init_latents = torch.cat(init_latents, dim=0)
+                else:
+                    init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+                if self.vae.config.force_upcast:
+                    self.vae.to(dtype)
+
+                init_latents = init_latents.to(dtype)
+                init_latents = self.vae.config.scaling_factor * init_latents
+
+            if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+                # expand init_latents for batch_size
+                additional_image_per_prompt = batch_size // init_latents.shape[0]
+                init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+            elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+                raise ValueError(
+                    f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+                )
+            else:
+                init_latents = torch.cat([init_latents], dim=0)
+
+            if add_noise:
+                shape = init_latents.shape
+                noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+                # get latents
+                init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+
+            latents = init_latents
+            return latents
+
+        else:
+            shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+            if isinstance(generator, list) and len(generator) != batch_size:
+                raise ValueError(
+                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                )
+
+            if (image is None or timestep is None) and not is_strength_max:
+                raise ValueError(
+                    "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
+                    "However, either the image or the noise timestep has not been provided."
+                )
+
+            if image.shape[1] == 4:
+                image_latents = image.to(device=device, dtype=dtype)
+                image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
+            elif return_image_latents or (latents is None and not is_strength_max):
+                image = image.to(device=device, dtype=dtype)
+                image_latents = self._encode_vae_image(image=image, generator=generator)
+                image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
+
+            if latents is None and add_noise:
+                noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+                # if strength is 1. then initialise the latents to noise, else initial to image + noise
+                latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
+                # if pure noise then scale the initial latents by the  Scheduler's init sigma
+                latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+            elif add_noise:
+                noise = latents.to(device)
+                latents = noise * self.scheduler.init_noise_sigma
+            else:
+                noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+                latents = image_latents.to(device)
+
+            outputs = (latents,)
+
+            if return_noise:
+                outputs += (noise,)
+
+            if return_image_latents:
+                outputs += (image_latents,)
+
+            return outputs
+
+    def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+        dtype = image.dtype
+        if self.vae.config.force_upcast:
+            image = image.float()
+            self.vae.to(dtype=torch.float32)
+
+        if isinstance(generator, list):
+            image_latents = [
+                retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+                for i in range(image.shape[0])
+            ]
+            image_latents = torch.cat(image_latents, dim=0)
+        else:
+            image_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+        if self.vae.config.force_upcast:
+            self.vae.to(dtype)
+
+        image_latents = image_latents.to(dtype)
+        image_latents = self.vae.config.scaling_factor * image_latents
+
+        return image_latents
+
+    def prepare_mask_latents(
+        self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+    ):
+        # resize the mask to latents shape as we concatenate the mask to the latents
+        # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+        # and half precision
+        mask = torch.nn.functional.interpolate(
+            mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+        )
+        mask = mask.to(device=device, dtype=dtype)
+
+        # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+        if mask.shape[0] < batch_size:
+            if not batch_size % mask.shape[0] == 0:
+                raise ValueError(
+                    "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+                    f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+                    " of masks that you pass is divisible by the total requested batch size."
+                )
+            mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+
+        mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+
+        if masked_image is not None and masked_image.shape[1] == 4:
+            masked_image_latents = masked_image
+        else:
+            masked_image_latents = None
+
+        if masked_image is not None:
+            if masked_image_latents is None:
+                masked_image = masked_image.to(device=device, dtype=dtype)
+                masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
+
+            if masked_image_latents.shape[0] < batch_size:
+                if not batch_size % masked_image_latents.shape[0] == 0:
+                    raise ValueError(
+                        "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+                        f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+                        " Make sure the number of images that you pass is divisible by the total requested batch size."
+                    )
+                masked_image_latents = masked_image_latents.repeat(
+                    batch_size // masked_image_latents.shape[0], 1, 1, 1
+                )
+
+            masked_image_latents = (
+                torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+            )
+
+            # aligning device to prevent device errors when concating it with the latent model input
+            masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+
+        return mask, masked_image_latents
 
     def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
         add_time_ids = list(original_size + crops_coords_top_left + target_size)
@@ -934,15 +1241,52 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
             self.vae.decoder.conv_in.to(dtype)
             self.vae.decoder.mid_block.to(dtype)
 
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def guidance_rescale(self):
+        return self._guidance_rescale
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+    @property
+    def cross_attention_kwargs(self):
+        return self._cross_attention_kwargs
+
+    @property
+    def denoising_end(self):
+        return self._denoising_end
+
+    @property
+    def denoising_start(self):
+        return self._denoising_start
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
         self,
         prompt: str = None,
         prompt_2: Optional[str] = None,
+        image: Optional[PipelineImageInput] = None,
+        mask_image: Optional[PipelineImageInput] = None,
+        masked_image_latents: Optional[torch.FloatTensor] = None,
         height: Optional[int] = None,
         width: Optional[int] = None,
+        strength: float = 0.8,
         num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        denoising_start: Optional[float] = None,
         denoising_end: Optional[float] = None,
         guidance_scale: float = 5.0,
         negative_prompt: Optional[str] = None,
@@ -975,20 +1319,46 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
             prompt_2 (`str`):
                 The prompt to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                 used in both text-encoders
+            image (`PipelineImageInput`, *optional*):
+                `Image`, or tensor representing an image batch, that will be used as the starting point for the
+                process.
+            mask_image (`PipelineImageInput`, *optional*):
+                `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+                replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
+                PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should
+                contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.
             height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The height in pixels of the generated image.
             width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                 The width in pixels of the generated image.
+            strength (`float`, *optional*, defaults to 0.8):
+                Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1.
+                `image` will be used as a starting point, adding more noise to it the larger the `strength`. The
+                number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
+                noise will be maximum and the denoising process will run for the full number of iterations specified in
+                `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
             num_inference_steps (`int`, *optional*, defaults to 50):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            denoising_start (`float`, *optional*):
+                When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be
+                bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and
+                it is assumed that the passed `image` is a partly denoised image. Note that when this is specified,
+                strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline
+                is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refine Image
+                Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality).
             denoising_end (`float`, *optional*):
                 When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
                 completed before it is intentionally prematurely terminated. As a result, the returned sample will
-                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
-                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
-                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
-                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+                still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be
+                denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the
+                final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline
+                forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refine Image
+                Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality).
             guidance_scale (`float`, *optional*, defaults to 5.0):
                 Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                 `guidance_scale` is defined as `w` of equation 2. of [Imagen
@@ -1084,6 +1454,7 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
             prompt_2,
             height,
             width,
+            strength,
             callback_steps,
             negative_prompt,
             negative_prompt_2,
@@ -1093,6 +1464,12 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
             negative_pooled_prompt_embeds,
         )
 
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._denoising_end = denoising_end
+        self._denoising_start = denoising_start
+
         # 2. Define call parameters
         if prompt is not None and isinstance(prompt, str):
             batch_size = 1
@@ -1121,28 +1498,126 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
         ) = get_weighted_text_embeddings_sdxl(
             pipe=self, prompt=prompt, neg_prompt=negative_prompt, num_images_per_prompt=num_images_per_prompt
         )
+        dtype = prompt_embeds.dtype
+
+        if isinstance(image, Image.Image):
+            image = self.image_processor.preprocess(image, height=height, width=width)
+        if image is not None:
+            image = image.to(device=self.device, dtype=dtype)
+
+        if isinstance(mask_image, Image.Image):
+            mask = self.mask_processor.preprocess(mask_image, height=height, width=width)
+        else:
+            mask = mask_image
+        if mask_image is not None:
+            mask = mask.to(device=self.device, dtype=dtype)
+
+            if masked_image_latents is not None:
+                masked_image = masked_image_latents
+            elif image.shape[1] == 4:
+                # if image is in latent space, we can't mask it
+                masked_image = None
+            else:
+                masked_image = image * (mask < 0.5)
+        else:
+            mask = None
 
         # 4. Prepare timesteps
-        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        def denoising_value_valid(dnv):
+            return isinstance(self.denoising_end, float) and 0 < dnv < 1
 
-        timesteps = self.scheduler.timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        if image is not None:
+            timesteps, num_inference_steps = self.get_timesteps(
+                num_inference_steps,
+                strength,
+                device,
+                denoising_start=self.denoising_start if denoising_value_valid else None,
+            )
+
+            # check that number of inference steps is not < 1 - as this doesn't make sense
+            if num_inference_steps < 1:
+                raise ValueError(
+                    f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
+                    f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
+                )
+
+        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+        is_strength_max = strength == 1.0
+        add_noise = True if self.denoising_start is None else False
 
         # 5. Prepare latent variables
-        num_channels_latents = self.unet.config.in_channels
+        num_channels_latents = self.vae.config.latent_channels
+        num_channels_unet = self.unet.config.in_channels
+        return_image_latents = num_channels_unet == 4
+
         latents = self.prepare_latents(
-            batch_size * num_images_per_prompt,
-            num_channels_latents,
-            height,
-            width,
-            prompt_embeds.dtype,
-            device,
-            generator,
-            latents,
+            image=image,
+            mask=mask,
+            width=width,
+            height=height,
+            num_channels_latents=num_channels_unet,
+            timestep=latent_timestep,
+            batch_size=batch_size,
+            num_images_per_prompt=num_images_per_prompt,
+            dtype=prompt_embeds.dtype,
+            device=device,
+            generator=generator,
+            add_noise=add_noise,
+            latents=latents,
+            is_strength_max=is_strength_max,
+            return_noise=True,
+            return_image_latents=return_image_latents,
         )
 
+        if mask is not None:
+            if return_image_latents:
+                latents, noise, image_latents = latents
+            else:
+                latents, noise = latents
+
+        # 5.1. Prepare mask latent variables
+        if mask is not None:
+            mask, masked_image_latents = self.prepare_mask_latents(
+                mask=mask,
+                masked_image=masked_image,
+                batch_size=batch_size * num_images_per_prompt,
+                height=height,
+                width=width,
+                dtype=prompt_embeds.dtype,
+                device=device,
+                generator=generator,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+            )
+
+            # 8. Check that sizes of mask, masked image and latents match
+            if num_channels_unet == 9:
+                # default case for runwayml/stable-diffusion-inpainting
+                num_channels_mask = mask.shape[1]
+                num_channels_masked_image = masked_image_latents.shape[1]
+                if num_channels_latents + num_channels_mask + num_channels_masked_image != num_channels_unet:
+                    raise ValueError(
+                        f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+                        f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+                        f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+                        f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                        " `pipeline.unet` or your `mask_image` or `image` input."
+                    )
+            elif num_channels_unet != 4:
+                raise ValueError(
+                    f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
+                )
+
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        height, width = latents.shape[-2:]
+        height = height * self.vae_scale_factor
+        width = width * self.vae_scale_factor
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
         # 7. Prepare added time ids & embeddings
         add_text_embeds = pooled_prompt_embeds
         add_time_ids = self._get_add_time_ids(
@@ -1158,20 +1633,41 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
         add_text_embeds = add_text_embeds.to(device)
         add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
 
-        # 8. Denoising loop
         num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
 
         # 7.1 Apply denoising_end
-        if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
+        if (
+            self.denoising_end is not None
+            and self.denoising_start is not None
+            and denoising_value_valid(self.denoising_end)
+            and denoising_value_valid(self.denoising_start)
+            and self.denoising_start >= self.denoising_end
+        ):
+            raise ValueError(
+                f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: "
+                + f" {self.denoising_end} when using type float."
+            )
+        elif self.denoising_end is not None and denoising_value_valid(self.denoising_end):
             discrete_timestep_cutoff = int(
                 round(
                     self.scheduler.config.num_train_timesteps
-                    - (denoising_end * self.scheduler.config.num_train_timesteps)
+                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
                 )
             )
             num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
             timesteps = timesteps[:num_inference_steps]
 
+        # 8. Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        self._num_timesteps = len(timesteps)
+
+        # 9. Denoising loop
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
                 # expand the latents if we are doing classifier free guidance
@@ -1179,13 +1675,17 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
 
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
+                if mask is not None and num_channels_unet == 9:
+                    latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
                 # predict the noise residual
                 added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
                 noise_pred = self.unet(
                     latent_model_input,
                     t,
                     encoder_hidden_states=prompt_embeds,
-                    cross_attention_kwargs=cross_attention_kwargs,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
                     added_cond_kwargs=added_cond_kwargs,
                     return_dict=False,
                 )[0]
@@ -1202,6 +1702,22 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
                 # compute the previous noisy sample x_t -> x_t-1
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
 
+                if mask is not None and num_channels_unet == 4:
+                    init_latents_proper = image_latents
+
+                    if self.do_classifier_free_guidance:
+                        init_mask, _ = mask.chunk(2)
+                    else:
+                        init_mask = mask
+
+                    if i < len(timesteps) - 1:
+                        noise_timestep = timesteps[i + 1]
+                        init_latents_proper = self.scheduler.add_noise(
+                            init_latents_proper, noise, torch.tensor([noise_timestep])
+                        )
+
+                    latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
                 # call the callback, if provided
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
@@ -1241,6 +1757,204 @@ class SDXLLongPromptWeightingPipeline(DiffusionPipeline, FromSingleFileMixin, Lo
 
         return StableDiffusionXLPipelineOutput(images=image)
 
+    def text2img(
+        self,
+        prompt: str = None,
+        prompt_2: Optional[str] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        denoising_start: Optional[float] = None,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Optional[Tuple[int, int]] = None,
+    ):
+        return self.__call__(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            height=height,
+            width=width,
+            num_inference_steps=num_inference_steps,
+            timesteps=timesteps,
+            denoising_start=denoising_start,
+            denoising_end=denoising_end,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            num_images_per_prompt=num_images_per_prompt,
+            eta=eta,
+            generator=generator,
+            latents=latents,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            output_type=output_type,
+            return_dict=return_dict,
+            callback=callback,
+            callback_steps=callback_steps,
+            cross_attention_kwargs=cross_attention_kwargs,
+            guidance_rescale=guidance_rescale,
+            original_size=original_size,
+            crops_coords_top_left=crops_coords_top_left,
+            target_size=target_size,
+        )
+
+    def img2img(
+        self,
+        prompt: str = None,
+        prompt_2: Optional[str] = None,
+        image: Optional[PipelineImageInput] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        strength: float = 0.8,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        denoising_start: Optional[float] = None,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Optional[Tuple[int, int]] = None,
+    ):
+        return self.__call__(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            image=image,
+            height=height,
+            width=width,
+            strength=strength,
+            num_inference_steps=num_inference_steps,
+            timesteps=timesteps,
+            denoising_start=denoising_start,
+            denoising_end=denoising_end,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            num_images_per_prompt=num_images_per_prompt,
+            eta=eta,
+            generator=generator,
+            latents=latents,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            output_type=output_type,
+            return_dict=return_dict,
+            callback=callback,
+            callback_steps=callback_steps,
+            cross_attention_kwargs=cross_attention_kwargs,
+            guidance_rescale=guidance_rescale,
+            original_size=original_size,
+            crops_coords_top_left=crops_coords_top_left,
+            target_size=target_size,
+        )
+
+    def inpaint(
+        self,
+        prompt: str = None,
+        prompt_2: Optional[str] = None,
+        image: Optional[PipelineImageInput] = None,
+        mask_image: Optional[PipelineImageInput] = None,
+        masked_image_latents: Optional[torch.FloatTensor] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        strength: float = 0.8,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        denoising_start: Optional[float] = None,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Optional[Tuple[int, int]] = None,
+    ):
+        return self.__call__(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            image=image,
+            mask_image=mask_image,
+            masked_image_latents=masked_image_latents,
+            height=height,
+            width=width,
+            strength=strength,
+            num_inference_steps=num_inference_steps,
+            timesteps=timesteps,
+            denoising_start=denoising_start,
+            denoising_end=denoising_end,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            num_images_per_prompt=num_images_per_prompt,
+            eta=eta,
+            generator=generator,
+            latents=latents,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            output_type=output_type,
+            return_dict=return_dict,
+            callback=callback,
+            callback_steps=callback_steps,
+            cross_attention_kwargs=cross_attention_kwargs,
+            guidance_rescale=guidance_rescale,
+            original_size=original_size,
+            crops_coords_top_left=crops_coords_top_left,
+            target_size=target_size,
+        )
+
     # Overrride to properly handle the loading and unloading of the additional text encoder.
     def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
         # We could have accessed the unet config from `lora_state_dict()` too. We pass

examples/community/regional_prompting_stable_diffusion.py

@@ -73,7 +73,14 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
         requires_safety_checker: bool = True,
     ):
         super().__init__(
-            vae, text_encoder, tokenizer, unet, scheduler, safety_checker, feature_extractor, requires_safety_checker
+            vae,
+            text_encoder,
+            tokenizer,
+            unet,
+            scheduler,
+            safety_checker,
+            feature_extractor,
+            requires_safety_checker,
         )
         self.register_modules(
             vae=vae,
@@ -102,22 +109,22 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
         return_dict: bool = True,
         rp_args: Dict[str, str] = None,
     ):
-        active = KBRK in prompt[0] if type(prompt) == list else KBRK in prompt  # noqa: E721
+        active = KBRK in prompt[0] if isinstance(prompt, list) else KBRK in prompt
         if negative_prompt is None:
-            negative_prompt = "" if type(prompt) == str else [""] * len(prompt)  # noqa: E721
+            negative_prompt = "" if isinstance(prompt, str) else [""] * len(prompt)
 
         device = self._execution_device
         regions = 0
 
         self.power = int(rp_args["power"]) if "power" in rp_args else 1
 
-        prompts = prompt if type(prompt) == list else [prompt]  # noqa: E721
-        n_prompts = negative_prompt if type(negative_prompt) == list else [negative_prompt]  # noqa: E721
+        prompts = prompt if isinstance(prompt, list) else [prompt]
+        n_prompts = negative_prompt if isinstance(prompt, str) else [negative_prompt]
         self.batch = batch = num_images_per_prompt * len(prompts)
         all_prompts_cn, all_prompts_p = promptsmaker(prompts, num_images_per_prompt)
         all_n_prompts_cn, _ = promptsmaker(n_prompts, num_images_per_prompt)
 
-        cn = len(all_prompts_cn) == len(all_n_prompts_cn)
+        equal = len(all_prompts_cn) == len(all_n_prompts_cn)
 
         if Compel:
             compel = Compel(tokenizer=self.tokenizer, text_encoder=self.text_encoder)
@@ -129,7 +136,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                 return torch.cat(embl)
 
             conds = getcompelembs(all_prompts_cn)
-            unconds = getcompelembs(all_n_prompts_cn) if cn else getcompelembs(n_prompts)
+            unconds = getcompelembs(all_n_prompts_cn)
             embs = getcompelembs(prompts)
             n_embs = getcompelembs(n_prompts)
             prompt = negative_prompt = None
@@ -137,7 +144,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
             conds = self.encode_prompt(prompts, device, 1, True)[0]
             unconds = (
                 self.encode_prompt(n_prompts, device, 1, True)[0]
-                if cn
+                if equal
                 else self.encode_prompt(all_n_prompts_cn, device, 1, True)[0]
             )
             embs = n_embs = None
@@ -206,8 +213,11 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                     else:
                         px, nx = hidden_states.chunk(2)
 
-                    if cn:
-                        hidden_states = torch.cat([px for i in range(regions)] + [nx for i in range(regions)], 0)
+                    if equal:
+                        hidden_states = torch.cat(
+                            [px for i in range(regions)] + [nx for i in range(regions)],
+                            0,
+                        )
                         encoder_hidden_states = torch.cat([conds] + [unconds])
                     else:
                         hidden_states = torch.cat([px for i in range(regions)] + [nx], 0)
@@ -289,9 +299,9 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                     if any(x in mode for x in ["COL", "ROW"]):
                         reshaped = hidden_states.reshape(hidden_states.size()[0], h, w, hidden_states.size()[2])
                         center = reshaped.shape[0] // 2
-                        px = reshaped[0:center] if cn else reshaped[0:-batch]
-                        nx = reshaped[center:] if cn else reshaped[-batch:]
-                        outs = [px, nx] if cn else [px]
+                        px = reshaped[0:center] if equal else reshaped[0:-batch]
+                        nx = reshaped[center:] if equal else reshaped[-batch:]
+                        outs = [px, nx] if equal else [px]
                         for out in outs:
                             c = 0
                             for i, ocell in enumerate(ocells):
@@ -321,15 +331,16 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                                             :,
                                         ]
                                     c += 1
-                        px, nx = (px[0:batch], nx[0:batch]) if cn else (px[0:batch], nx)
+                        px, nx = (px[0:batch], nx[0:batch]) if equal else (px[0:batch], nx)
                         hidden_states = torch.cat([nx, px], 0) if revers else torch.cat([px, nx], 0)
                         hidden_states = hidden_states.reshape(xshape)
 
                     #### Regional Prompting Prompt mode
                     elif "PRO" in mode:
-                        center = reshaped.shape[0] // 2
-                        px = reshaped[0:center] if cn else reshaped[0:-batch]
-                        nx = reshaped[center:] if cn else reshaped[-batch:]
+                        px, nx = (
+                            torch.chunk(hidden_states) if equal else hidden_states[0:-batch],
+                            hidden_states[-batch:],
+                        )
 
                         if (h, w) in self.attnmasks and self.maskready:
 
@@ -340,8 +351,8 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                                         out[b] = out[b] + out[r * batch + b]
                                 return out
 
-                            px, nx = (mask(px), mask(nx)) if cn else (mask(px), nx)
-                        px, nx = (px[0:batch], nx[0:batch]) if cn else (px[0:batch], nx)
+                            px, nx = (mask(px), mask(nx)) if equal else (mask(px), nx)
+                        px, nx = (px[0:batch], nx[0:batch]) if equal else (px[0:batch], nx)
                         hidden_states = torch.cat([nx, px], 0) if revers else torch.cat([px, nx], 0)
                     return hidden_states
 
@@ -378,7 +389,15 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
             save_mask = False
 
         if mode == "PROMPT" and save_mask:
-            saveattnmaps(self, output, height, width, thresholds, num_inference_steps // 2, regions)
+            saveattnmaps(
+                self,
+                output,
+                height,
+                width,
+                thresholds,
+                num_inference_steps // 2,
+                regions,
+            )
 
         return output
 
@@ -437,7 +456,11 @@ def make_cells(ratios):
 def make_emblist(self, prompts):
     with torch.no_grad():
         tokens = self.tokenizer(
-            prompts, max_length=self.tokenizer.model_max_length, padding=True, truncation=True, return_tensors="pt"
+            prompts,
+            max_length=self.tokenizer.model_max_length,
+            padding=True,
+            truncation=True,
+            return_tensors="pt",
         ).input_ids.to(self.device)
         embs = self.text_encoder(tokens, output_hidden_states=True).last_hidden_state.to(self.device, dtype=self.dtype)
     return embs
@@ -563,7 +586,15 @@ def tokendealer(self, all_prompts):
 
 
 def scaled_dot_product_attention(
-    self, query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, scale=None, getattn=False
+    self,
+    query,
+    key,
+    value,
+    attn_mask=None,
+    dropout_p=0.0,
+    is_causal=False,
+    scale=None,
+    getattn=False,
 ) -> torch.Tensor:
     # Efficient implementation equivalent to the following:
     L, S = query.size(-2), key.size(-2)

examples/community/sde_drag.py

@@ -0,0 +1,594 @@
+import math
+import tempfile
+from typing import List, Optional
+
+import numpy as np
+import PIL.Image
+import torch
+from accelerate import Accelerator
+from torchvision import transforms
+from tqdm.auto import tqdm
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from diffusers import AutoencoderKL, DiffusionPipeline, DPMSolverMultistepScheduler, UNet2DConditionModel
+from diffusers.loaders import AttnProcsLayers, LoraLoaderMixin
+from diffusers.models.attention_processor import (
+    AttnAddedKVProcessor,
+    AttnAddedKVProcessor2_0,
+    LoRAAttnAddedKVProcessor,
+    LoRAAttnProcessor,
+    LoRAAttnProcessor2_0,
+    SlicedAttnAddedKVProcessor,
+)
+from diffusers.optimization import get_scheduler
+
+
+class SdeDragPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for image drag-and-drop editing using stochastic differential equations: https://arxiv.org/abs/2311.01410.
+    Please refer to the [official repository](https://github.com/ML-GSAI/SDE-Drag) for more information.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Please use
+            [`DDIMScheduler`].
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: DPMSolverMultistepScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: str,
+        image: PIL.Image.Image,
+        mask_image: PIL.Image.Image,
+        source_points: List[List[int]],
+        target_points: List[List[int]],
+        t0: Optional[float] = 0.6,
+        steps: Optional[int] = 200,
+        step_size: Optional[int] = 2,
+        image_scale: Optional[float] = 0.3,
+        adapt_radius: Optional[int] = 5,
+        min_lora_scale: Optional[float] = 0.5,
+        generator: Optional[torch.Generator] = None,
+    ):
+        r"""
+        Function invoked when calling the pipeline for image editing.
+        Args:
+            prompt (`str`, *required*):
+                The prompt to guide the image editing.
+            image (`PIL.Image.Image`, *required*):
+                Which will be edited, parts of the image will be masked out with `mask_image` and edited
+                according to `prompt`.
+            mask_image (`PIL.Image.Image`, *required*):
+                To mask `image`. White pixels in the mask will be edited, while black pixels will be preserved.
+            source_points (`List[List[int]]`, *required*):
+                Used to mark the starting positions of drag editing in the image, with each pixel represented as a
+                `List[int]` of length 2.
+            target_points (`List[List[int]]`, *required*):
+                Used to mark the target positions of drag editing in the image, with each pixel represented as a
+                `List[int]` of length 2.
+            t0 (`float`, *optional*, defaults to 0.6):
+                The time parameter. Higher t0 improves the fidelity while lowering the faithfulness of the edited images
+                and vice versa.
+            steps (`int`, *optional*, defaults to 200):
+                The number of sampling iterations.
+            step_size (`int`, *optional*, defaults to 2):
+                The drag diatance of each drag step.
+            image_scale (`float`, *optional*, defaults to 0.3):
+                To avoid duplicating the content, use image_scale to perturbs the source.
+            adapt_radius (`int`, *optional*, defaults to 5):
+                The size of the region for copy and paste operations during each step of the drag process.
+            min_lora_scale (`float`, *optional*, defaults to 0.5):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+                min_lora_scale specifies the minimum LoRA scale during the image drag-editing process.
+            generator ('torch.Generator', *optional*, defaults to None):
+                To make generation deterministic(https://pytorch.org/docs/stable/generated/torch.Generator.html).
+        Examples:
+        ```py
+        >>> import PIL
+        >>> import torch
+        >>> from diffusers import DDIMScheduler, DiffusionPipeline
+
+        >>> # Load the pipeline
+        >>> model_path = "runwayml/stable-diffusion-v1-5"
+        >>> scheduler = DDIMScheduler.from_pretrained(model_path, subfolder="scheduler")
+        >>> pipe = DiffusionPipeline.from_pretrained(model_path, scheduler=scheduler, custom_pipeline="sde_drag")
+        >>> pipe.to('cuda')
+
+        >>> # To save GPU memory, torch.float16 can be used, but it may compromise image quality.
+        >>> # If not training LoRA, please avoid using torch.float16
+        >>> # pipe.to(torch.float16)
+
+        >>> # Provide prompt, image, mask image, and the starting and target points for drag editing.
+        >>> prompt = "prompt of the image"
+        >>> image = PIL.Image.open('/path/to/image')
+        >>> mask_image = PIL.Image.open('/path/to/mask_image')
+        >>> source_points = [[123, 456]]
+        >>> target_points = [[234, 567]]
+
+        >>> # train_lora is optional, and in most cases, using train_lora can better preserve consistency with the original image.
+        >>> pipe.train_lora(prompt, image)
+
+        >>> output = pipe(prompt, image, mask_image, source_points, target_points)
+        >>> output_image = PIL.Image.fromarray(output)
+        >>> output_image.save("./output.png")
+        ```
+        """
+
+        self.scheduler.set_timesteps(steps)
+
+        noise_scale = (1 - image_scale**2) ** (0.5)
+
+        text_embeddings = self._get_text_embed(prompt)
+        uncond_embeddings = self._get_text_embed([""])
+        text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        latent = self._get_img_latent(image)
+
+        mask = mask_image.resize((latent.shape[3], latent.shape[2]))
+        mask = torch.tensor(np.array(mask))
+        mask = mask.unsqueeze(0).expand_as(latent).to(self.device)
+
+        source_points = torch.tensor(source_points).div(torch.tensor([8]), rounding_mode="trunc")
+        target_points = torch.tensor(target_points).div(torch.tensor([8]), rounding_mode="trunc")
+
+        distance = target_points - source_points
+        distance_norm_max = torch.norm(distance.float(), dim=1, keepdim=True).max()
+
+        if distance_norm_max <= step_size:
+            drag_num = 1
+        else:
+            drag_num = distance_norm_max.div(torch.tensor([step_size]), rounding_mode="trunc")
+            if (distance_norm_max / drag_num - step_size).abs() > (
+                distance_norm_max / (drag_num + 1) - step_size
+            ).abs():
+                drag_num += 1
+
+        latents = []
+        for i in tqdm(range(int(drag_num)), desc="SDE Drag"):
+            source_new = source_points + (i / drag_num * distance).to(torch.int)
+            target_new = source_points + ((i + 1) / drag_num * distance).to(torch.int)
+
+            latent, noises, hook_latents, lora_scales, cfg_scales = self._forward(
+                latent, steps, t0, min_lora_scale, text_embeddings, generator
+            )
+            latent = self._copy_and_paste(
+                latent,
+                source_new,
+                target_new,
+                adapt_radius,
+                latent.shape[2] - 1,
+                latent.shape[3] - 1,
+                image_scale,
+                noise_scale,
+                generator,
+            )
+            latent = self._backward(
+                latent, mask, steps, t0, noises, hook_latents, lora_scales, cfg_scales, text_embeddings, generator
+            )
+
+            latents.append(latent)
+
+        result_image = 1 / 0.18215 * latents[-1]
+
+        with torch.no_grad():
+            result_image = self.vae.decode(result_image).sample
+
+        result_image = (result_image / 2 + 0.5).clamp(0, 1)
+        result_image = result_image.cpu().permute(0, 2, 3, 1).numpy()[0]
+        result_image = (result_image * 255).astype(np.uint8)
+
+        return result_image
+
+    def train_lora(self, prompt, image, lora_step=100, lora_rank=16, generator=None):
+        accelerator = Accelerator(gradient_accumulation_steps=1, mixed_precision="fp16")
+
+        self.vae.requires_grad_(False)
+        self.text_encoder.requires_grad_(False)
+        self.unet.requires_grad_(False)
+
+        unet_lora_attn_procs = {}
+        for name, attn_processor in self.unet.attn_processors.items():
+            cross_attention_dim = None if name.endswith("attn1.processor") else self.unet.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = self.unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(self.unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = self.unet.config.block_out_channels[block_id]
+            else:
+                raise NotImplementedError("name must start with up_blocks, mid_blocks, or down_blocks")
+
+            if isinstance(attn_processor, (AttnAddedKVProcessor, SlicedAttnAddedKVProcessor, AttnAddedKVProcessor2_0)):
+                lora_attn_processor_class = LoRAAttnAddedKVProcessor
+            else:
+                lora_attn_processor_class = (
+                    LoRAAttnProcessor2_0
+                    if hasattr(torch.nn.functional, "scaled_dot_product_attention")
+                    else LoRAAttnProcessor
+                )
+            unet_lora_attn_procs[name] = lora_attn_processor_class(
+                hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, rank=lora_rank
+            )
+
+        self.unet.set_attn_processor(unet_lora_attn_procs)
+        unet_lora_layers = AttnProcsLayers(self.unet.attn_processors)
+        params_to_optimize = unet_lora_layers.parameters()
+
+        optimizer = torch.optim.AdamW(
+            params_to_optimize,
+            lr=2e-4,
+            betas=(0.9, 0.999),
+            weight_decay=1e-2,
+            eps=1e-08,
+        )
+
+        lr_scheduler = get_scheduler(
+            "constant",
+            optimizer=optimizer,
+            num_warmup_steps=0,
+            num_training_steps=lora_step,
+            num_cycles=1,
+            power=1.0,
+        )
+
+        unet_lora_layers = accelerator.prepare_model(unet_lora_layers)
+        optimizer = accelerator.prepare_optimizer(optimizer)
+        lr_scheduler = accelerator.prepare_scheduler(lr_scheduler)
+
+        with torch.no_grad():
+            text_inputs = self._tokenize_prompt(prompt, tokenizer_max_length=None)
+            text_embedding = self._encode_prompt(
+                text_inputs.input_ids, text_inputs.attention_mask, text_encoder_use_attention_mask=False
+            )
+
+        image_transforms = transforms.Compose(
+            [
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+
+        image = image_transforms(image).to(self.device, dtype=self.vae.dtype)
+        image = image.unsqueeze(dim=0)
+        latents_dist = self.vae.encode(image).latent_dist
+
+        for _ in tqdm(range(lora_step), desc="Train LoRA"):
+            self.unet.train()
+            model_input = latents_dist.sample() * self.vae.config.scaling_factor
+
+            # Sample noise that we'll add to the latents
+            noise = torch.randn(
+                model_input.size(),
+                dtype=model_input.dtype,
+                layout=model_input.layout,
+                device=model_input.device,
+                generator=generator,
+            )
+            bsz, channels, height, width = model_input.shape
+
+            # Sample a random timestep for each image
+            timesteps = torch.randint(
+                0, self.scheduler.config.num_train_timesteps, (bsz,), device=model_input.device, generator=generator
+            )
+            timesteps = timesteps.long()
+
+            # Add noise to the model input according to the noise magnitude at each timestep
+            # (this is the forward diffusion process)
+            noisy_model_input = self.scheduler.add_noise(model_input, noise, timesteps)
+
+            # Predict the noise residual
+            model_pred = self.unet(noisy_model_input, timesteps, text_embedding).sample
+
+            # Get the target for loss depending on the prediction type
+            if self.scheduler.config.prediction_type == "epsilon":
+                target = noise
+            elif self.scheduler.config.prediction_type == "v_prediction":
+                target = self.scheduler.get_velocity(model_input, noise, timesteps)
+            else:
+                raise ValueError(f"Unknown prediction type {self.scheduler.config.prediction_type}")
+
+            loss = torch.nn.functional.mse_loss(model_pred.float(), target.float(), reduction="mean")
+            accelerator.backward(loss)
+            optimizer.step()
+            lr_scheduler.step()
+            optimizer.zero_grad()
+
+        with tempfile.TemporaryDirectory() as save_lora_dir:
+            LoraLoaderMixin.save_lora_weights(
+                save_directory=save_lora_dir,
+                unet_lora_layers=unet_lora_layers,
+                text_encoder_lora_layers=None,
+            )
+
+            self.unet.load_attn_procs(save_lora_dir)
+
+    def _tokenize_prompt(self, prompt, tokenizer_max_length=None):
+        if tokenizer_max_length is not None:
+            max_length = tokenizer_max_length
+        else:
+            max_length = self.tokenizer.model_max_length
+
+        text_inputs = self.tokenizer(
+            prompt,
+            truncation=True,
+            padding="max_length",
+            max_length=max_length,
+            return_tensors="pt",
+        )
+
+        return text_inputs
+
+    def _encode_prompt(self, input_ids, attention_mask, text_encoder_use_attention_mask=False):
+        text_input_ids = input_ids.to(self.device)
+
+        if text_encoder_use_attention_mask:
+            attention_mask = attention_mask.to(self.device)
+        else:
+            attention_mask = None
+
+        prompt_embeds = self.text_encoder(
+            text_input_ids,
+            attention_mask=attention_mask,
+        )
+        prompt_embeds = prompt_embeds[0]
+
+        return prompt_embeds
+
+    @torch.no_grad()
+    def _get_text_embed(self, prompt):
+        text_input = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]
+        return text_embeddings
+
+    def _copy_and_paste(
+        self, latent, source_new, target_new, adapt_radius, max_height, max_width, image_scale, noise_scale, generator
+    ):
+        def adaption_r(source, target, adapt_radius, max_height, max_width):
+            r_x_lower = min(adapt_radius, source[0], target[0])
+            r_x_upper = min(adapt_radius, max_width - source[0], max_width - target[0])
+            r_y_lower = min(adapt_radius, source[1], target[1])
+            r_y_upper = min(adapt_radius, max_height - source[1], max_height - target[1])
+            return r_x_lower, r_x_upper, r_y_lower, r_y_upper
+
+        for source_, target_ in zip(source_new, target_new):
+            r_x_lower, r_x_upper, r_y_lower, r_y_upper = adaption_r(
+                source_, target_, adapt_radius, max_height, max_width
+            )
+
+            source_feature = latent[
+                :, :, source_[1] - r_y_lower : source_[1] + r_y_upper, source_[0] - r_x_lower : source_[0] + r_x_upper
+            ].clone()
+
+            latent[
+                :, :, source_[1] - r_y_lower : source_[1] + r_y_upper, source_[0] - r_x_lower : source_[0] + r_x_upper
+            ] = image_scale * source_feature + noise_scale * torch.randn(
+                latent.shape[0],
+                4,
+                r_y_lower + r_y_upper,
+                r_x_lower + r_x_upper,
+                device=self.device,
+                generator=generator,
+            )
+
+            latent[
+                :, :, target_[1] - r_y_lower : target_[1] + r_y_upper, target_[0] - r_x_lower : target_[0] + r_x_upper
+            ] = source_feature * 1.1
+        return latent
+
+    @torch.no_grad()
+    def _get_img_latent(self, image, height=None, weight=None):
+        data = image.convert("RGB")
+        if height is not None:
+            data = data.resize((weight, height))
+        transform = transforms.ToTensor()
+        data = transform(data).unsqueeze(0)
+        data = (data * 2.0) - 1.0
+        data = data.to(self.device, dtype=self.vae.dtype)
+        latent = self.vae.encode(data).latent_dist.sample()
+        latent = 0.18215 * latent
+        return latent
+
+    @torch.no_grad()
+    def _get_eps(self, latent, timestep, guidance_scale, text_embeddings, lora_scale=None):
+        latent_model_input = torch.cat([latent] * 2) if guidance_scale > 1.0 else latent
+        text_embeddings = text_embeddings if guidance_scale > 1.0 else text_embeddings.chunk(2)[1]
+
+        cross_attention_kwargs = None if lora_scale is None else {"scale": lora_scale}
+
+        with torch.no_grad():
+            noise_pred = self.unet(
+                latent_model_input,
+                timestep,
+                encoder_hidden_states=text_embeddings,
+                cross_attention_kwargs=cross_attention_kwargs,
+            ).sample
+
+        if guidance_scale > 1.0:
+            noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+        elif guidance_scale == 1.0:
+            noise_pred_text = noise_pred
+            noise_pred_uncond = 0.0
+        else:
+            raise NotImplementedError(guidance_scale)
+        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+        return noise_pred
+
+    def _forward_sde(
+        self, timestep, sample, guidance_scale, text_embeddings, steps, eta=1.0, lora_scale=None, generator=None
+    ):
+        num_train_timesteps = len(self.scheduler)
+        alphas_cumprod = self.scheduler.alphas_cumprod
+        initial_alpha_cumprod = torch.tensor(1.0)
+
+        prev_timestep = timestep + num_train_timesteps // steps
+
+        alpha_prod_t = alphas_cumprod[timestep] if timestep >= 0 else initial_alpha_cumprod
+        alpha_prod_t_prev = alphas_cumprod[prev_timestep]
+
+        beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+        x_prev = (alpha_prod_t_prev / alpha_prod_t) ** (0.5) * sample + (1 - alpha_prod_t_prev / alpha_prod_t) ** (
+            0.5
+        ) * torch.randn(
+            sample.size(), dtype=sample.dtype, layout=sample.layout, device=self.device, generator=generator
+        )
+        eps = self._get_eps(x_prev, prev_timestep, guidance_scale, text_embeddings, lora_scale)
+
+        sigma_t_prev = (
+            eta
+            * (1 - alpha_prod_t) ** (0.5)
+            * (1 - alpha_prod_t_prev / (1 - alpha_prod_t_prev) * (1 - alpha_prod_t) / alpha_prod_t) ** (0.5)
+        )
+
+        pred_original_sample = (x_prev - beta_prod_t_prev ** (0.5) * eps) / alpha_prod_t_prev ** (0.5)
+        pred_sample_direction_coeff = (1 - alpha_prod_t - sigma_t_prev**2) ** (0.5)
+
+        noise = (
+            sample - alpha_prod_t ** (0.5) * pred_original_sample - pred_sample_direction_coeff * eps
+        ) / sigma_t_prev
+
+        return x_prev, noise
+
+    def _sample(
+        self,
+        timestep,
+        sample,
+        guidance_scale,
+        text_embeddings,
+        steps,
+        sde=False,
+        noise=None,
+        eta=1.0,
+        lora_scale=None,
+        generator=None,
+    ):
+        num_train_timesteps = len(self.scheduler)
+        alphas_cumprod = self.scheduler.alphas_cumprod
+        final_alpha_cumprod = torch.tensor(1.0)
+
+        eps = self._get_eps(sample, timestep, guidance_scale, text_embeddings, lora_scale)
+
+        prev_timestep = timestep - num_train_timesteps // steps
+
+        alpha_prod_t = alphas_cumprod[timestep]
+        alpha_prod_t_prev = alphas_cumprod[prev_timestep] if prev_timestep >= 0 else final_alpha_cumprod
+
+        beta_prod_t = 1 - alpha_prod_t
+
+        sigma_t = (
+            eta
+            * ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) ** (0.5)
+            * (1 - alpha_prod_t / alpha_prod_t_prev) ** (0.5)
+            if sde
+            else 0
+        )
+
+        pred_original_sample = (sample - beta_prod_t ** (0.5) * eps) / alpha_prod_t ** (0.5)
+        pred_sample_direction_coeff = (1 - alpha_prod_t_prev - sigma_t**2) ** (0.5)
+
+        noise = (
+            torch.randn(
+                sample.size(), dtype=sample.dtype, layout=sample.layout, device=self.device, generator=generator
+            )
+            if noise is None
+            else noise
+        )
+        latent = (
+            alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction_coeff * eps + sigma_t * noise
+        )
+
+        return latent
+
+    def _forward(self, latent, steps, t0, lora_scale_min, text_embeddings, generator):
+        def scale_schedule(begin, end, n, length, type="linear"):
+            if type == "constant":
+                return end
+            elif type == "linear":
+                return begin + (end - begin) * n / length
+            elif type == "cos":
+                factor = (1 - math.cos(n * math.pi / length)) / 2
+                return (1 - factor) * begin + factor * end
+            else:
+                raise NotImplementedError(type)
+
+        noises = []
+        latents = []
+        lora_scales = []
+        cfg_scales = []
+        latents.append(latent)
+        t0 = int(t0 * steps)
+        t_begin = steps - t0
+
+        length = len(self.scheduler.timesteps[t_begin - 1 : -1]) - 1
+        index = 1
+        for t in self.scheduler.timesteps[t_begin:].flip(dims=[0]):
+            lora_scale = scale_schedule(1, lora_scale_min, index, length, type="cos")
+            cfg_scale = scale_schedule(1, 3.0, index, length, type="linear")
+            latent, noise = self._forward_sde(
+                t, latent, cfg_scale, text_embeddings, steps, lora_scale=lora_scale, generator=generator
+            )
+
+            noises.append(noise)
+            latents.append(latent)
+            lora_scales.append(lora_scale)
+            cfg_scales.append(cfg_scale)
+            index += 1
+        return latent, noises, latents, lora_scales, cfg_scales
+
+    def _backward(
+        self, latent, mask, steps, t0, noises, hook_latents, lora_scales, cfg_scales, text_embeddings, generator
+    ):
+        t0 = int(t0 * steps)
+        t_begin = steps - t0
+
+        hook_latent = hook_latents.pop()
+        latent = torch.where(mask > 128, latent, hook_latent)
+        for t in self.scheduler.timesteps[t_begin - 1 : -1]:
+            latent = self._sample(
+                t,
+                latent,
+                cfg_scales.pop(),
+                text_embeddings,
+                steps,
+                sde=True,
+                noise=noises.pop(),
+                lora_scale=lora_scales.pop(),
+                generator=generator,
+            )
+            hook_latent = hook_latents.pop()
+            latent = torch.where(mask > 128, latent, hook_latent)
+        return latent

examples/consistency_distillation/README.md

@@ -1,6 +1,6 @@
 # Latent Consistency Distillation Example:
 
-[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is method to distill latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use the latent consistency distillation to distill stable-diffusion-v1.5 for less timestep inference.
+[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is a method to distill a latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use latent consistency distillation to distill stable-diffusion-v1.5 for inference with few timesteps.
 
 ## Full model distillation
 
@@ -24,7 +24,7 @@ Then cd in the example folder and run
 pip install -r requirements.txt
 ```
 
-And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+And initialize an [🤗 Accelerate](https://github.com/huggingface/accelerate/) environment with:
 
 ```bash
 accelerate config
@@ -46,12 +46,16 @@ write_basic_config()
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
 
 
-#### Example with LAION-A6+ dataset
+#### Example
+
+The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example, and for illustrative purposes only. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/). You may also need to search the hyperparameter space according to the dataset you use.
 
 ```bash
-runwayml/stable-diffusion-v1-5
-PROGRAM="train_lcm_distill_sd_wds.py \
-    --pretrained_teacher_model=$MODEL_DIR \
+export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+export OUTPUT_DIR="path/to/saved/model"
+
+accelerate launch train_lcm_distill_sd_wds.py \
+    --pretrained_teacher_model=$MODEL_NAME \
     --output_dir=$OUTPUT_DIR \
     --mixed_precision=fp16 \
     --resolution=512 \
@@ -59,7 +63,7 @@ PROGRAM="train_lcm_distill_sd_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
+    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \
@@ -69,19 +73,23 @@ PROGRAM="train_lcm_distill_sd_wds.py \
     --resume_from_checkpoint=latest \
     --report_to=wandb \
     --seed=453645634 \
-    --push_to_hub \
+    --push_to_hub
 ```
 
 ## LCM-LoRA
 
 Instead of fine-tuning the full model, we can also just train a LoRA that can be injected into any SDXL model.
 
-### Example with LAION-A6+ dataset
-    
+### Example
+
+The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/).
+
 ```bash
-runwayml/stable-diffusion-v1-5
-PROGRAM="train_lcm_distill_lora_sd_wds.py \
-    --pretrained_teacher_model=$MODEL_DIR \
+export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+export OUTPUT_DIR="path/to/saved/model"
+
+accelerate launch train_lcm_distill_lora_sd_wds.py \
+    --pretrained_teacher_model=$MODEL_NAME \
     --output_dir=$OUTPUT_DIR \
     --mixed_precision=fp16 \
     --resolution=512 \
@@ -90,7 +98,7 @@ PROGRAM="train_lcm_distill_lora_sd_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
+    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \

examples/consistency_distillation/README_sdxl.md

@@ -1,6 +1,6 @@
 # Latent Consistency Distillation Example:
 
-[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is method to distill latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use the latent consistency distillation to distill SDXL for less timestep inference.
+[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is a method to distill a latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use latent consistency distillation to distill SDXL for inference with few timesteps.
 
 ## Full model distillation
 
@@ -24,7 +24,7 @@ Then cd in the example folder and run
 pip install -r requirements.txt
 ```
 
-And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+And initialize an [🤗 Accelerate](https://github.com/huggingface/accelerate/) environment with:
 
 ```bash
 accelerate config
@@ -46,12 +46,16 @@ write_basic_config()
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
 
 
-#### Example with LAION-A6+ dataset
+#### Example
+
+The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example, and for illustrative purposes only. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/). You may also need to search the hyperparameter space according to the dataset you use.
 
 ```bash
-export MODEL_DIR="stabilityai/stable-diffusion-xl-base-1.0"
-PROGRAM="train_lcm_distill_sdxl_wds.py \
-    --pretrained_teacher_model=$MODEL_DIR \
+export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0"
+export OUTPUT_DIR="path/to/saved/model"
+
+accelerate launch train_lcm_distill_sdxl_wds.py \
+    --pretrained_teacher_model=$MODEL_NAME \
     --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
     --output_dir=$OUTPUT_DIR \
     --mixed_precision=fp16 \
@@ -60,7 +64,7 @@ PROGRAM="train_lcm_distill_sdxl_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
+    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \
@@ -77,11 +81,15 @@ PROGRAM="train_lcm_distill_sdxl_wds.py \
 
 Instead of fine-tuning the full model, we can also just train a LoRA that can be injected into any SDXL model.
 
-### Example with LAION-A6+ dataset
-    
+### Example
+
+The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/).
+
 ```bash
-export MODEL_DIR="stabilityai/stable-diffusion-xl-base-1.0"
-PROGRAM="train_lcm_distill_lora_sdxl_wds.py \
+export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0"
+export OUTPUT_DIR="path/to/saved/model"
+
+accelerate launch train_lcm_distill_lora_sdxl_wds.py \
     --pretrained_teacher_model=$MODEL_DIR \
     --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
     --output_dir=$OUTPUT_DIR \
@@ -92,7 +100,7 @@ PROGRAM="train_lcm_distill_lora_sdxl_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
+    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \

examples/consistency_distillation/train_lcm_distill_lora_sd_wds.py

@@ -156,7 +156,7 @@ class WebdatasetFilter:
             return False
 
 
-class Text2ImageDataset:
+class SDText2ImageDataset:
     def __init__(
         self,
         train_shards_path_or_url: Union[str, List[str]],
@@ -359,19 +359,43 @@ def scalings_for_boundary_conditions(timestep, sigma_data=0.5, timestep_scaling=
 
 
 # Compare LCMScheduler.step, Step 4
-def predicted_origin(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+def get_predicted_original_sample(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+    alphas = extract_into_tensor(alphas, timesteps, sample.shape)
+    sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
     if prediction_type == "epsilon":
-        sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
-        alphas = extract_into_tensor(alphas, timesteps, sample.shape)
         pred_x_0 = (sample - sigmas * model_output) / alphas
+    elif prediction_type == "sample":
+        pred_x_0 = model_output
     elif prediction_type == "v_prediction":
-        pred_x_0 = alphas[timesteps] * sample - sigmas[timesteps] * model_output
+        pred_x_0 = alphas * sample - sigmas * model_output
     else:
-        raise ValueError(f"Prediction type {prediction_type} currently not supported.")
+        raise ValueError(
+            f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`"
+            f" are supported."
+        )
 
     return pred_x_0
 
 
+# Based on step 4 in DDIMScheduler.step
+def get_predicted_noise(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+    alphas = extract_into_tensor(alphas, timesteps, sample.shape)
+    sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
+    if prediction_type == "epsilon":
+        pred_epsilon = model_output
+    elif prediction_type == "sample":
+        pred_epsilon = (sample - alphas * model_output) / sigmas
+    elif prediction_type == "v_prediction":
+        pred_epsilon = alphas * model_output + sigmas * sample
+    else:
+        raise ValueError(
+            f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`"
+            f" are supported."
+        )
+
+    return pred_epsilon
+
+
 def extract_into_tensor(a, t, x_shape):
     b, *_ = t.shape
     out = a.gather(-1, t)
@@ -835,34 +859,35 @@ def main(args):
         args.pretrained_teacher_model, subfolder="scheduler", revision=args.teacher_revision
     )
 
-    # The scheduler calculates the alpha and sigma schedule for us
+    # DDPMScheduler calculates the alpha and sigma noise schedules (based on the alpha bars) for us
     alpha_schedule = torch.sqrt(noise_scheduler.alphas_cumprod)
     sigma_schedule = torch.sqrt(1 - noise_scheduler.alphas_cumprod)
+    # Initialize the DDIM ODE solver for distillation.
     solver = DDIMSolver(
         noise_scheduler.alphas_cumprod.numpy(),
         timesteps=noise_scheduler.config.num_train_timesteps,
         ddim_timesteps=args.num_ddim_timesteps,
     )
 
-    # 2. Load tokenizers from SD-XL checkpoint.
+    # 2. Load tokenizers from SD 1.X/2.X checkpoint.
     tokenizer = AutoTokenizer.from_pretrained(
         args.pretrained_teacher_model, subfolder="tokenizer", revision=args.teacher_revision, use_fast=False
     )
 
-    # 3. Load text encoders from SD-1.5 checkpoint.
+    # 3. Load text encoders from SD 1.X/2.X checkpoint.
     # import correct text encoder classes
     text_encoder = CLIPTextModel.from_pretrained(
         args.pretrained_teacher_model, subfolder="text_encoder", revision=args.teacher_revision
     )
 
-    # 4. Load VAE from SD-XL checkpoint (or more stable VAE)
+    # 4. Load VAE from SD 1.X/2.X checkpoint
     vae = AutoencoderKL.from_pretrained(
         args.pretrained_teacher_model,
         subfolder="vae",
         revision=args.teacher_revision,
     )
 
-    # 5. Load teacher U-Net from SD-XL checkpoint
+    # 5. Load teacher U-Net from SD 1.X/2.X checkpoint
     teacher_unet = UNet2DConditionModel.from_pretrained(
         args.pretrained_teacher_model, subfolder="unet", revision=args.teacher_revision
     )
@@ -872,7 +897,7 @@ def main(args):
     text_encoder.requires_grad_(False)
     teacher_unet.requires_grad_(False)
 
-    # 7. Create online (`unet`) student U-Nets.
+    # 7. Create online student U-Net.
     unet = UNet2DConditionModel.from_pretrained(
         args.pretrained_teacher_model, subfolder="unet", revision=args.teacher_revision
     )
@@ -935,6 +960,7 @@ def main(args):
     # Also move the alpha and sigma noise schedules to accelerator.device.
     alpha_schedule = alpha_schedule.to(accelerator.device)
     sigma_schedule = sigma_schedule.to(accelerator.device)
+    # Move the ODE solver to accelerator.device.
     solver = solver.to(accelerator.device)
 
     # 10. Handle saving and loading of checkpoints
@@ -1011,13 +1037,14 @@ def main(args):
         eps=args.adam_epsilon,
     )
 
+    # 13. Dataset creation and data processing
     # Here, we compute not just the text embeddings but also the additional embeddings
     # needed for the SD XL UNet to operate.
     def compute_embeddings(prompt_batch, proportion_empty_prompts, text_encoder, tokenizer, is_train=True):
         prompt_embeds = encode_prompt(prompt_batch, text_encoder, tokenizer, proportion_empty_prompts, is_train)
         return {"prompt_embeds": prompt_embeds}
 
-    dataset = Text2ImageDataset(
+    dataset = SDText2ImageDataset(
         train_shards_path_or_url=args.train_shards_path_or_url,
         num_train_examples=args.max_train_samples,
         per_gpu_batch_size=args.train_batch_size,
@@ -1037,6 +1064,7 @@ def main(args):
         tokenizer=tokenizer,
     )
 
+    # 14. LR Scheduler creation
     # Scheduler and math around the number of training steps.
     overrode_max_train_steps = False
     num_update_steps_per_epoch = math.ceil(train_dataloader.num_batches / args.gradient_accumulation_steps)
@@ -1051,6 +1079,7 @@ def main(args):
         num_training_steps=args.max_train_steps,
     )
 
+    # 15. Prepare for training
     # Prepare everything with our `accelerator`.
     unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
 
@@ -1072,7 +1101,7 @@ def main(args):
     ).input_ids.to(accelerator.device)
     uncond_prompt_embeds = text_encoder(uncond_input_ids)[0]
 
-    # Train!
+    # 16. Train!
     total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
 
     logger.info("***** Running training *****")
@@ -1123,7 +1152,8 @@ def main(args):
     for epoch in range(first_epoch, args.num_train_epochs):
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
-                image, text, _, _ = batch
+                # 1. Load and process the image and text conditioning
+                image, text = batch
 
                 image = image.to(accelerator.device, non_blocking=True)
                 encoded_text = compute_embeddings_fn(text)
@@ -1140,37 +1170,37 @@ def main(args):
 
                 latents = latents * vae.config.scaling_factor
                 latents = latents.to(weight_dtype)
-
-                # Sample noise that we'll add to the latents
-                noise = torch.randn_like(latents)
                 bsz = latents.shape[0]
 
-                # Sample a random timestep for each image t_n ~ U[0, N - k - 1] without bias.
+                # 2. Sample a random timestep for each image t_n from the ODE solver timesteps without bias.
+                # For the DDIM solver, the timestep schedule is [T - 1, T - k - 1, T - 2 * k - 1, ...]
                 topk = noise_scheduler.config.num_train_timesteps // args.num_ddim_timesteps
                 index = torch.randint(0, args.num_ddim_timesteps, (bsz,), device=latents.device).long()
                 start_timesteps = solver.ddim_timesteps[index]
                 timesteps = start_timesteps - topk
                 timesteps = torch.where(timesteps < 0, torch.zeros_like(timesteps), timesteps)
 
-                # 20.4.4. Get boundary scalings for start_timesteps and (end) timesteps.
+                # 3. Get boundary scalings for start_timesteps and (end) timesteps.
                 c_skip_start, c_out_start = scalings_for_boundary_conditions(start_timesteps)
                 c_skip_start, c_out_start = [append_dims(x, latents.ndim) for x in [c_skip_start, c_out_start]]
                 c_skip, c_out = scalings_for_boundary_conditions(timesteps)
                 c_skip, c_out = [append_dims(x, latents.ndim) for x in [c_skip, c_out]]
 
-                # 20.4.5. Add noise to the latents according to the noise magnitude at each timestep
-                # (this is the forward diffusion process) [z_{t_{n + k}} in Algorithm 1]
+                # 4. Sample noise from the prior and add it to the latents according to the noise magnitude at each
+                # timestep (this is the forward diffusion process) [z_{t_{n + k}} in Algorithm 1]
+                noise = torch.randn_like(latents)
                 noisy_model_input = noise_scheduler.add_noise(latents, noise, start_timesteps)
 
-                # 20.4.6. Sample a random guidance scale w from U[w_min, w_max] and embed it
+                # 5. Sample a random guidance scale w from U[w_min, w_max]
+                # Note that for LCM-LoRA distillation it is not necessary to use a guidance scale embedding
                 w = (args.w_max - args.w_min) * torch.rand((bsz,)) + args.w_min
                 w = w.reshape(bsz, 1, 1, 1)
                 w = w.to(device=latents.device, dtype=latents.dtype)
 
-                # 20.4.8. Prepare prompt embeds and unet_added_conditions
+                # 6. Prepare prompt embeds and unet_added_conditions
                 prompt_embeds = encoded_text.pop("prompt_embeds")
 
-                # 20.4.9. Get online LCM prediction on z_{t_{n + k}}, w, c, t_{n + k}
+                # 7. Get online LCM prediction on z_{t_{n + k}} (noisy_model_input), w, c, t_{n + k} (start_timesteps)
                 noise_pred = unet(
                     noisy_model_input,
                     start_timesteps,
@@ -1179,7 +1209,7 @@ def main(args):
                     added_cond_kwargs=encoded_text,
                 ).sample
 
-                pred_x_0 = predicted_origin(
+                pred_x_0 = get_predicted_original_sample(
                     noise_pred,
                     start_timesteps,
                     noisy_model_input,
@@ -1190,17 +1220,27 @@ def main(args):
 
                 model_pred = c_skip_start * noisy_model_input + c_out_start * pred_x_0
 
-                # 20.4.10. Use the ODE solver to predict the kth step in the augmented PF-ODE trajectory after
-                # noisy_latents with both the conditioning embedding c and unconditional embedding 0
-                # Get teacher model prediction on noisy_latents and conditional embedding
+                # 8. Compute the conditional and unconditional teacher model predictions to get CFG estimates of the
+                # predicted noise eps_0 and predicted original sample x_0, then run the ODE solver using these
+                # estimates to predict the data point in the augmented PF-ODE trajectory corresponding to the next ODE
+                # solver timestep.
                 with torch.no_grad():
                     with torch.autocast("cuda"):
+                        # 1. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and conditional embedding c
                         cond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
                             start_timesteps,
                             encoder_hidden_states=prompt_embeds.to(weight_dtype),
                         ).sample
-                        cond_pred_x0 = predicted_origin(
+                        cond_pred_x0 = get_predicted_original_sample(
+                            cond_teacher_output,
+                            start_timesteps,
+                            noisy_model_input,
+                            noise_scheduler.config.prediction_type,
+                            alpha_schedule,
+                            sigma_schedule,
+                        )
+                        cond_pred_noise = get_predicted_noise(
                             cond_teacher_output,
                             start_timesteps,
                             noisy_model_input,
@@ -1209,13 +1249,21 @@ def main(args):
                             sigma_schedule,
                         )
 
-                        # Get teacher model prediction on noisy_latents and unconditional embedding
+                        # 2. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and unconditional embedding 0
                         uncond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
                             start_timesteps,
                             encoder_hidden_states=uncond_prompt_embeds.to(weight_dtype),
                         ).sample
-                        uncond_pred_x0 = predicted_origin(
+                        uncond_pred_x0 = get_predicted_original_sample(
+                            uncond_teacher_output,
+                            start_timesteps,
+                            noisy_model_input,
+                            noise_scheduler.config.prediction_type,
+                            alpha_schedule,
+                            sigma_schedule,
+                        )
+                        uncond_pred_noise = get_predicted_noise(
                             uncond_teacher_output,
                             start_timesteps,
                             noisy_model_input,
@@ -1224,12 +1272,17 @@ def main(args):
                             sigma_schedule,
                         )
 
-                        # 20.4.11. Perform "CFG" to get x_prev estimate (using the LCM paper's CFG formulation)
+                        # 3. Calculate the CFG estimate of x_0 (pred_x0) and eps_0 (pred_noise)
+                        # Note that this uses the LCM paper's CFG formulation rather than the Imagen CFG formulation
                         pred_x0 = cond_pred_x0 + w * (cond_pred_x0 - uncond_pred_x0)
-                        pred_noise = cond_teacher_output + w * (cond_teacher_output - uncond_teacher_output)
+                        pred_noise = cond_pred_noise + w * (cond_pred_noise - uncond_pred_noise)
+                        # 4. Run one step of the ODE solver to estimate the next point x_prev on the
+                        # augmented PF-ODE trajectory (solving backward in time)
+                        # Note that the DDIM step depends on both the predicted x_0 and source noise eps_0.
                         x_prev = solver.ddim_step(pred_x0, pred_noise, index)
 
-                # 20.4.12. Get target LCM prediction on x_prev, w, c, t_n
+                # 9. Get target LCM prediction on x_prev, w, c, t_n (timesteps)
+                # Note that we do not use a separate target network for LCM-LoRA distillation.
                 with torch.no_grad():
                     with torch.autocast("cuda", dtype=weight_dtype):
                         target_noise_pred = unet(
@@ -1238,7 +1291,7 @@ def main(args):
                             timestep_cond=None,
                             encoder_hidden_states=prompt_embeds.float(),
                         ).sample
-                    pred_x_0 = predicted_origin(
+                    pred_x_0 = get_predicted_original_sample(
                         target_noise_pred,
                         timesteps,
                         x_prev,
@@ -1248,7 +1301,7 @@ def main(args):
                     )
                     target = c_skip * x_prev + c_out * pred_x_0
 
-                # 20.4.13. Calculate loss
+                # 10. Calculate loss
                 if args.loss_type == "l2":
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
                 elif args.loss_type == "huber":
@@ -1256,7 +1309,7 @@ def main(args):
                         torch.sqrt((model_pred.float() - target.float()) ** 2 + args.huber_c**2) - args.huber_c
                     )
 
-                # 20.4.14. Backpropagate on the online student model (`unet`)
+                # 11. Backpropagate on the online student model (`unet`)
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
                     accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)

examples/consistency_distillation/train_lcm_distill_lora_sdxl_wds.py

@@ -68,6 +68,11 @@ from diffusers.utils.import_utils import is_xformers_available
 
 MAX_SEQ_LENGTH = 77
 
+# Adjust for your dataset
+WDS_JSON_WIDTH = "width"  # original_width for LAION
+WDS_JSON_HEIGHT = "height"  # original_height for LAION
+MIN_SIZE = 700  # ~960 for LAION, ideal: 1024 if the dataset contains large images
+
 if is_wandb_available():
     import wandb
 
@@ -146,10 +151,10 @@ class WebdatasetFilter:
         try:
             if "json" in x:
                 x_json = json.loads(x["json"])
-                filter_size = (x_json.get("original_width", 0.0) or 0.0) >= self.min_size and x_json.get(
-                    "original_height", 0
+                filter_size = (x_json.get(WDS_JSON_WIDTH, 0.0) or 0.0) >= self.min_size and x_json.get(
+                    WDS_JSON_HEIGHT, 0
                 ) >= self.min_size
-                filter_watermark = (x_json.get("pwatermark", 1.0) or 1.0) <= self.max_pwatermark
+                filter_watermark = (x_json.get("pwatermark", 0.0) or 0.0) <= self.max_pwatermark
                 return filter_size and filter_watermark
             else:
                 return False
@@ -157,7 +162,7 @@ class WebdatasetFilter:
             return False
 
 
-class Text2ImageDataset:
+class SDXLText2ImageDataset:
     def __init__(
         self,
         train_shards_path_or_url: Union[str, List[str]],
@@ -180,7 +185,7 @@ class Text2ImageDataset:
             if use_fix_crop_and_size:
                 return (resolution, resolution)
             else:
-                return (int(json.get("original_width", 0.0)), int(json.get("original_height", 0.0)))
+                return (int(json.get(WDS_JSON_WIDTH, 0.0)), int(json.get(WDS_JSON_HEIGHT, 0.0)))
 
         def transform(example):
             # resize image
@@ -212,7 +217,7 @@ class Text2ImageDataset:
         pipeline = [
             wds.ResampledShards(train_shards_path_or_url),
             tarfile_to_samples_nothrow,
-            wds.select(WebdatasetFilter(min_size=960)),
+            wds.select(WebdatasetFilter(min_size=MIN_SIZE)),
             wds.shuffle(shuffle_buffer_size),
             *processing_pipeline,
             wds.batched(per_gpu_batch_size, partial=False, collation_fn=default_collate),
@@ -341,19 +346,43 @@ def scalings_for_boundary_conditions(timestep, sigma_data=0.5, timestep_scaling=
 
 
 # Compare LCMScheduler.step, Step 4
-def predicted_origin(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+def get_predicted_original_sample(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+    alphas = extract_into_tensor(alphas, timesteps, sample.shape)
+    sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
     if prediction_type == "epsilon":
-        sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
-        alphas = extract_into_tensor(alphas, timesteps, sample.shape)
         pred_x_0 = (sample - sigmas * model_output) / alphas
+    elif prediction_type == "sample":
+        pred_x_0 = model_output
     elif prediction_type == "v_prediction":
-        pred_x_0 = alphas[timesteps] * sample - sigmas[timesteps] * model_output
+        pred_x_0 = alphas * sample - sigmas * model_output
     else:
-        raise ValueError(f"Prediction type {prediction_type} currently not supported.")
+        raise ValueError(
+            f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`"
+            f" are supported."
+        )
 
     return pred_x_0
 
 
+# Based on step 4 in DDIMScheduler.step
+def get_predicted_noise(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+    alphas = extract_into_tensor(alphas, timesteps, sample.shape)
+    sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
+    if prediction_type == "epsilon":
+        pred_epsilon = model_output
+    elif prediction_type == "sample":
+        pred_epsilon = (sample - alphas * model_output) / sigmas
+    elif prediction_type == "v_prediction":
+        pred_epsilon = alphas * model_output + sigmas * sample
+    else:
+        raise ValueError(
+            f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`"
+            f" are supported."
+        )
+
+    return pred_epsilon
+
+
 def extract_into_tensor(a, t, x_shape):
     b, *_ = t.shape
     out = a.gather(-1, t)
@@ -825,9 +854,10 @@ def main(args):
         args.pretrained_teacher_model, subfolder="scheduler", revision=args.teacher_revision
     )
 
-    # The scheduler calculates the alpha and sigma schedule for us
+    # DDPMScheduler calculates the alpha and sigma noise schedules (based on the alpha bars) for us
     alpha_schedule = torch.sqrt(noise_scheduler.alphas_cumprod)
     sigma_schedule = torch.sqrt(1 - noise_scheduler.alphas_cumprod)
+    # Initialize the DDIM ODE solver for distillation.
     solver = DDIMSolver(
         noise_scheduler.alphas_cumprod.numpy(),
         timesteps=noise_scheduler.config.num_train_timesteps,
@@ -881,7 +911,7 @@ def main(args):
     text_encoder_two.requires_grad_(False)
     teacher_unet.requires_grad_(False)
 
-    # 7. Create online (`unet`) student U-Nets.
+    # 7. Create online student U-Net.
     unet = UNet2DConditionModel.from_pretrained(
         args.pretrained_teacher_model, subfolder="unet", revision=args.teacher_revision
     )
@@ -945,6 +975,7 @@ def main(args):
     # Also move the alpha and sigma noise schedules to accelerator.device.
     alpha_schedule = alpha_schedule.to(accelerator.device)
     sigma_schedule = sigma_schedule.to(accelerator.device)
+    # Move the ODE solver to accelerator.device.
     solver = solver.to(accelerator.device)
 
     # 10. Handle saving and loading of checkpoints
@@ -1052,7 +1083,7 @@ def main(args):
 
         return {"prompt_embeds": prompt_embeds, **unet_added_cond_kwargs}
 
-    dataset = Text2ImageDataset(
+    dataset = SDXLText2ImageDataset(
         train_shards_path_or_url=args.train_shards_path_or_url,
         num_train_examples=args.max_train_samples,
         per_gpu_batch_size=args.train_batch_size,
@@ -1170,6 +1201,7 @@ def main(args):
     for epoch in range(first_epoch, args.num_train_epochs):
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
+                # 1. Load and process the image, text, and micro-conditioning (original image size, crop coordinates)
                 image, text, orig_size, crop_coords = batch
 
                 image = image.to(accelerator.device, non_blocking=True)
@@ -1191,37 +1223,37 @@ def main(args):
                 latents = latents * vae.config.scaling_factor
                 if args.pretrained_vae_model_name_or_path is None:
                     latents = latents.to(weight_dtype)
-
-                # Sample noise that we'll add to the latents
-                noise = torch.randn_like(latents)
                 bsz = latents.shape[0]
 
-                # Sample a random timestep for each image t_n ~ U[0, N - k - 1] without bias.
+                # 2. Sample a random timestep for each image t_n from the ODE solver timesteps without bias.
+                # For the DDIM solver, the timestep schedule is [T - 1, T - k - 1, T - 2 * k - 1, ...]
                 topk = noise_scheduler.config.num_train_timesteps // args.num_ddim_timesteps
                 index = torch.randint(0, args.num_ddim_timesteps, (bsz,), device=latents.device).long()
                 start_timesteps = solver.ddim_timesteps[index]
                 timesteps = start_timesteps - topk
                 timesteps = torch.where(timesteps < 0, torch.zeros_like(timesteps), timesteps)
 
-                # 20.4.4. Get boundary scalings for start_timesteps and (end) timesteps.
+                # 3. Get boundary scalings for start_timesteps and (end) timesteps.
                 c_skip_start, c_out_start = scalings_for_boundary_conditions(start_timesteps)
                 c_skip_start, c_out_start = [append_dims(x, latents.ndim) for x in [c_skip_start, c_out_start]]
                 c_skip, c_out = scalings_for_boundary_conditions(timesteps)
                 c_skip, c_out = [append_dims(x, latents.ndim) for x in [c_skip, c_out]]
 
-                # 20.4.5. Add noise to the latents according to the noise magnitude at each timestep
-                # (this is the forward diffusion process) [z_{t_{n + k}} in Algorithm 1]
+                # 4. Sample noise from the prior and add it to the latents according to the noise magnitude at each
+                # timestep (this is the forward diffusion process) [z_{t_{n + k}} in Algorithm 1]
+                noise = torch.randn_like(latents)
                 noisy_model_input = noise_scheduler.add_noise(latents, noise, start_timesteps)
 
-                # 20.4.6. Sample a random guidance scale w from U[w_min, w_max] and embed it
+                # 5. Sample a random guidance scale w from U[w_min, w_max]
+                # Note that for LCM-LoRA distillation it is not necessary to use a guidance scale embedding
                 w = (args.w_max - args.w_min) * torch.rand((bsz,)) + args.w_min
                 w = w.reshape(bsz, 1, 1, 1)
                 w = w.to(device=latents.device, dtype=latents.dtype)
 
-                # 20.4.8. Prepare prompt embeds and unet_added_conditions
+                # 6. Prepare prompt embeds and unet_added_conditions
                 prompt_embeds = encoded_text.pop("prompt_embeds")
 
-                # 20.4.9. Get online LCM prediction on z_{t_{n + k}}, w, c, t_{n + k}
+                # 7. Get online LCM prediction on z_{t_{n + k}} (noisy_model_input), w, c, t_{n + k} (start_timesteps)
                 noise_pred = unet(
                     noisy_model_input,
                     start_timesteps,
@@ -1230,7 +1262,7 @@ def main(args):
                     added_cond_kwargs=encoded_text,
                 ).sample
 
-                pred_x_0 = predicted_origin(
+                pred_x_0 = get_predicted_original_sample(
                     noise_pred,
                     start_timesteps,
                     noisy_model_input,
@@ -1241,18 +1273,28 @@ def main(args):
 
                 model_pred = c_skip_start * noisy_model_input + c_out_start * pred_x_0
 
-                # 20.4.10. Use the ODE solver to predict the kth step in the augmented PF-ODE trajectory after
-                # noisy_latents with both the conditioning embedding c and unconditional embedding 0
-                # Get teacher model prediction on noisy_latents and conditional embedding
+                # 8. Compute the conditional and unconditional teacher model predictions to get CFG estimates of the
+                # predicted noise eps_0 and predicted original sample x_0, then run the ODE solver using these
+                # estimates to predict the data point in the augmented PF-ODE trajectory corresponding to the next ODE
+                # solver timestep.
                 with torch.no_grad():
                     with torch.autocast("cuda"):
+                        # 1. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and conditional embedding c
                         cond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
                             start_timesteps,
                             encoder_hidden_states=prompt_embeds.to(weight_dtype),
                             added_cond_kwargs={k: v.to(weight_dtype) for k, v in encoded_text.items()},
                         ).sample
-                        cond_pred_x0 = predicted_origin(
+                        cond_pred_x0 = get_predicted_original_sample(
+                            cond_teacher_output,
+                            start_timesteps,
+                            noisy_model_input,
+                            noise_scheduler.config.prediction_type,
+                            alpha_schedule,
+                            sigma_schedule,
+                        )
+                        cond_pred_noise = get_predicted_noise(
                             cond_teacher_output,
                             start_timesteps,
                             noisy_model_input,
@@ -1261,7 +1303,7 @@ def main(args):
                             sigma_schedule,
                         )
 
-                        # Get teacher model prediction on noisy_latents and unconditional embedding
+                        # 2. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and unconditional embedding 0
                         uncond_added_conditions = copy.deepcopy(encoded_text)
                         uncond_added_conditions["text_embeds"] = uncond_pooled_prompt_embeds
                         uncond_teacher_output = teacher_unet(
@@ -1270,7 +1312,15 @@ def main(args):
                             encoder_hidden_states=uncond_prompt_embeds.to(weight_dtype),
                             added_cond_kwargs={k: v.to(weight_dtype) for k, v in uncond_added_conditions.items()},
                         ).sample
-                        uncond_pred_x0 = predicted_origin(
+                        uncond_pred_x0 = get_predicted_original_sample(
+                            uncond_teacher_output,
+                            start_timesteps,
+                            noisy_model_input,
+                            noise_scheduler.config.prediction_type,
+                            alpha_schedule,
+                            sigma_schedule,
+                        )
+                        uncond_pred_noise = get_predicted_noise(
                             uncond_teacher_output,
                             start_timesteps,
                             noisy_model_input,
@@ -1279,12 +1329,17 @@ def main(args):
                             sigma_schedule,
                         )
 
-                        # 20.4.11. Perform "CFG" to get x_prev estimate (using the LCM paper's CFG formulation)
+                        # 3. Calculate the CFG estimate of x_0 (pred_x0) and eps_0 (pred_noise)
+                        # Note that this uses the LCM paper's CFG formulation rather than the Imagen CFG formulation
                         pred_x0 = cond_pred_x0 + w * (cond_pred_x0 - uncond_pred_x0)
-                        pred_noise = cond_teacher_output + w * (cond_teacher_output - uncond_teacher_output)
+                        pred_noise = cond_pred_noise + w * (cond_pred_noise - uncond_pred_noise)
+                        # 4. Run one step of the ODE solver to estimate the next point x_prev on the
+                        # augmented PF-ODE trajectory (solving backward in time)
+                        # Note that the DDIM step depends on both the predicted x_0 and source noise eps_0.
                         x_prev = solver.ddim_step(pred_x0, pred_noise, index)
 
-                # 20.4.12. Get target LCM prediction on x_prev, w, c, t_n
+                # 9. Get target LCM prediction on x_prev, w, c, t_n (timesteps)
+                # Note that we do not use a separate target network for LCM-LoRA distillation.
                 with torch.no_grad():
                     with torch.autocast("cuda", enabled=True, dtype=weight_dtype):
                         target_noise_pred = unet(
@@ -1294,7 +1349,7 @@ def main(args):
                             encoder_hidden_states=prompt_embeds.float(),
                             added_cond_kwargs=encoded_text,
                         ).sample
-                    pred_x_0 = predicted_origin(
+                    pred_x_0 = get_predicted_original_sample(
                         target_noise_pred,
                         timesteps,
                         x_prev,
@@ -1304,7 +1359,7 @@ def main(args):
                     )
                     target = c_skip * x_prev + c_out * pred_x_0
 
-                # 20.4.13. Calculate loss
+                # 10. Calculate loss
                 if args.loss_type == "l2":
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
                 elif args.loss_type == "huber":
@@ -1312,7 +1367,7 @@ def main(args):
                         torch.sqrt((model_pred.float() - target.float()) ** 2 + args.huber_c**2) - args.huber_c
                     )
 
-                # 20.4.14. Backpropagate on the online student model (`unet`)
+                # 11. Backpropagate on the online student model (`unet`)
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
                     accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)

examples/consistency_distillation/train_lcm_distill_sd_wds.py

@@ -138,7 +138,7 @@ class WebdatasetFilter:
             return False
 
 
-class Text2ImageDataset:
+class SDText2ImageDataset:
     def __init__(
         self,
         train_shards_path_or_url: Union[str, List[str]],
@@ -336,19 +336,43 @@ def scalings_for_boundary_conditions(timestep, sigma_data=0.5, timestep_scaling=
 
 
 # Compare LCMScheduler.step, Step 4
-def predicted_origin(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+def get_predicted_original_sample(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+    alphas = extract_into_tensor(alphas, timesteps, sample.shape)
+    sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
     if prediction_type == "epsilon":
-        sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
-        alphas = extract_into_tensor(alphas, timesteps, sample.shape)
         pred_x_0 = (sample - sigmas * model_output) / alphas
+    elif prediction_type == "sample":
+        pred_x_0 = model_output
     elif prediction_type == "v_prediction":
-        pred_x_0 = alphas[timesteps] * sample - sigmas[timesteps] * model_output
+        pred_x_0 = alphas * sample - sigmas * model_output
     else:
-        raise ValueError(f"Prediction type {prediction_type} currently not supported.")
+        raise ValueError(
+            f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`"
+            f" are supported."
+        )
 
     return pred_x_0
 
 
+# Based on step 4 in DDIMScheduler.step
+def get_predicted_noise(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+    alphas = extract_into_tensor(alphas, timesteps, sample.shape)
+    sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
+    if prediction_type == "epsilon":
+        pred_epsilon = model_output
+    elif prediction_type == "sample":
+        pred_epsilon = (sample - alphas * model_output) / sigmas
+    elif prediction_type == "v_prediction":
+        pred_epsilon = alphas * model_output + sigmas * sample
+    else:
+        raise ValueError(
+            f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`"
+            f" are supported."
+        )
+
+    return pred_epsilon
+
+
 def extract_into_tensor(a, t, x_shape):
     b, *_ = t.shape
     out = a.gather(-1, t)
@@ -823,34 +847,35 @@ def main(args):
         args.pretrained_teacher_model, subfolder="scheduler", revision=args.teacher_revision
     )
 
-    # The scheduler calculates the alpha and sigma schedule for us
+    # DDPMScheduler calculates the alpha and sigma noise schedules (based on the alpha bars) for us
     alpha_schedule = torch.sqrt(noise_scheduler.alphas_cumprod)
     sigma_schedule = torch.sqrt(1 - noise_scheduler.alphas_cumprod)
+    # Initialize the DDIM ODE solver for distillation.
     solver = DDIMSolver(
         noise_scheduler.alphas_cumprod.numpy(),
         timesteps=noise_scheduler.config.num_train_timesteps,
         ddim_timesteps=args.num_ddim_timesteps,
     )
 
-    # 2. Load tokenizers from SD-XL checkpoint.
+    # 2. Load tokenizers from SD 1.X/2.X checkpoint.
     tokenizer = AutoTokenizer.from_pretrained(
         args.pretrained_teacher_model, subfolder="tokenizer", revision=args.teacher_revision, use_fast=False
     )
 
-    # 3. Load text encoders from SD-1.5 checkpoint.
+    # 3. Load text encoders from SD 1.X/2.X checkpoint.
     # import correct text encoder classes
     text_encoder = CLIPTextModel.from_pretrained(
         args.pretrained_teacher_model, subfolder="text_encoder", revision=args.teacher_revision
     )
 
-    # 4. Load VAE from SD-XL checkpoint (or more stable VAE)
+    # 4. Load VAE from SD 1.X/2.X checkpoint
     vae = AutoencoderKL.from_pretrained(
         args.pretrained_teacher_model,
         subfolder="vae",
         revision=args.teacher_revision,
     )
 
-    # 5. Load teacher U-Net from SD-XL checkpoint
+    # 5. Load teacher U-Net from SD 1.X/2.X checkpoint
     teacher_unet = UNet2DConditionModel.from_pretrained(
         args.pretrained_teacher_model, subfolder="unet", revision=args.teacher_revision
     )
@@ -860,7 +885,7 @@ def main(args):
     text_encoder.requires_grad_(False)
     teacher_unet.requires_grad_(False)
 
-    # 8. Create online (`unet`) student U-Nets. This will be updated by the optimizer (e.g. via backpropagation.)
+    # 7. Create online student U-Net. This will be updated by the optimizer (e.g. via backpropagation.)
     # Add `time_cond_proj_dim` to the student U-Net if `teacher_unet.config.time_cond_proj_dim` is None
     if teacher_unet.config.time_cond_proj_dim is None:
         teacher_unet.config["time_cond_proj_dim"] = args.unet_time_cond_proj_dim
@@ -869,8 +894,8 @@ def main(args):
     unet.load_state_dict(teacher_unet.state_dict(), strict=False)
     unet.train()
 
-    # 9. Create target (`ema_unet`) student U-Net parameters. This will be updated via EMA updates (polyak averaging).
-    # Initialize from unet
+    # 8. Create target student U-Net. This will be updated via EMA updates (polyak averaging).
+    # Initialize from (online) unet
     target_unet = UNet2DConditionModel(**teacher_unet.config)
     target_unet.load_state_dict(unet.state_dict())
     target_unet.train()
@@ -887,7 +912,7 @@ def main(args):
             f"Controlnet loaded as datatype {accelerator.unwrap_model(unet).dtype}. {low_precision_error_string}"
         )
 
-    # 10. Handle mixed precision and device placement
+    # 9. Handle mixed precision and device placement
     # For mixed precision training we cast all non-trainable weigths to half-precision
     # as these weights are only used for inference, keeping weights in full precision is not required.
     weight_dtype = torch.float32
@@ -914,7 +939,7 @@ def main(args):
     sigma_schedule = sigma_schedule.to(accelerator.device)
     solver = solver.to(accelerator.device)
 
-    # 11. Handle saving and loading of checkpoints
+    # 10. Handle saving and loading of checkpoints
     # `accelerate` 0.16.0 will have better support for customized saving
     if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
         # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
@@ -948,7 +973,7 @@ def main(args):
         accelerator.register_save_state_pre_hook(save_model_hook)
         accelerator.register_load_state_pre_hook(load_model_hook)
 
-    # 12. Enable optimizations
+    # 11. Enable optimizations
     if args.enable_xformers_memory_efficient_attention:
         if is_xformers_available():
             import xformers
@@ -994,13 +1019,14 @@ def main(args):
         eps=args.adam_epsilon,
     )
 
+    # 13. Dataset creation and data processing
     # Here, we compute not just the text embeddings but also the additional embeddings
     # needed for the SD XL UNet to operate.
     def compute_embeddings(prompt_batch, proportion_empty_prompts, text_encoder, tokenizer, is_train=True):
         prompt_embeds = encode_prompt(prompt_batch, text_encoder, tokenizer, proportion_empty_prompts, is_train)
         return {"prompt_embeds": prompt_embeds}
 
-    dataset = Text2ImageDataset(
+    dataset = SDText2ImageDataset(
         train_shards_path_or_url=args.train_shards_path_or_url,
         num_train_examples=args.max_train_samples,
         per_gpu_batch_size=args.train_batch_size,
@@ -1020,6 +1046,7 @@ def main(args):
         tokenizer=tokenizer,
     )
 
+    # 14. LR Scheduler creation
     # Scheduler and math around the number of training steps.
     overrode_max_train_steps = False
     num_update_steps_per_epoch = math.ceil(train_dataloader.num_batches / args.gradient_accumulation_steps)
@@ -1034,6 +1061,7 @@ def main(args):
         num_training_steps=args.max_train_steps,
     )
 
+    # 15. Prepare for training
     # Prepare everything with our `accelerator`.
     unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
 
@@ -1055,7 +1083,7 @@ def main(args):
     ).input_ids.to(accelerator.device)
     uncond_prompt_embeds = text_encoder(uncond_input_ids)[0]
 
-    # Train!
+    # 16. Train!
     total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
 
     logger.info("***** Running training *****")
@@ -1106,7 +1134,8 @@ def main(args):
     for epoch in range(first_epoch, args.num_train_epochs):
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
-                image, text, _, _ = batch
+                # 1. Load and process the image and text conditioning
+                image, text = batch
 
                 image = image.to(accelerator.device, non_blocking=True)
                 encoded_text = compute_embeddings_fn(text)
@@ -1123,29 +1152,28 @@ def main(args):
 
                 latents = latents * vae.config.scaling_factor
                 latents = latents.to(weight_dtype)
-
-                # Sample noise that we'll add to the latents
-                noise = torch.randn_like(latents)
                 bsz = latents.shape[0]
 
-                # Sample a random timestep for each image t_n ~ U[0, N - k - 1] without bias.
+                # 2. Sample a random timestep for each image t_n from the ODE solver timesteps without bias.
+                # For the DDIM solver, the timestep schedule is [T - 1, T - k - 1, T - 2 * k - 1, ...]
                 topk = noise_scheduler.config.num_train_timesteps // args.num_ddim_timesteps
                 index = torch.randint(0, args.num_ddim_timesteps, (bsz,), device=latents.device).long()
                 start_timesteps = solver.ddim_timesteps[index]
                 timesteps = start_timesteps - topk
                 timesteps = torch.where(timesteps < 0, torch.zeros_like(timesteps), timesteps)
 
-                # 20.4.4. Get boundary scalings for start_timesteps and (end) timesteps.
+                # 3. Get boundary scalings for start_timesteps and (end) timesteps.
                 c_skip_start, c_out_start = scalings_for_boundary_conditions(start_timesteps)
                 c_skip_start, c_out_start = [append_dims(x, latents.ndim) for x in [c_skip_start, c_out_start]]
                 c_skip, c_out = scalings_for_boundary_conditions(timesteps)
                 c_skip, c_out = [append_dims(x, latents.ndim) for x in [c_skip, c_out]]
 
-                # 20.4.5. Add noise to the latents according to the noise magnitude at each timestep
-                # (this is the forward diffusion process) [z_{t_{n + k}} in Algorithm 1]
+                # 4. Sample noise from the prior and add it to the latents according to the noise magnitude at each
+                # timestep (this is the forward diffusion process) [z_{t_{n + k}} in Algorithm 1]
+                noise = torch.randn_like(latents)
                 noisy_model_input = noise_scheduler.add_noise(latents, noise, start_timesteps)
 
-                # 20.4.6. Sample a random guidance scale w from U[w_min, w_max] and embed it
+                # 5. Sample a random guidance scale w from U[w_min, w_max] and embed it
                 w = (args.w_max - args.w_min) * torch.rand((bsz,)) + args.w_min
                 w_embedding = guidance_scale_embedding(w, embedding_dim=unet.config.time_cond_proj_dim)
                 w = w.reshape(bsz, 1, 1, 1)
@@ -1153,10 +1181,10 @@ def main(args):
                 w = w.to(device=latents.device, dtype=latents.dtype)
                 w_embedding = w_embedding.to(device=latents.device, dtype=latents.dtype)
 
-                # 20.4.8. Prepare prompt embeds and unet_added_conditions
+                # 6. Prepare prompt embeds and unet_added_conditions
                 prompt_embeds = encoded_text.pop("prompt_embeds")
 
-                # 20.4.9. Get online LCM prediction on z_{t_{n + k}}, w, c, t_{n + k}
+                # 7. Get online LCM prediction on z_{t_{n + k}} (noisy_model_input), w, c, t_{n + k} (start_timesteps)
                 noise_pred = unet(
                     noisy_model_input,
                     start_timesteps,
@@ -1165,7 +1193,7 @@ def main(args):
                     added_cond_kwargs=encoded_text,
                 ).sample
 
-                pred_x_0 = predicted_origin(
+                pred_x_0 = get_predicted_original_sample(
                     noise_pred,
                     start_timesteps,
                     noisy_model_input,
@@ -1176,17 +1204,27 @@ def main(args):
 
                 model_pred = c_skip_start * noisy_model_input + c_out_start * pred_x_0
 
-                # 20.4.10. Use the ODE solver to predict the kth step in the augmented PF-ODE trajectory after
-                # noisy_latents with both the conditioning embedding c and unconditional embedding 0
-                # Get teacher model prediction on noisy_latents and conditional embedding
+                # 8. Compute the conditional and unconditional teacher model predictions to get CFG estimates of the
+                # predicted noise eps_0 and predicted original sample x_0, then run the ODE solver using these
+                # estimates to predict the data point in the augmented PF-ODE trajectory corresponding to the next ODE
+                # solver timestep.
                 with torch.no_grad():
                     with torch.autocast("cuda"):
+                        # 1. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and conditional embedding c
                         cond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
                             start_timesteps,
                             encoder_hidden_states=prompt_embeds.to(weight_dtype),
                         ).sample
-                        cond_pred_x0 = predicted_origin(
+                        cond_pred_x0 = get_predicted_original_sample(
+                            cond_teacher_output,
+                            start_timesteps,
+                            noisy_model_input,
+                            noise_scheduler.config.prediction_type,
+                            alpha_schedule,
+                            sigma_schedule,
+                        )
+                        cond_pred_noise = get_predicted_noise(
                             cond_teacher_output,
                             start_timesteps,
                             noisy_model_input,
@@ -1195,13 +1233,21 @@ def main(args):
                             sigma_schedule,
                         )
 
-                        # Get teacher model prediction on noisy_latents and unconditional embedding
+                        # 2. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and unconditional embedding 0
                         uncond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
                             start_timesteps,
                             encoder_hidden_states=uncond_prompt_embeds.to(weight_dtype),
                         ).sample
-                        uncond_pred_x0 = predicted_origin(
+                        uncond_pred_x0 = get_predicted_original_sample(
+                            uncond_teacher_output,
+                            start_timesteps,
+                            noisy_model_input,
+                            noise_scheduler.config.prediction_type,
+                            alpha_schedule,
+                            sigma_schedule,
+                        )
+                        uncond_pred_noise = get_predicted_noise(
                             uncond_teacher_output,
                             start_timesteps,
                             noisy_model_input,
@@ -1210,12 +1256,16 @@ def main(args):
                             sigma_schedule,
                         )
 
-                        # 20.4.11. Perform "CFG" to get x_prev estimate (using the LCM paper's CFG formulation)
+                        # 3. Calculate the CFG estimate of x_0 (pred_x0) and eps_0 (pred_noise)
+                        # Note that this uses the LCM paper's CFG formulation rather than the Imagen CFG formulation
                         pred_x0 = cond_pred_x0 + w * (cond_pred_x0 - uncond_pred_x0)
-                        pred_noise = cond_teacher_output + w * (cond_teacher_output - uncond_teacher_output)
+                        pred_noise = cond_pred_noise + w * (cond_pred_noise - uncond_pred_noise)
+                        # 4. Run one step of the ODE solver to estimate the next point x_prev on the
+                        # augmented PF-ODE trajectory (solving backward in time)
+                        # Note that the DDIM step depends on both the predicted x_0 and source noise eps_0.
                         x_prev = solver.ddim_step(pred_x0, pred_noise, index)
 
-                # 20.4.12. Get target LCM prediction on x_prev, w, c, t_n
+                # 9. Get target LCM prediction on x_prev, w, c, t_n (timesteps)
                 with torch.no_grad():
                     with torch.autocast("cuda", dtype=weight_dtype):
                         target_noise_pred = target_unet(
@@ -1224,7 +1274,7 @@ def main(args):
                             timestep_cond=w_embedding,
                             encoder_hidden_states=prompt_embeds.float(),
                         ).sample
-                    pred_x_0 = predicted_origin(
+                    pred_x_0 = get_predicted_original_sample(
                         target_noise_pred,
                         timesteps,
                         x_prev,
@@ -1234,7 +1284,7 @@ def main(args):
                     )
                     target = c_skip * x_prev + c_out * pred_x_0
 
-                # 20.4.13. Calculate loss
+                # 10. Calculate loss
                 if args.loss_type == "l2":
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
                 elif args.loss_type == "huber":
@@ -1242,7 +1292,7 @@ def main(args):
                         torch.sqrt((model_pred.float() - target.float()) ** 2 + args.huber_c**2) - args.huber_c
                     )
 
-                # 20.4.14. Backpropagate on the online student model (`unet`)
+                # 11. Backpropagate on the online student model (`unet`)
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
                     accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
@@ -1252,7 +1302,7 @@ def main(args):
 
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
-                # 20.4.15. Make EMA update to target student model parameters
+                # 12. Make EMA update to target student model parameters (`target_unet`)
                 update_ema(target_unet.parameters(), unet.parameters(), args.ema_decay)
                 progress_bar.update(1)
                 global_step += 1

examples/consistency_distillation/train_lcm_distill_sdxl_wds.py

@@ -67,6 +67,11 @@ from diffusers.utils.import_utils import is_xformers_available
 
 MAX_SEQ_LENGTH = 77
 
+# Adjust for your dataset
+WDS_JSON_WIDTH = "width"  # original_width for LAION
+WDS_JSON_HEIGHT = "height"  # original_height for LAION
+MIN_SIZE = 700  # ~960 for LAION, ideal: 1024 if the dataset contains large images
+
 if is_wandb_available():
     import wandb
 
@@ -128,10 +133,10 @@ class WebdatasetFilter:
         try:
             if "json" in x:
                 x_json = json.loads(x["json"])
-                filter_size = (x_json.get("original_width", 0.0) or 0.0) >= self.min_size and x_json.get(
-                    "original_height", 0
+                filter_size = (x_json.get(WDS_JSON_WIDTH, 0.0) or 0.0) >= self.min_size and x_json.get(
+                    WDS_JSON_HEIGHT, 0
                 ) >= self.min_size
-                filter_watermark = (x_json.get("pwatermark", 1.0) or 1.0) <= self.max_pwatermark
+                filter_watermark = (x_json.get("pwatermark", 0.0) or 0.0) <= self.max_pwatermark
                 return filter_size and filter_watermark
             else:
                 return False
@@ -139,7 +144,7 @@ class WebdatasetFilter:
             return False
 
 
-class Text2ImageDataset:
+class SDXLText2ImageDataset:
     def __init__(
         self,
         train_shards_path_or_url: Union[str, List[str]],
@@ -162,7 +167,7 @@ class Text2ImageDataset:
             if use_fix_crop_and_size:
                 return (resolution, resolution)
             else:
-                return (int(json.get("original_width", 0.0)), int(json.get("original_height", 0.0)))
+                return (int(json.get(WDS_JSON_WIDTH, 0.0)), int(json.get(WDS_JSON_HEIGHT, 0.0)))
 
         def transform(example):
             # resize image
@@ -194,7 +199,7 @@ class Text2ImageDataset:
         pipeline = [
             wds.ResampledShards(train_shards_path_or_url),
             tarfile_to_samples_nothrow,
-            wds.select(WebdatasetFilter(min_size=960)),
+            wds.select(WebdatasetFilter(min_size=MIN_SIZE)),
             wds.shuffle(shuffle_buffer_size),
             *processing_pipeline,
             wds.batched(per_gpu_batch_size, partial=False, collation_fn=default_collate),
@@ -319,19 +324,43 @@ def scalings_for_boundary_conditions(timestep, sigma_data=0.5, timestep_scaling=
 
 
 # Compare LCMScheduler.step, Step 4
-def predicted_origin(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+def get_predicted_original_sample(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+    alphas = extract_into_tensor(alphas, timesteps, sample.shape)
+    sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
     if prediction_type == "epsilon":
-        sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
-        alphas = extract_into_tensor(alphas, timesteps, sample.shape)
         pred_x_0 = (sample - sigmas * model_output) / alphas
+    elif prediction_type == "sample":
+        pred_x_0 = model_output
     elif prediction_type == "v_prediction":
-        pred_x_0 = alphas[timesteps] * sample - sigmas[timesteps] * model_output
+        pred_x_0 = alphas * sample - sigmas * model_output
     else:
-        raise ValueError(f"Prediction type {prediction_type} currently not supported.")
+        raise ValueError(
+            f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`"
+            f" are supported."
+        )
 
     return pred_x_0
 
 
+# Based on step 4 in DDIMScheduler.step
+def get_predicted_noise(model_output, timesteps, sample, prediction_type, alphas, sigmas):
+    alphas = extract_into_tensor(alphas, timesteps, sample.shape)
+    sigmas = extract_into_tensor(sigmas, timesteps, sample.shape)
+    if prediction_type == "epsilon":
+        pred_epsilon = model_output
+    elif prediction_type == "sample":
+        pred_epsilon = (sample - alphas * model_output) / sigmas
+    elif prediction_type == "v_prediction":
+        pred_epsilon = alphas * model_output + sigmas * sample
+    else:
+        raise ValueError(
+            f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`"
+            f" are supported."
+        )
+
+    return pred_epsilon
+
+
 def extract_into_tensor(a, t, x_shape):
     b, *_ = t.shape
     out = a.gather(-1, t)
@@ -858,9 +887,10 @@ def main(args):
         args.pretrained_teacher_model, subfolder="scheduler", revision=args.teacher_revision
     )
 
-    # The scheduler calculates the alpha and sigma schedule for us
+    # DDPMScheduler calculates the alpha and sigma noise schedules (based on the alpha bars) for us
     alpha_schedule = torch.sqrt(noise_scheduler.alphas_cumprod)
     sigma_schedule = torch.sqrt(1 - noise_scheduler.alphas_cumprod)
+    # Initialize the DDIM ODE solver for distillation.
     solver = DDIMSolver(
         noise_scheduler.alphas_cumprod.numpy(),
         timesteps=noise_scheduler.config.num_train_timesteps,
@@ -914,7 +944,7 @@ def main(args):
     text_encoder_two.requires_grad_(False)
     teacher_unet.requires_grad_(False)
 
-    # 8. Create online (`unet`) student U-Nets. This will be updated by the optimizer (e.g. via backpropagation.)
+    # 7. Create online student U-Net. This will be updated by the optimizer (e.g. via backpropagation.)
     # Add `time_cond_proj_dim` to the student U-Net if `teacher_unet.config.time_cond_proj_dim` is None
     if teacher_unet.config.time_cond_proj_dim is None:
         teacher_unet.config["time_cond_proj_dim"] = args.unet_time_cond_proj_dim
@@ -923,8 +953,8 @@ def main(args):
     unet.load_state_dict(teacher_unet.state_dict(), strict=False)
     unet.train()
 
-    # 9. Create target (`ema_unet`) student U-Net parameters. This will be updated via EMA updates (polyak averaging).
-    # Initialize from unet
+    # 8. Create target student U-Net. This will be updated via EMA updates (polyak averaging).
+    # Initialize from (online) unet
     target_unet = UNet2DConditionModel(**teacher_unet.config)
     target_unet.load_state_dict(unet.state_dict())
     target_unet.train()
@@ -966,6 +996,7 @@ def main(args):
     # Also move the alpha and sigma noise schedules to accelerator.device.
     alpha_schedule = alpha_schedule.to(accelerator.device)
     sigma_schedule = sigma_schedule.to(accelerator.device)
+    # Move the ODE solver to accelerator.device.
     solver = solver.to(accelerator.device)
 
     # 10. Handle saving and loading of checkpoints
@@ -1079,7 +1110,7 @@ def main(args):
 
         return {"prompt_embeds": prompt_embeds, **unet_added_cond_kwargs}
 
-    dataset = Text2ImageDataset(
+    dataset = SDXLText2ImageDataset(
         train_shards_path_or_url=args.train_shards_path_or_url,
         num_train_examples=args.max_train_samples,
         per_gpu_batch_size=args.train_batch_size,
@@ -1197,6 +1228,7 @@ def main(args):
     for epoch in range(first_epoch, args.num_train_epochs):
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
+                # 1. Load and process the image, text, and micro-conditioning (original image size, crop coordinates)
                 image, text, orig_size, crop_coords = batch
 
                 image = image.to(accelerator.device, non_blocking=True)
@@ -1218,38 +1250,39 @@ def main(args):
                 latents = latents * vae.config.scaling_factor
                 if args.pretrained_vae_model_name_or_path is None:
                     latents = latents.to(weight_dtype)
-
-                # Sample noise that we'll add to the latents
-                noise = torch.randn_like(latents)
                 bsz = latents.shape[0]
 
-                # Sample a random timestep for each image t_n ~ U[0, N - k - 1] without bias.
+                # 2. Sample a random timestep for each image t_n from the ODE solver timesteps without bias.
+                # For the DDIM solver, the timestep schedule is [T - 1, T - k - 1, T - 2 * k - 1, ...]
                 topk = noise_scheduler.config.num_train_timesteps // args.num_ddim_timesteps
                 index = torch.randint(0, args.num_ddim_timesteps, (bsz,), device=latents.device).long()
                 start_timesteps = solver.ddim_timesteps[index]
                 timesteps = start_timesteps - topk
                 timesteps = torch.where(timesteps < 0, torch.zeros_like(timesteps), timesteps)
 
-                # 20.4.4. Get boundary scalings for start_timesteps and (end) timesteps.
+                # 3. Get boundary scalings for start_timesteps and (end) timesteps.
                 c_skip_start, c_out_start = scalings_for_boundary_conditions(start_timesteps)
                 c_skip_start, c_out_start = [append_dims(x, latents.ndim) for x in [c_skip_start, c_out_start]]
                 c_skip, c_out = scalings_for_boundary_conditions(timesteps)
                 c_skip, c_out = [append_dims(x, latents.ndim) for x in [c_skip, c_out]]
 
-                # 20.4.5. Add noise to the latents according to the noise magnitude at each timestep
-                # (this is the forward diffusion process) [z_{t_{n + k}} in Algorithm 1]
+                # 4. Sample noise from the prior and add it to the latents according to the noise magnitude at each
+                # timestep (this is the forward diffusion process) [z_{t_{n + k}} in Algorithm 1]
+                noise = torch.randn_like(latents)
                 noisy_model_input = noise_scheduler.add_noise(latents, noise, start_timesteps)
 
-                # 20.4.6. Sample a random guidance scale w from U[w_min, w_max] and embed it
+                # 5. Sample a random guidance scale w from U[w_min, w_max] and embed it
                 w = (args.w_max - args.w_min) * torch.rand((bsz,)) + args.w_min
                 w_embedding = guidance_scale_embedding(w, embedding_dim=unet.config.time_cond_proj_dim)
                 w = w.reshape(bsz, 1, 1, 1)
+                # Move to U-Net device and dtype
                 w = w.to(device=latents.device, dtype=latents.dtype)
+                w_embedding = w_embedding.to(device=latents.device, dtype=latents.dtype)
 
-                # 20.4.8. Prepare prompt embeds and unet_added_conditions
+                # 6. Prepare prompt embeds and unet_added_conditions
                 prompt_embeds = encoded_text.pop("prompt_embeds")
 
-                # 20.4.9. Get online LCM prediction on z_{t_{n + k}}, w, c, t_{n + k}
+                # 7. Get online LCM prediction on z_{t_{n + k}} (noisy_model_input), w, c, t_{n + k} (start_timesteps)
                 noise_pred = unet(
                     noisy_model_input,
                     start_timesteps,
@@ -1258,7 +1291,7 @@ def main(args):
                     added_cond_kwargs=encoded_text,
                 ).sample
 
-                pred_x_0 = predicted_origin(
+                pred_x_0 = get_predicted_original_sample(
                     noise_pred,
                     start_timesteps,
                     noisy_model_input,
@@ -1269,18 +1302,28 @@ def main(args):
 
                 model_pred = c_skip_start * noisy_model_input + c_out_start * pred_x_0
 
-                # 20.4.10. Use the ODE solver to predict the kth step in the augmented PF-ODE trajectory after
-                # noisy_latents with both the conditioning embedding c and unconditional embedding 0
-                # Get teacher model prediction on noisy_latents and conditional embedding
+                # 8. Compute the conditional and unconditional teacher model predictions to get CFG estimates of the
+                # predicted noise eps_0 and predicted original sample x_0, then run the ODE solver using these
+                # estimates to predict the data point in the augmented PF-ODE trajectory corresponding to the next ODE
+                # solver timestep.
                 with torch.no_grad():
                     with torch.autocast("cuda"):
+                        # 1. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and conditional embedding c
                         cond_teacher_output = teacher_unet(
                             noisy_model_input.to(weight_dtype),
                             start_timesteps,
                             encoder_hidden_states=prompt_embeds.to(weight_dtype),
                             added_cond_kwargs={k: v.to(weight_dtype) for k, v in encoded_text.items()},
                         ).sample
-                        cond_pred_x0 = predicted_origin(
+                        cond_pred_x0 = get_predicted_original_sample(
+                            cond_teacher_output,
+                            start_timesteps,
+                            noisy_model_input,
+                            noise_scheduler.config.prediction_type,
+                            alpha_schedule,
+                            sigma_schedule,
+                        )
+                        cond_pred_noise = get_predicted_noise(
                             cond_teacher_output,
                             start_timesteps,
                             noisy_model_input,
@@ -1289,7 +1332,7 @@ def main(args):
                             sigma_schedule,
                         )
 
-                        # Get teacher model prediction on noisy_latents and unconditional embedding
+                        # 2. Get teacher model prediction on noisy_model_input z_{t_{n + k}} and unconditional embedding 0
                         uncond_added_conditions = copy.deepcopy(encoded_text)
                         uncond_added_conditions["text_embeds"] = uncond_pooled_prompt_embeds
                         uncond_teacher_output = teacher_unet(
@@ -1298,7 +1341,15 @@ def main(args):
                             encoder_hidden_states=uncond_prompt_embeds.to(weight_dtype),
                             added_cond_kwargs={k: v.to(weight_dtype) for k, v in uncond_added_conditions.items()},
                         ).sample
-                        uncond_pred_x0 = predicted_origin(
+                        uncond_pred_x0 = get_predicted_original_sample(
+                            uncond_teacher_output,
+                            start_timesteps,
+                            noisy_model_input,
+                            noise_scheduler.config.prediction_type,
+                            alpha_schedule,
+                            sigma_schedule,
+                        )
+                        uncond_pred_noise = get_predicted_noise(
                             uncond_teacher_output,
                             start_timesteps,
                             noisy_model_input,
@@ -1307,12 +1358,16 @@ def main(args):
                             sigma_schedule,
                         )
 
-                        # 20.4.11. Perform "CFG" to get x_prev estimate (using the LCM paper's CFG formulation)
+                        # 3. Calculate the CFG estimate of x_0 (pred_x0) and eps_0 (pred_noise)
+                        # Note that this uses the LCM paper's CFG formulation rather than the Imagen CFG formulation
                         pred_x0 = cond_pred_x0 + w * (cond_pred_x0 - uncond_pred_x0)
-                        pred_noise = cond_teacher_output + w * (cond_teacher_output - uncond_teacher_output)
+                        pred_noise = cond_pred_noise + w * (cond_pred_noise - uncond_pred_noise)
+                        # 4. Run one step of the ODE solver to estimate the next point x_prev on the
+                        # augmented PF-ODE trajectory (solving backward in time)
+                        # Note that the DDIM step depends on both the predicted x_0 and source noise eps_0.
                         x_prev = solver.ddim_step(pred_x0, pred_noise, index)
 
-                # 20.4.12. Get target LCM prediction on x_prev, w, c, t_n
+                # 9. Get target LCM prediction on x_prev, w, c, t_n (timesteps)
                 with torch.no_grad():
                     with torch.autocast("cuda", dtype=weight_dtype):
                         target_noise_pred = target_unet(
@@ -1322,7 +1377,7 @@ def main(args):
                             encoder_hidden_states=prompt_embeds.float(),
                             added_cond_kwargs=encoded_text,
                         ).sample
-                    pred_x_0 = predicted_origin(
+                    pred_x_0 = get_predicted_original_sample(
                         target_noise_pred,
                         timesteps,
                         x_prev,
@@ -1332,7 +1387,7 @@ def main(args):
                     )
                     target = c_skip * x_prev + c_out * pred_x_0
 
-                # 20.4.13. Calculate loss
+                # 10. Calculate loss
                 if args.loss_type == "l2":
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
                 elif args.loss_type == "huber":
@@ -1340,7 +1395,7 @@ def main(args):
                         torch.sqrt((model_pred.float() - target.float()) ** 2 + args.huber_c**2) - args.huber_c
                     )
 
-                # 20.4.14. Backpropagate on the online student model (`unet`)
+                # 11. Backpropagate on the online student model (`unet`)
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
                     accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
@@ -1350,7 +1405,7 @@ def main(args):
 
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
-                # 20.4.15. Make EMA update to target student model parameters
+                # 12. Make EMA update to target student model parameters (`target_unet`)
                 update_ema(target_unet.parameters(), unet.parameters(), args.ema_decay)
                 progress_bar.update(1)
                 global_step += 1

examples/dreambooth/README.md

@@ -44,6 +44,7 @@ write_basic_config()
 ```
 
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups. 
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
 
 ### Dog toy example
 

examples/dreambooth/README_sdxl.md

@@ -47,6 +47,7 @@ write_basic_config()
 ```
 
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups. 
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
 
 ### Dog toy example
 

examples/dreambooth/requirements.txt

@@ -4,3 +4,4 @@ transformers>=4.25.1
 ftfy
 tensorboard
 Jinja2
+peft==0.7.0

examples/dreambooth/requirements_sdxl.txt

@@ -4,3 +4,4 @@ transformers>=4.25.1
 ftfy
 tensorboard
 Jinja2
+peft==0.7.0

examples/dreambooth/train_dreambooth_lora.py

@@ -16,7 +16,6 @@
 import argparse
 import copy
 import gc
-import itertools
 import logging
 import math
 import os
@@ -35,6 +34,8 @@ from accelerate.utils import ProjectConfiguration, set_seed
 from huggingface_hub import create_repo, upload_folder
 from huggingface_hub.utils import insecure_hashlib
 from packaging import version
+from peft import LoraConfig
+from peft.utils import get_peft_model_state_dict
 from PIL import Image
 from PIL.ImageOps import exif_transpose
 from torch.utils.data import Dataset
@@ -52,14 +53,7 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.loaders import LoraLoaderMixin
-from diffusers.models.attention_processor import (
-    AttnAddedKVProcessor,
-    AttnAddedKVProcessor2_0,
-    SlicedAttnAddedKVProcessor,
-)
-from diffusers.models.lora import LoRALinearLayer
 from diffusers.optimization import get_scheduler
-from diffusers.training_utils import unet_lora_state_dict
 from diffusers.utils import check_min_version, is_wandb_available
 from diffusers.utils.import_utils import is_xformers_available
 
@@ -70,39 +64,6 @@ check_min_version("0.25.0.dev0")
 logger = get_logger(__name__)
 
 
-# TODO: This function should be removed once training scripts are rewritten in PEFT
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    def text_encoder_attn_modules(text_encoder):
-        from transformers import CLIPTextModel, CLIPTextModelWithProjection
-
-        attn_modules = []
-
-        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-                name = f"text_model.encoder.layers.{i}.self_attn"
-                mod = layer.self_attn
-                attn_modules.append((name, mod))
-
-        return attn_modules
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
 def save_model_card(
     repo_id: str,
     images=None,
@@ -864,79 +825,19 @@ def main(args):
             text_encoder.gradient_checkpointing_enable()
 
     # now we will add new LoRA weights to the attention layers
-    # It's important to realize here how many attention weights will be added and of which sizes
-    # The sizes of the attention layers consist only of two different variables:
-    # 1) - the "hidden_size", which is increased according to `unet.config.block_out_channels`.
-    # 2) - the "cross attention size", which is set to `unet.config.cross_attention_dim`.
+    unet_lora_config = LoraConfig(
+        r=args.rank,
+        init_lora_weights="gaussian",
+        target_modules=["to_k", "to_q", "to_v", "to_out.0", "add_k_proj", "add_v_proj"],
+    )
+    unet.add_adapter(unet_lora_config)
 
-    # Let's first see how many attention processors we will have to set.
-    # For Stable Diffusion, it should be equal to:
-    # - down blocks (2x attention layers) * (2x transformer layers) * (3x down blocks) = 12
-    # - mid blocks (2x attention layers) * (1x transformer layers) * (1x mid blocks) = 2
-    # - up blocks (2x attention layers) * (3x transformer layers) * (3x up blocks) = 18
-    # => 32 layers
-
-    # Set correct lora layers
-    unet_lora_parameters = []
-    for attn_processor_name, attn_processor in unet.attn_processors.items():
-        # Parse the attention module.
-        attn_module = unet
-        for n in attn_processor_name.split(".")[:-1]:
-            attn_module = getattr(attn_module, n)
-
-        # Set the `lora_layer` attribute of the attention-related matrices.
-        attn_module.to_q.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_k.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_v.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_out[0].set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_out[0].in_features,
-                out_features=attn_module.to_out[0].out_features,
-                rank=args.rank,
-            )
-        )
-
-        # Accumulate the LoRA params to optimize.
-        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
-
-        if isinstance(attn_processor, (AttnAddedKVProcessor, SlicedAttnAddedKVProcessor, AttnAddedKVProcessor2_0)):
-            attn_module.add_k_proj.set_lora_layer(
-                LoRALinearLayer(
-                    in_features=attn_module.add_k_proj.in_features,
-                    out_features=attn_module.add_k_proj.out_features,
-                    rank=args.rank,
-                )
-            )
-            attn_module.add_v_proj.set_lora_layer(
-                LoRALinearLayer(
-                    in_features=attn_module.add_v_proj.in_features,
-                    out_features=attn_module.add_v_proj.out_features,
-                    rank=args.rank,
-                )
-            )
-            unet_lora_parameters.extend(attn_module.add_k_proj.lora_layer.parameters())
-            unet_lora_parameters.extend(attn_module.add_v_proj.lora_layer.parameters())
-
-    # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
-    # So, instead, we monkey-patch the forward calls of its attention-blocks.
+    # The text encoder comes from 🤗 transformers, we will also attach adapters to it.
     if args.train_text_encoder:
-        # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-        text_lora_parameters = LoraLoaderMixin._modify_text_encoder(text_encoder, dtype=torch.float32, rank=args.rank)
+        text_lora_config = LoraConfig(
+            r=args.rank, init_lora_weights="gaussian", target_modules=["q_proj", "k_proj", "v_proj", "out_proj"]
+        )
+        text_encoder.add_adapter(text_lora_config)
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
@@ -948,9 +849,9 @@ def main(args):
 
             for model in models:
                 if isinstance(model, type(accelerator.unwrap_model(unet))):
-                    unet_lora_layers_to_save = unet_lora_state_dict(model)
+                    unet_lora_layers_to_save = get_peft_model_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder))):
-                    text_encoder_lora_layers_to_save = text_encoder_lora_state_dict(model)
+                    text_encoder_lora_layers_to_save = get_peft_model_state_dict(model)
                 else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 
@@ -1010,11 +911,10 @@ def main(args):
         optimizer_class = torch.optim.AdamW
 
     # Optimizer creation
-    params_to_optimize = (
-        itertools.chain(unet_lora_parameters, text_lora_parameters)
-        if args.train_text_encoder
-        else unet_lora_parameters
-    )
+    params_to_optimize = list(filter(lambda p: p.requires_grad, unet.parameters()))
+    if args.train_text_encoder:
+        params_to_optimize = params_to_optimize + list(filter(lambda p: p.requires_grad, text_encoder.parameters()))
+
     optimizer = optimizer_class(
         params_to_optimize,
         lr=args.learning_rate,
@@ -1257,12 +1157,7 @@ def main(args):
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    params_to_clip = (
-                        itertools.chain(unet_lora_parameters, text_lora_parameters)
-                        if args.train_text_encoder
-                        else unet_lora_parameters
-                    )
-                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+                    accelerator.clip_grad_norm_(params_to_optimize, args.max_grad_norm)
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()
@@ -1385,19 +1280,19 @@ def main(args):
     if accelerator.is_main_process:
         unet = accelerator.unwrap_model(unet)
         unet = unet.to(torch.float32)
-        unet_lora_layers = unet_lora_state_dict(unet)
 
-        if text_encoder is not None and args.train_text_encoder:
+        unet_lora_state_dict = get_peft_model_state_dict(unet)
+
+        if args.train_text_encoder:
             text_encoder = accelerator.unwrap_model(text_encoder)
-            text_encoder = text_encoder.to(torch.float32)
-            text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder)
+            text_encoder_state_dict = get_peft_model_state_dict(text_encoder)
         else:
-            text_encoder_lora_layers = None
+            text_encoder_state_dict = None
 
         LoraLoaderMixin.save_lora_weights(
             save_directory=args.output_dir,
-            unet_lora_layers=unet_lora_layers,
-            text_encoder_lora_layers=text_encoder_lora_layers,
+            unet_lora_layers=unet_lora_state_dict,
+            text_encoder_lora_layers=text_encoder_state_dict,
         )
 
         # Final inference

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -34,6 +34,8 @@ from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration
 from huggingface_hub import create_repo, upload_folder
 from huggingface_hub.utils import insecure_hashlib
 from packaging import version
+from peft import LoraConfig
+from peft.utils import get_peft_model_state_dict
 from PIL import Image
 from PIL.ImageOps import exif_transpose
 from torch.utils.data import Dataset
@@ -50,9 +52,8 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.loaders import LoraLoaderMixin
-from diffusers.models.lora import LoRALinearLayer
 from diffusers.optimization import get_scheduler
-from diffusers.training_utils import compute_snr, unet_lora_state_dict
+from diffusers.training_utils import compute_snr
 from diffusers.utils import check_min_version, is_wandb_available
 from diffusers.utils.import_utils import is_xformers_available
 
@@ -63,39 +64,6 @@ check_min_version("0.25.0.dev0")
 logger = get_logger(__name__)
 
 
-# TODO: This function should be removed once training scripts are rewritten in PEFT
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    def text_encoder_attn_modules(text_encoder):
-        from transformers import CLIPTextModel, CLIPTextModelWithProjection
-
-        attn_modules = []
-
-        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-                name = f"text_model.encoder.layers.{i}.self_attn"
-                mod = layer.self_attn
-                attn_modules.append((name, mod))
-
-        return attn_modules
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
 def save_model_card(
     repo_id: str,
     images=None,
@@ -1009,54 +977,30 @@ def main(args):
             text_encoder_two.gradient_checkpointing_enable()
 
     # now we will add new LoRA weights to the attention layers
-    # Set correct lora layers
-    unet_lora_parameters = []
-    for attn_processor_name, attn_processor in unet.attn_processors.items():
-        # Parse the attention module.
-        attn_module = unet
-        for n in attn_processor_name.split(".")[:-1]:
-            attn_module = getattr(attn_module, n)
-
-        # Set the `lora_layer` attribute of the attention-related matrices.
-        attn_module.to_q.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_k.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_v.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_out[0].set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_out[0].in_features,
-                out_features=attn_module.to_out[0].out_features,
-                rank=args.rank,
-            )
-        )
-
-        # Accumulate the LoRA params to optimize.
-        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
+    unet_lora_config = LoraConfig(
+        r=args.rank, init_lora_weights="gaussian", target_modules=["to_k", "to_q", "to_v", "to_out.0"]
+    )
+    unet.add_adapter(unet_lora_config)
 
     # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
     # So, instead, we monkey-patch the forward calls of its attention-blocks.
     if args.train_text_encoder:
-        # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-        text_lora_parameters_one = LoraLoaderMixin._modify_text_encoder(
-            text_encoder_one, dtype=torch.float32, rank=args.rank
-        )
-        text_lora_parameters_two = LoraLoaderMixin._modify_text_encoder(
-            text_encoder_two, dtype=torch.float32, rank=args.rank
+        text_lora_config = LoraConfig(
+            r=args.rank, init_lora_weights="gaussian", target_modules=["q_proj", "k_proj", "v_proj", "out_proj"]
         )
+        text_encoder_one.add_adapter(text_lora_config)
+        text_encoder_two.add_adapter(text_lora_config)
+
+    # Make sure the trainable params are in float32.
+    if args.mixed_precision == "fp16":
+        models = [unet]
+        if args.train_text_encoder:
+            models.extend([text_encoder_one, text_encoder_two])
+        for model in models:
+            for param in model.parameters():
+                # only upcast trainable parameters (LoRA) into fp32
+                if param.requires_grad:
+                    param.data = param.to(torch.float32)
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
@@ -1069,11 +1013,11 @@ def main(args):
 
             for model in models:
                 if isinstance(model, type(accelerator.unwrap_model(unet))):
-                    unet_lora_layers_to_save = unet_lora_state_dict(model)
+                    unet_lora_layers_to_save = get_peft_model_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
-                    text_encoder_one_lora_layers_to_save = text_encoder_lora_state_dict(model)
+                    text_encoder_one_lora_layers_to_save = get_peft_model_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
-                    text_encoder_two_lora_layers_to_save = text_encoder_lora_state_dict(model)
+                    text_encoder_two_lora_layers_to_save = get_peft_model_state_dict(model)
                 else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 
@@ -1130,6 +1074,12 @@ def main(args):
             args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
         )
 
+    unet_lora_parameters = list(filter(lambda p: p.requires_grad, unet.parameters()))
+
+    if args.train_text_encoder:
+        text_lora_parameters_one = list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
+        text_lora_parameters_two = list(filter(lambda p: p.requires_grad, text_encoder_two.parameters()))
+
     # Optimization parameters
     unet_lora_parameters_with_lr = {"params": unet_lora_parameters, "lr": args.learning_rate}
     if args.train_text_encoder:
@@ -1194,26 +1144,10 @@ def main(args):
 
         optimizer_class = prodigyopt.Prodigy
 
-        if args.learning_rate <= 0.1:
-            logger.warn(
-                "Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
-            )
-        if args.train_text_encoder and args.text_encoder_lr:
-            logger.warn(
-                f"Learning rates were provided both for the unet and the text encoder- e.g. text_encoder_lr:"
-                f" {args.text_encoder_lr} and learning_rate: {args.learning_rate}. "
-                f"When using prodigy only learning_rate is used as the initial learning rate."
-            )
-            # changes the learning rate of text_encoder_parameters_one and text_encoder_parameters_two to be
-            # --learning_rate
-            params_to_optimize[1]["lr"] = args.learning_rate
-            params_to_optimize[2]["lr"] = args.learning_rate
-
         optimizer = optimizer_class(
             params_to_optimize,
             lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
-            beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,
             eps=args.adam_epsilon,
             decouple=args.prodigy_decouple,
@@ -1659,13 +1593,13 @@ def main(args):
     if accelerator.is_main_process:
         unet = accelerator.unwrap_model(unet)
         unet = unet.to(torch.float32)
-        unet_lora_layers = unet_lora_state_dict(unet)
+        unet_lora_layers = get_peft_model_state_dict(unet)
 
         if args.train_text_encoder:
             text_encoder_one = accelerator.unwrap_model(text_encoder_one)
-            text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder_one.to(torch.float32))
+            text_encoder_lora_layers = get_peft_model_state_dict(text_encoder_one.to(torch.float32))
             text_encoder_two = accelerator.unwrap_model(text_encoder_two)
-            text_encoder_2_lora_layers = text_encoder_lora_state_dict(text_encoder_two.to(torch.float32))
+            text_encoder_2_lora_layers = get_peft_model_state_dict(text_encoder_two.to(torch.float32))
         else:
             text_encoder_lora_layers = None
             text_encoder_2_lora_layers = None

examples/research_projects/README.md

@@ -1,7 +1,7 @@
 # Research projects
 
-This folder contains various research projects using 🧨 Diffusers. 
-They are not really maintained by the core maintainers of this library and often require a specific version of Diffusers that is indicated in the requirements file of each folder. 
+This folder contains various research projects using 🧨 Diffusers.
+They are not really maintained by the core maintainers of this library and often require a specific version of Diffusers that is indicated in the requirements file of each folder.
 Updating them to the most recent version of the library will require some work.
 
 To use any of them, just run the command

examples/research_projects/multi_token_textual_inversion/README.md

@@ -1,6 +1,6 @@
 ## [Deprecated] Multi Token Textual Inversion
 
-**IMPORTART: This research project is deprecated. Multi Token Textual Inversion is now supported natively in [the officail textual inversion example](https://github.com/huggingface/diffusers/tree/main/examples/textual_inversion#running-locally-with-pytorch).**
+**IMPORTART: This research project is deprecated. Multi Token Textual Inversion is now supported natively in [the official textual inversion example](https://github.com/huggingface/diffusers/tree/main/examples/textual_inversion#running-locally-with-pytorch).**
 
 The author of this project is [Isamu Isozaki](https://github.com/isamu-isozaki) - please make sure to tag the author for issue and PRs as well as @patrickvonplaten.
 
@@ -17,9 +17,9 @@ Feel free to add these options to your training! In practice num_vec_per_token a
 [Textual inversion](https://arxiv.org/abs/2208.01618) is a method to personalize text2image models like stable diffusion on your own images using just 3-5 examples.
 The `textual_inversion.py` script shows how to implement the training procedure and adapt it for stable diffusion.
 
-## Running on Colab 
+## Running on Colab
 
-Colab for training 
+Colab for training
 [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb)
 
 Colab for inference
@@ -53,7 +53,7 @@ accelerate config
 
 ### Cat toy example
 
-You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-5`, so you'll need to visit [its card](https://huggingface.co/runwayml/stable-diffusion-v1-5), read the license and tick the checkbox if you agree. 
+You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-5`, so you'll need to visit [its card](https://huggingface.co/runwayml/stable-diffusion-v1-5), read the license and tick the checkbox if you agree.
 
 You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
 
@@ -63,7 +63,7 @@ Run the following command to authenticate your token
 huggingface-cli login
 ```
 
-If you have already cloned the repo, then you won't need to go through these steps. 
+If you have already cloned the repo, then you won't need to go through these steps.
 
 <br>
 

examples/research_projects/onnxruntime/README.md

@@ -2,4 +2,4 @@
 
 **This research project is not actively maintained by the diffusers team. For any questions or comments, please contact Prathik Rao (prathikr), Sunghoon Choi (hanbitmyths), Ashwini Khade (askhade), or Peng Wang (pengwa) on github with any questions.**
 
-This aims to provide diffusers examples with ONNXRuntime optimizations for training/fine-tuning unconditional image generation, text to image, and textual inversion. Please see individual directories for more details on how to run each task using ONNXRuntime.
+This aims to provide diffusers examples with ONNXRuntime optimizations for training/fine-tuning unconditional image generation, text to image, and textual inversion. Please see individual directories for more details on how to run each task using ONNXRuntime.

examples/research_projects/onnxruntime/text_to_image/README.md

@@ -34,7 +34,7 @@ accelerate config
 
 ### Pokemon example
 
-You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree. 
+You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree.
 
 You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
 
@@ -68,7 +68,7 @@ accelerate launch --mixed_precision="fp16"  train_text_to_image.py \
   --learning_rate=1e-05 \
   --max_grad_norm=1 \
   --lr_scheduler="constant" --lr_warmup_steps=0 \
-  --output_dir="sd-pokemon-model" 
+  --output_dir="sd-pokemon-model"
 ```
 
 Please contact Prathik Rao (prathikr), Sunghoon Choi (hanbitmyths), Ashwini Khade (askhade), or Peng Wang (pengwa) on github with any questions.

examples/research_projects/onnxruntime/textual_inversion/README.md

@@ -3,9 +3,9 @@
 [Textual inversion](https://arxiv.org/abs/2208.01618) is a method to personalize text2image models like stable diffusion on your own images using just 3-5 examples.
 The `textual_inversion.py` script shows how to implement the training procedure and adapt it for stable diffusion.
 
-## Running on Colab 
+## Running on Colab
 
-Colab for training 
+Colab for training
 [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb)
 
 Colab for inference
@@ -39,7 +39,7 @@ accelerate config
 
 ### Cat toy example
 
-You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-5`, so you'll need to visit [its card](https://huggingface.co/runwayml/stable-diffusion-v1-5), read the license and tick the checkbox if you agree. 
+You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-5`, so you'll need to visit [its card](https://huggingface.co/runwayml/stable-diffusion-v1-5), read the license and tick the checkbox if you agree.
 
 You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
 
@@ -49,7 +49,7 @@ Run the following command to authenticate your token
 huggingface-cli login
 ```
 
-If you have already cloned the repo, then you won't need to go through these steps. 
+If you have already cloned the repo, then you won't need to go through these steps.
 
 <br>
 

examples/research_projects/realfill/README.md

@@ -35,7 +35,7 @@ from accelerate.utils import write_basic_config
 write_basic_config()
 ```
 
-When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups. 
+When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
 
 ### Toy example
 

examples/research_projects/sdxl_flax/README.md

@@ -72,8 +72,8 @@ params = jax.tree_util.tree_map(lambda x: x.astype(jnp.bfloat16), params)
 params["scheduler"] = scheduler_state
 ```
 This section adjusts the data types of the model parameters.
-We convert all parameters to `bfloat16` to speed-up the computation with model weights. 
-**Note** that the scheduler parameters are **not** converted to `blfoat16` as the loss 
+We convert all parameters to `bfloat16` to speed-up the computation with model weights.
+**Note** that the scheduler parameters are **not** converted to `blfoat16` as the loss
 in precision is degrading the pipeline's performance too significantly.
 
 **3. Define Inputs to Pipeline**
@@ -146,12 +146,12 @@ For this we will be using a JAX feature called [Ahead of Time](https://jax.readt
 
 In [sdxl_single_aot.py](./sdxl_single_aot.py) we give a simple example of how to write our own parallelization logic for text-to-image generation pipeline in JAX using [StabilityAI's Stable Diffusion XL](stabilityai/stable-diffusion-xl-base-1.0)
 
-We add a `aot_compile` function that compiles the `pipeline._generate` function 
+We add a `aot_compile` function that compiles the `pipeline._generate` function
 telling JAX which input arguments are static, that is, arguments that
-are known at compile time and won't change. In our case, it is num_inference_steps, 
+are known at compile time and won't change. In our case, it is num_inference_steps,
 height, width and return_latents.
 
-Once the function is compiled, these parameters are omitted from future calls and 
+Once the function is compiled, these parameters are omitted from future calls and
 cannot be changed without modifying the code and recompiling.
 
 ```python
@@ -205,9 +205,9 @@ def generate(
     g = jnp.array([guidance_scale] * prompt_ids.shape[0], dtype=jnp.float32)
     g = g[:, None]
     images = p_generate(
-        prompt_ids, 
-        p_params, 
-        rng, 
+        prompt_ids,
+        p_params,
+        rng,
         g,
         None,
         neg_prompt_ids)
@@ -220,7 +220,7 @@ def generate(
 The first forward pass after AOT compilation still takes a while longer than
 subsequent passes, this is because on the first pass, JAX uses Python dispatch, which
 Fills the C++ dispatch cache.
-When using jit, this extra step is done automatically, but when using AOT compilation, 
+When using jit, this extra step is done automatically, but when using AOT compilation,
 it doesn't happen until the function call is made.
 
 ```python

examples/t2i_adapter/README_sdxl.md

@@ -42,7 +42,7 @@ from accelerate.utils import write_basic_config
 write_basic_config()
 ```
 
-When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups. 
+When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
 
 ## Circle filling dataset
 
@@ -85,7 +85,7 @@ accelerate launch train_t2i_adapter_sdxl.py \
 To better track our training experiments, we're using the following flags in the command above:
 
 * `report_to="wandb` will ensure the training runs are tracked on Weights and Biases. To use it, be sure to install `wandb` with `pip install wandb`.
-* `validation_image`, `validation_prompt`, and `validation_steps` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected. 
+* `validation_image`, `validation_prompt`, and `validation_steps` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
 
 Our experiments were conducted on a single 40GB A100 GPU.
 

examples/text_to_image/README.md

@@ -32,9 +32,11 @@ And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) e
 accelerate config
 ```
 
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
+
 ### Pokemon example
 
-You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree. 
+You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree.
 
 You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
 
@@ -69,7 +71,7 @@ accelerate launch --mixed_precision="fp16"  train_text_to_image.py \
   --learning_rate=1e-05 \
   --max_grad_norm=1 \
   --lr_scheduler="constant" --lr_warmup_steps=0 \
-  --output_dir="sd-pokemon-model" 
+  --output_dir="sd-pokemon-model"
 ```
 <!-- accelerate_snippet_end -->
 
@@ -99,8 +101,8 @@ accelerate launch --mixed_precision="fp16" train_text_to_image.py \
 
 Once the training is finished the model will be saved in the `output_dir` specified in the command. In this example it's `sd-pokemon-model`. To load the fine-tuned model for inference just pass that path to `StableDiffusionPipeline`
 
-
 ```python
+import torch
 from diffusers import StableDiffusionPipeline
 
 model_path = "path_to_saved_model"
@@ -112,12 +114,13 @@ image.save("yoda-pokemon.png")
 ```
 
 Checkpoints only save the unet, so to run inference from a checkpoint, just load the unet
+
 ```python
+import torch
 from diffusers import StableDiffusionPipeline, UNet2DConditionModel
 
 model_path = "path_to_saved_model"
-
-unet = UNet2DConditionModel.from_pretrained(model_path + "/checkpoint-<N>/unet")
+unet = UNet2DConditionModel.from_pretrained(model_path + "/checkpoint-<N>/unet", torch_dtype=torch.float16)
 
 pipe = StableDiffusionPipeline.from_pretrained("<initial model>", unet=unet, torch_dtype=torch.float16)
 pipe.to("cuda")
@@ -143,11 +146,11 @@ accelerate launch --mixed_precision="fp16" --multi_gpu  train_text_to_image.py \
   --train_batch_size=1 \
   --gradient_accumulation_steps=4 \
   --gradient_checkpointing \
-  --max_train_steps=15000 \ 
+  --max_train_steps=15000 \
   --learning_rate=1e-05 \
   --max_grad_norm=1 \
   --lr_scheduler="constant" --lr_warmup_steps=0 \
-  --output_dir="sd-pokemon-model" 
+  --output_dir="sd-pokemon-model"
 ```
 
 
@@ -155,7 +158,7 @@ accelerate launch --mixed_precision="fp16" --multi_gpu  train_text_to_image.py \
 
 We support training with the Min-SNR weighting strategy proposed in [Efficient Diffusion Training via Min-SNR Weighting Strategy](https://arxiv.org/abs/2303.09556) which helps to achieve faster convergence
 by rebalancing the loss. In order to use it, one needs to set the `--snr_gamma` argument. The recommended
-value when using it is 5.0. 
+value when using it is 5.0.
 
 You can find [this project on Weights and Biases](https://wandb.ai/sayakpaul/text2image-finetune-minsnr) that compares the loss surfaces of the following setups:
 
@@ -165,7 +168,7 @@ You can find [this project on Weights and Biases](https://wandb.ai/sayakpaul/tex
 
 For our small Pokemons dataset, the effects of Min-SNR weighting strategy might not appear to be pronounced, but for larger datasets, we believe the effects will be more pronounced.
 
-Also, note that in this example, we either predict `epsilon` (i.e., the noise) or the `v_prediction`. For both of these cases, the formulation of the Min-SNR weighting strategy that we have used holds. 
+Also, note that in this example, we either predict `epsilon` (i.e., the noise) or the `v_prediction`. For both of these cases, the formulation of the Min-SNR weighting strategy that we have used holds.
 
 ## Training with LoRA
 
@@ -184,7 +187,7 @@ on consumer GPUs like Tesla T4, Tesla V100.
 
 ### Training
 
-First, you need to set up your development environment as is explained in the [installation section](#installing-the-dependencies). Make sure to set the `MODEL_NAME` and `DATASET_NAME` environment variables. Here, we will use [Stable Diffusion v1-4](https://hf.co/CompVis/stable-diffusion-v1-4) and the [Pokemons dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions).  
+First, you need to set up your development environment as is explained in the [installation section](#installing-the-dependencies). Make sure to set the `MODEL_NAME` and `DATASET_NAME` environment variables. Here, we will use [Stable Diffusion v1-4](https://hf.co/CompVis/stable-diffusion-v1-4) and the [Pokemons dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions).
 
 **___Note: Change the `resolution` to 768 if you are using the [stable-diffusion-2](https://huggingface.co/stabilityai/stable-diffusion-2) 768x768 model.___**
 
@@ -195,7 +198,7 @@ export MODEL_NAME="CompVis/stable-diffusion-v1-4"
 export DATASET_NAME="lambdalabs/pokemon-blip-captions"
 ```
 
-For this example we want to directly store the trained LoRA embeddings on the Hub, so 
+For this example we want to directly store the trained LoRA embeddings on the Hub, so
 we need to be logged in and add the `--push_to_hub` flag.
 
 ```bash
@@ -223,11 +226,11 @@ The above command will also run inference as fine-tuning progresses and log the
 
 The final LoRA embedding weights have been uploaded to [sayakpaul/sd-model-finetuned-lora-t4](https://huggingface.co/sayakpaul/sd-model-finetuned-lora-t4). **___Note: [The final weights](https://huggingface.co/sayakpaul/sd-model-finetuned-lora-t4/blob/main/pytorch_lora_weights.bin) are only 3 MB in size, which is orders of magnitudes smaller than the original model.___**
 
-You can check some inference samples that were logged during the course of the fine-tuning process [here](https://wandb.ai/sayakpaul/text2image-fine-tune/runs/q4lc0xsw). 
+You can check some inference samples that were logged during the course of the fine-tuning process [here](https://wandb.ai/sayakpaul/text2image-fine-tune/runs/q4lc0xsw).
 
 ### Inference
 
-Once you have trained a model using above command, the inference can be done simply using the `StableDiffusionPipeline` after loading the trained LoRA weights.  You 
+Once you have trained a model using above command, the inference can be done simply using the `StableDiffusionPipeline` after loading the trained LoRA weights.  You
 need to pass the `output_dir` for loading the LoRA weights which, in this case, is `sd-pokemon-model-lora`.
 
 ```python
@@ -246,9 +249,9 @@ image.save("pokemon.png")
 
 If you are loading the LoRA parameters from the Hub and if the Hub repository has
 a `base_model` tag (such as [this](https://huggingface.co/sayakpaul/sd-model-finetuned-lora-t4/blob/main/README.md?code=true#L4)), then
-you can do: 
+you can do:
 
-```py 
+```py
 from huggingface_hub.repocard import RepoCard
 
 lora_model_id = "sayakpaul/sd-model-finetuned-lora-t4"
@@ -285,7 +288,7 @@ python train_text_to_image_flax.py \
   --max_train_steps=15000 \
   --learning_rate=1e-05 \
   --max_grad_norm=1 \
-  --output_dir="sd-pokemon-model" 
+  --output_dir="sd-pokemon-model"
 ```
 
 To run on your own training files prepare the dataset according to the format required by `datasets`, you can find the instructions for how to do that in this [document](https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder-with-metadata).
@@ -319,5 +322,5 @@ According to [this issue](https://github.com/huggingface/diffusers/issues/2234#i
 
 ## Stable Diffusion XL
 
-* We support fine-tuning the UNet shipped in [Stable Diffusion XL](https://huggingface.co/papers/2307.01952) via the `train_text_to_image_sdxl.py` script. Please refer to the docs [here](./README_sdxl.md). 
-* We also support fine-tuning of the UNet and Text Encoder shipped in [Stable Diffusion XL](https://huggingface.co/papers/2307.01952) with LoRA via the `train_text_to_image_lora_sdxl.py` script. Please refer to the docs [here](./README_sdxl.md). 
+* We support fine-tuning the UNet shipped in [Stable Diffusion XL](https://huggingface.co/papers/2307.01952) via the `train_text_to_image_sdxl.py` script. Please refer to the docs [here](./README_sdxl.md).
+* We also support fine-tuning of the UNet and Text Encoder shipped in [Stable Diffusion XL](https://huggingface.co/papers/2307.01952) with LoRA via the `train_text_to_image_lora_sdxl.py` script. Please refer to the docs [here](./README_sdxl.md).

examples/text_to_image/README_sdxl.md

@@ -45,6 +45,7 @@ write_basic_config()
 ```
 
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
 
 ### Training
 

examples/text_to_image/requirements.txt

@@ -5,3 +5,4 @@ datasets
 ftfy
 tensorboard
 Jinja2
+peft==0.7.0

examples/text_to_image/requirements_sdxl.txt

@@ -5,3 +5,4 @@ ftfy
 tensorboard
 Jinja2
 datasets
+peft==0.7.0

examples/text_to_image/train_text_to_image_lora.py

@@ -34,13 +34,14 @@ from accelerate.utils import ProjectConfiguration, set_seed
 from datasets import load_dataset
 from huggingface_hub import create_repo, upload_folder
 from packaging import version
+from peft import LoraConfig
+from peft.utils import get_peft_model_state_dict
 from torchvision import transforms
 from tqdm.auto import tqdm
 from transformers import CLIPTextModel, CLIPTokenizer
 
 import diffusers
-from diffusers import AutoencoderKL, DDPMScheduler, DiffusionPipeline, UNet2DConditionModel
-from diffusers.models.lora import LoRALinearLayer
+from diffusers import AutoencoderKL, DDPMScheduler, DiffusionPipeline, StableDiffusionPipeline, UNet2DConditionModel
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_snr
 from diffusers.utils import check_min_version, is_wandb_available
@@ -53,39 +54,6 @@ check_min_version("0.25.0.dev0")
 logger = get_logger(__name__, log_level="INFO")
 
 
-# TODO: This function should be removed once training scripts are rewritten in PEFT
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    def text_encoder_attn_modules(text_encoder):
-        from transformers import CLIPTextModel, CLIPTextModelWithProjection
-
-        attn_modules = []
-
-        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-                name = f"text_model.encoder.layers.{i}.self_attn"
-                mod = layer.self_attn
-                attn_modules.append((name, mod))
-
-        return attn_modules
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
 def save_model_card(repo_id: str, images=None, base_model=str, dataset_name=str, repo_folder=None):
     img_str = ""
     for i, image in enumerate(images):
@@ -479,62 +447,26 @@ def main():
     elif accelerator.mixed_precision == "bf16":
         weight_dtype = torch.bfloat16
 
+    # Freeze the unet parameters before adding adapters
+    for param in unet.parameters():
+        param.requires_grad_(False)
+
+    unet_lora_config = LoraConfig(
+        r=args.rank, init_lora_weights="gaussian", target_modules=["to_k", "to_q", "to_v", "to_out.0"]
+    )
+
     # Move unet, vae and text_encoder to device and cast to weight_dtype
     unet.to(accelerator.device, dtype=weight_dtype)
     vae.to(accelerator.device, dtype=weight_dtype)
     text_encoder.to(accelerator.device, dtype=weight_dtype)
 
-    # now we will add new LoRA weights to the attention layers
-    # It's important to realize here how many attention weights will be added and of which sizes
-    # The sizes of the attention layers consist only of two different variables:
-    # 1) - the "hidden_size", which is increased according to `unet.config.block_out_channels`.
-    # 2) - the "cross attention size", which is set to `unet.config.cross_attention_dim`.
-
-    # Let's first see how many attention processors we will have to set.
-    # For Stable Diffusion, it should be equal to:
-    # - down blocks (2x attention layers) * (2x transformer layers) * (3x down blocks) = 12
-    # - mid blocks (2x attention layers) * (1x transformer layers) * (1x mid blocks) = 2
-    # - up blocks (2x attention layers) * (3x transformer layers) * (3x down blocks) = 18
-    # => 32 layers
-
-    # Set correct lora layers
-    unet_lora_parameters = []
-    for attn_processor_name, attn_processor in unet.attn_processors.items():
-        # Parse the attention module.
-        attn_module = unet
-        for n in attn_processor_name.split(".")[:-1]:
-            attn_module = getattr(attn_module, n)
-
-        # Set the `lora_layer` attribute of the attention-related matrices.
-        attn_module.to_q.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_k.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
-            )
-        )
-
-        attn_module.to_v.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_out[0].set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_out[0].in_features,
-                out_features=attn_module.to_out[0].out_features,
-                rank=args.rank,
-            )
-        )
-
-        # Accumulate the LoRA params to optimize.
-        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
+    # Add adapter and make sure the trainable params are in float32.
+    unet.add_adapter(unet_lora_config)
+    if args.mixed_precision == "fp16":
+        for param in unet.parameters():
+            # only upcast trainable parameters (LoRA) into fp32
+            if param.requires_grad:
+                param.data = param.to(torch.float32)
 
     if args.enable_xformers_memory_efficient_attention:
         if is_xformers_available():
@@ -549,6 +481,8 @@ def main():
         else:
             raise ValueError("xformers is not available. Make sure it is installed correctly")
 
+    lora_layers = filter(lambda p: p.requires_grad, unet.parameters())
+
     # Enable TF32 for faster training on Ampere GPUs,
     # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
     if args.allow_tf32:
@@ -573,7 +507,7 @@ def main():
         optimizer_cls = torch.optim.AdamW
 
     optimizer = optimizer_cls(
-        unet_lora_parameters,
+        lora_layers,
         lr=args.learning_rate,
         betas=(args.adam_beta1, args.adam_beta2),
         weight_decay=args.adam_weight_decay,
@@ -700,8 +634,8 @@ def main():
     )
 
     # Prepare everything with our `accelerator`.
-    unet_lora_parameters, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        unet_lora_parameters, optimizer, train_dataloader, lr_scheduler
+    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        unet, optimizer, train_dataloader, lr_scheduler
     )
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
@@ -833,7 +767,7 @@ def main():
                 # Backpropagate
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    params_to_clip = unet_lora_parameters
+                    params_to_clip = lora_layers
                     accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                 optimizer.step()
                 lr_scheduler.step()
@@ -870,6 +804,16 @@ def main():
 
                         save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                         accelerator.save_state(save_path)
+
+                        unwrapped_unet = accelerator.unwrap_model(unet)
+                        unet_lora_state_dict = get_peft_model_state_dict(unwrapped_unet)
+
+                        StableDiffusionPipeline.save_lora_weights(
+                            save_directory=save_path,
+                            unet_lora_layers=unet_lora_state_dict,
+                            safe_serialization=True,
+                        )
+
                         logger.info(f"Saved state to {save_path}")
 
             logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
@@ -926,7 +870,14 @@ def main():
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         unet = unet.to(torch.float32)
-        unet.save_attn_procs(args.output_dir)
+
+        unwrapped_unet = accelerator.unwrap_model(unet)
+        unet_lora_state_dict = get_peft_model_state_dict(unwrapped_unet)
+        StableDiffusionPipeline.save_lora_weights(
+            save_directory=args.output_dir,
+            unet_lora_layers=unet_lora_state_dict,
+            safe_serialization=True,
+        )
 
         if args.push_to_hub:
             save_model_card(
@@ -943,39 +894,42 @@ def main():
                 ignore_patterns=["step_*", "epoch_*"],
             )
 
-    # Final inference
-    # Load previous pipeline
-    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)
+        # 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.unet.load_attn_procs(args.output_dir)
+            # 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 = []
-    for _ in range(args.num_validation_images):
-        images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0])
+            # run inference
+            generator = torch.Generator(device=accelerator.device)
+            if args.seed is not None:
+                generator = generator.manual_seed(args.seed)
+            images = []
+            for _ in range(args.num_validation_images):
+                images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0])
 
-    if accelerator.is_main_process:
-        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)
-                            ]
-                        }
-                    )
+            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

@@ -16,7 +16,6 @@
 """Fine-tuning script for Stable Diffusion XL for text2image with support for LoRA."""
 
 import argparse
-import itertools
 import logging
 import math
 import os
@@ -37,6 +36,8 @@ from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration
 from datasets import load_dataset
 from huggingface_hub import create_repo, upload_folder
 from packaging import version
+from peft import LoraConfig
+from peft.utils import get_peft_model_state_dict
 from torchvision import transforms
 from torchvision.transforms.functional import crop
 from tqdm.auto import tqdm
@@ -50,7 +51,6 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.loaders import LoraLoaderMixin
-from diffusers.models.lora import LoRALinearLayer
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_snr
 from diffusers.utils import check_min_version, is_wandb_available
@@ -63,39 +63,6 @@ check_min_version("0.25.0.dev0")
 logger = get_logger(__name__)
 
 
-# TODO: This function should be removed once training scripts are rewritten in PEFT
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    def text_encoder_attn_modules(text_encoder):
-        from transformers import CLIPTextModel, CLIPTextModelWithProjection
-
-        attn_modules = []
-
-        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-                name = f"text_model.encoder.layers.{i}.self_attn"
-                mod = layer.self_attn
-                attn_modules.append((name, mod))
-
-        return attn_modules
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
 def save_model_card(
     repo_id: str,
     images=None,
@@ -658,53 +625,20 @@ def main(args):
 
     # now we will add new LoRA weights to the attention layers
     # Set correct lora layers
-    unet_lora_parameters = []
-    for attn_processor_name, attn_processor in unet.attn_processors.items():
-        # Parse the attention module.
-        attn_module = unet
-        for n in attn_processor_name.split(".")[:-1]:
-            attn_module = getattr(attn_module, n)
+    unet_lora_config = LoraConfig(
+        r=args.rank, init_lora_weights="gaussian", target_modules=["to_k", "to_q", "to_v", "to_out.0"]
+    )
 
-        # Set the `lora_layer` attribute of the attention-related matrices.
-        attn_module.to_q.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_k.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_v.set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
-            )
-        )
-        attn_module.to_out[0].set_lora_layer(
-            LoRALinearLayer(
-                in_features=attn_module.to_out[0].in_features,
-                out_features=attn_module.to_out[0].out_features,
-                rank=args.rank,
-            )
-        )
+    unet.add_adapter(unet_lora_config)
 
-        # Accumulate the LoRA params to optimize.
-        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
-        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
-
-    # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
-    # So, instead, we monkey-patch the forward calls of its attention-blocks.
+    # The text encoder comes from 🤗 transformers, we will also attach adapters to it.
     if args.train_text_encoder:
         # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-        text_lora_parameters_one = LoraLoaderMixin._modify_text_encoder(
-            text_encoder_one, dtype=torch.float32, rank=args.rank
-        )
-        text_lora_parameters_two = LoraLoaderMixin._modify_text_encoder(
-            text_encoder_two, dtype=torch.float32, rank=args.rank
+        text_lora_config = LoraConfig(
+            r=args.rank, init_lora_weights="gaussian", target_modules=["q_proj", "k_proj", "v_proj", "out_proj"]
         )
+        text_encoder_one.add_adapter(text_lora_config)
+        text_encoder_two.add_adapter(text_lora_config)
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
@@ -717,11 +651,11 @@ def main(args):
 
             for model in models:
                 if isinstance(model, type(accelerator.unwrap_model(unet))):
-                    unet_lora_layers_to_save = unet_attn_processors_state_dict(model)
+                    unet_lora_layers_to_save = get_peft_model_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
-                    text_encoder_one_lora_layers_to_save = text_encoder_lora_state_dict(model)
+                    text_encoder_one_lora_layers_to_save = get_peft_model_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
-                    text_encoder_two_lora_layers_to_save = text_encoder_lora_state_dict(model)
+                    text_encoder_two_lora_layers_to_save = get_peft_model_state_dict(model)
                 else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 
@@ -792,11 +726,13 @@ def main(args):
         optimizer_class = torch.optim.AdamW
 
     # Optimizer creation
-    params_to_optimize = (
-        itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
-        if args.train_text_encoder
-        else unet_lora_parameters
-    )
+    params_to_optimize = list(filter(lambda p: p.requires_grad, unet.parameters()))
+    if args.train_text_encoder:
+        params_to_optimize = (
+            params_to_optimize
+            + list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
+            + list(filter(lambda p: p.requires_grad, text_encoder_two.parameters()))
+        )
     optimizer = optimizer_class(
         params_to_optimize,
         lr=args.learning_rate,
@@ -1128,12 +1064,7 @@ def main(args):
                 # Backpropagate
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    params_to_clip = (
-                        itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
-                        if args.train_text_encoder
-                        else unet_lora_parameters
-                    )
-                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+                    accelerator.clip_grad_norm_(params_to_optimize, args.max_grad_norm)
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()
@@ -1229,20 +1160,21 @@ def main(args):
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         unet = accelerator.unwrap_model(unet)
-        unet_lora_layers = unet_attn_processors_state_dict(unet)
+        unet_lora_state_dict = get_peft_model_state_dict(unet)
 
         if args.train_text_encoder:
             text_encoder_one = accelerator.unwrap_model(text_encoder_one)
-            text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder_one)
             text_encoder_two = accelerator.unwrap_model(text_encoder_two)
-            text_encoder_2_lora_layers = text_encoder_lora_state_dict(text_encoder_two)
+
+            text_encoder_lora_layers = get_peft_model_state_dict(text_encoder_one)
+            text_encoder_2_lora_layers = get_peft_model_state_dict(text_encoder_two)
         else:
             text_encoder_lora_layers = None
             text_encoder_2_lora_layers = None
 
         StableDiffusionXLPipeline.save_lora_weights(
             save_directory=args.output_dir,
-            unet_lora_layers=unet_lora_layers,
+            unet_lora_layers=unet_lora_state_dict,
             text_encoder_lora_layers=text_encoder_lora_layers,
             text_encoder_2_lora_layers=text_encoder_2_lora_layers,
         )

examples/textual_inversion/README.md

@@ -3,9 +3,9 @@
 [Textual inversion](https://arxiv.org/abs/2208.01618) is a method to personalize text2image models like stable diffusion on your own images using just 3-5 examples.
 The `textual_inversion.py` script shows how to implement the training procedure and adapt it for stable diffusion.
 
-## Running on Colab 
+## Running on Colab
 
-Colab for training 
+Colab for training
 [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb)
 
 Colab for inference
@@ -84,11 +84,11 @@ accelerate launch textual_inversion.py \
 
 A full training run takes ~1 hour on one V100 GPU.
 
-**Note**: As described in [the official paper](https://arxiv.org/abs/2208.01618) 
+**Note**: As described in [the official paper](https://arxiv.org/abs/2208.01618)
 only one embedding vector is used for the placeholder token, *e.g.* `"<cat-toy>"`.
-However, one can also add multiple embedding vectors for the placeholder token 
-to increase the number of fine-tuneable parameters. This can help the model to learn 
-more complex details. To use multiple embedding vectors, you should define `--num_vectors` 
+However, one can also add multiple embedding vectors for the placeholder token
+to increase the number of fine-tuneable parameters. This can help the model to learn
+more complex details. To use multiple embedding vectors, you should define `--num_vectors`
 to a number larger than one, *e.g.*:
 ```bash
 --num_vectors 5

examples/unconditional_image_generation/README.md

@@ -27,7 +27,7 @@ And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) e
 accelerate config
 ```
 
-### Unconditional Flowers  
+### Unconditional Flowers
 
 The command to train a DDPM UNet model on the Oxford Flowers dataset:
 
@@ -52,7 +52,7 @@ A full training run takes 2 hours on 4xV100 GPUs.
 <img src="https://user-images.githubusercontent.com/26864830/180248660-a0b143d0-b89a-42c5-8656-2ebf6ece7e52.png" width="700" />
 
 
-### Unconditional Pokemon 
+### Unconditional Pokemon
 
 The command to train a DDPM UNet model on the Pokemon dataset:
 
@@ -96,7 +96,7 @@ accelerate launch --mixed_precision="fp16" --multi_gpu train_unconditional.py \
   --logger="wandb"
 ```
 
-To be able to use Weights and Biases (`wandb`) as a logger you need to install the library: `pip install wandb`. 
+To be able to use Weights and Biases (`wandb`) as a logger you need to install the library: `pip install wandb`.
 
 ### Using your own data
 

examples/wuerstchen/text_to_image/README.md

@@ -72,12 +72,12 @@ In a nutshell, LoRA allows adapting pretrained models by adding pairs of rank-de
 
 ### Prior Training
 
-First, you need to set up your development environment as explained in the [installation](#Running-locally-with-PyTorch) section. Make sure to set the `DATASET_NAME` environment variable. Here, we will use the [Pokemon captions dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions).  
+First, you need to set up your development environment as explained in the [installation](#Running-locally-with-PyTorch) section. Make sure to set the `DATASET_NAME` environment variable. Here, we will use the [Pokemon captions dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions).
 
 ```bash
 export DATASET_NAME="lambdalabs/pokemon-blip-captions"
 
-accelerate launch train_text_to_image_prior_lora.py \
+accelerate launch train_text_to_image_lora_prior.py \
   --mixed_precision="fp16" \
   --dataset_name=$DATASET_NAME --caption_column="text" \
   --resolution=768 \

scripts/convert_consistency_decoder.py

@@ -12,9 +12,9 @@ from safetensors.torch import load_file as stl
 from tqdm import tqdm
 
 from diffusers import AutoencoderKL, ConsistencyDecoderVAE, DiffusionPipeline, StableDiffusionPipeline, UNet2DModel
+from diffusers.models.autoencoders.vae import Encoder
 from diffusers.models.embeddings import TimestepEmbedding
 from diffusers.models.unet_2d_blocks import ResnetDownsampleBlock2D, ResnetUpsampleBlock2D, UNetMidBlock2D
-from diffusers.models.vae import Encoder
 
 
 args = ArgumentParser()

scripts/convert_diffusers_to_original_stable_diffusion.py

@@ -159,6 +159,14 @@ vae_conversion_map_attn = [
     ("proj_out.", "proj_attn."),
 ]
 
+# This is probably not the most ideal solution, but it does work.
+vae_extra_conversion_map = [
+    ("to_q", "q"),
+    ("to_k", "k"),
+    ("to_v", "v"),
+    ("to_out.0", "proj_out"),
+]
+
 
 def reshape_weight_for_sd(w):
     # convert HF linear weights to SD conv2d weights
@@ -178,11 +186,20 @@ def convert_vae_state_dict(vae_state_dict):
             mapping[k] = v
     new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
     weights_to_convert = ["q", "k", "v", "proj_out"]
+    keys_to_rename = {}
     for k, v in new_state_dict.items():
         for weight_name in weights_to_convert:
             if f"mid.attn_1.{weight_name}.weight" in k:
                 print(f"Reshaping {k} for SD format")
                 new_state_dict[k] = reshape_weight_for_sd(v)
+        for weight_name, real_weight_name in vae_extra_conversion_map:
+            if f"mid.attn_1.{weight_name}.weight" in k or f"mid.attn_1.{weight_name}.bias" in k:
+                keys_to_rename[k] = k.replace(weight_name, real_weight_name)
+    for k, v in keys_to_rename.items():
+        if k in new_state_dict:
+            print(f"Renaming {k} to {v}")
+            new_state_dict[v] = reshape_weight_for_sd(new_state_dict[k])
+            del new_state_dict[k]
     return new_state_dict
 
 

setup.py

@@ -204,7 +204,7 @@ class DepsTableUpdateCommand(Command):
 extras = {}
 extras["quality"] = deps_list("urllib3", "isort", "ruff", "hf-doc-builder")
 extras["docs"] = deps_list("hf-doc-builder")
-extras["training"] = deps_list("accelerate", "datasets", "protobuf", "tensorboard", "Jinja2")
+extras["training"] = deps_list("accelerate", "datasets", "protobuf", "tensorboard", "Jinja2", "peft")
 extras["test"] = deps_list(
     "compel",
     "GitPython",

src/diffusers/__init__.py

@@ -80,6 +80,7 @@ else:
             "AutoencoderTiny",
             "ConsistencyDecoderVAE",
             "ControlNetModel",
+            "ControlNetXSModel",
             "Kandinsky3UNet",
             "ModelMixin",
             "MotionAdapter",
@@ -250,6 +251,7 @@ else:
             "StableDiffusionControlNetImg2ImgPipeline",
             "StableDiffusionControlNetInpaintPipeline",
             "StableDiffusionControlNetPipeline",
+            "StableDiffusionControlNetXSPipeline",
             "StableDiffusionDepth2ImgPipeline",
             "StableDiffusionDiffEditPipeline",
             "StableDiffusionGLIGENPipeline",
@@ -273,6 +275,7 @@ else:
             "StableDiffusionXLControlNetImg2ImgPipeline",
             "StableDiffusionXLControlNetInpaintPipeline",
             "StableDiffusionXLControlNetPipeline",
+            "StableDiffusionXLControlNetXSPipeline",
             "StableDiffusionXLImg2ImgPipeline",
             "StableDiffusionXLInpaintPipeline",
             "StableDiffusionXLInstructPix2PixPipeline",
@@ -454,6 +457,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderTiny,
             ConsistencyDecoderVAE,
             ControlNetModel,
+            ControlNetXSModel,
             Kandinsky3UNet,
             ModelMixin,
             MotionAdapter,
@@ -603,6 +607,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionControlNetImg2ImgPipeline,
             StableDiffusionControlNetInpaintPipeline,
             StableDiffusionControlNetPipeline,
+            StableDiffusionControlNetXSPipeline,
             StableDiffusionDepth2ImgPipeline,
             StableDiffusionDiffEditPipeline,
             StableDiffusionGLIGENPipeline,
@@ -626,6 +631,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionXLControlNetImg2ImgPipeline,
             StableDiffusionXLControlNetInpaintPipeline,
             StableDiffusionXLControlNetPipeline,
+            StableDiffusionXLControlNetXSPipeline,
             StableDiffusionXLImg2ImgPipeline,
             StableDiffusionXLInpaintPipeline,
             StableDiffusionXLInstructPix2PixPipeline,

src/diffusers/image_processor.py

@@ -88,7 +88,7 @@ class VaeImageProcessor(ConfigMixin):
             self.config.do_convert_rgb = False
 
     @staticmethod
-    def numpy_to_pil(images: np.ndarray) -> PIL.Image.Image:
+    def numpy_to_pil(images: np.ndarray) -> List[PIL.Image.Image]:
         """
         Convert a numpy image or a batch of images to a PIL image.
         """

src/diffusers/loaders/lora.py

@@ -18,6 +18,7 @@ from typing import Callable, Dict, List, Optional, Union
 import safetensors
 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
@@ -229,7 +230,9 @@ class LoraLoaderMixin:
                     # determine `weight_name`.
                     if weight_name is None:
                         weight_name = cls._best_guess_weight_name(
-                            pretrained_model_name_or_path_or_dict, file_extension=".safetensors"
+                            pretrained_model_name_or_path_or_dict,
+                            file_extension=".safetensors",
+                            local_files_only=local_files_only,
                         )
                     model_file = _get_model_file(
                         pretrained_model_name_or_path_or_dict,
@@ -255,7 +258,7 @@ class LoraLoaderMixin:
             if model_file is None:
                 if weight_name is None:
                     weight_name = cls._best_guess_weight_name(
-                        pretrained_model_name_or_path_or_dict, file_extension=".bin"
+                        pretrained_model_name_or_path_or_dict, file_extension=".bin", local_files_only=local_files_only
                     )
                 model_file = _get_model_file(
                     pretrained_model_name_or_path_or_dict,
@@ -294,7 +297,12 @@ class LoraLoaderMixin:
         return state_dict, network_alphas
 
     @classmethod
-    def _best_guess_weight_name(cls, pretrained_model_name_or_path_or_dict, file_extension=".safetensors"):
+    def _best_guess_weight_name(
+        cls, pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False
+    ):
+        if local_files_only or HF_HUB_OFFLINE:
+            raise ValueError("When using the offline mode, you must specify a `weight_name`.")
+
         targeted_files = []
 
         if os.path.isfile(pretrained_model_name_or_path_or_dict):

src/diffusers/loaders/single_file.py

@@ -169,10 +169,12 @@ class FromSingleFileMixin:
         load_safety_checker = kwargs.pop("load_safety_checker", True)
         prediction_type = kwargs.pop("prediction_type", None)
         text_encoder = kwargs.pop("text_encoder", None)
+        text_encoder_2 = kwargs.pop("text_encoder_2", None)
         vae = kwargs.pop("vae", None)
         controlnet = kwargs.pop("controlnet", None)
         adapter = kwargs.pop("adapter", None)
         tokenizer = kwargs.pop("tokenizer", None)
+        tokenizer_2 = kwargs.pop("tokenizer_2", None)
 
         torch_dtype = kwargs.pop("torch_dtype", None)
 
@@ -274,8 +276,10 @@ class FromSingleFileMixin:
             load_safety_checker=load_safety_checker,
             prediction_type=prediction_type,
             text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
             vae=vae,
             tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
             original_config_file=original_config_file,
             config_files=config_files,
             local_files_only=local_files_only,

src/diffusers/loaders/unet.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import os
-from collections import OrderedDict, defaultdict
+from collections import defaultdict
 from contextlib import nullcontext
 from typing import Callable, Dict, List, Optional, Union
 
@@ -22,7 +22,7 @@ import torch.nn.functional as F
 from huggingface_hub.utils import validate_hf_hub_args
 from torch import nn
 
-from ..models.embeddings import ImageProjection, Resampler
+from ..models.embeddings import ImageProjection, MLPProjection, Resampler
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
 from ..utils import (
     USE_PEFT_BACKEND,
@@ -664,6 +664,80 @@ class UNet2DConditionLoadersMixin:
             if hasattr(self, "peft_config"):
                 self.peft_config.pop(adapter_name, None)
 
+    def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict):
+        updated_state_dict = {}
+        image_projection = None
+
+        if "proj.weight" in state_dict:
+            # IP-Adapter
+            num_image_text_embeds = 4
+            clip_embeddings_dim = state_dict["proj.weight"].shape[-1]
+            cross_attention_dim = state_dict["proj.weight"].shape[0] // 4
+
+            image_projection = ImageProjection(
+                cross_attention_dim=cross_attention_dim,
+                image_embed_dim=clip_embeddings_dim,
+                num_image_text_embeds=num_image_text_embeds,
+            )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("proj", "image_embeds")
+                updated_state_dict[diffusers_name] = value
+
+        elif "proj.3.weight" in state_dict:
+            # IP-Adapter Full
+            clip_embeddings_dim = state_dict["proj.0.weight"].shape[0]
+            cross_attention_dim = state_dict["proj.3.weight"].shape[0]
+
+            image_projection = MLPProjection(
+                cross_attention_dim=cross_attention_dim, image_embed_dim=clip_embeddings_dim
+            )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("proj.0", "ff.net.0.proj")
+                diffusers_name = diffusers_name.replace("proj.2", "ff.net.2")
+                diffusers_name = diffusers_name.replace("proj.3", "norm")
+                updated_state_dict[diffusers_name] = value
+
+        else:
+            # IP-Adapter Plus
+            num_image_text_embeds = state_dict["latents"].shape[1]
+            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
+
+            image_projection = Resampler(
+                embed_dims=embed_dims,
+                output_dims=output_dims,
+                hidden_dims=hidden_dims,
+                heads=heads,
+                num_queries=num_image_text_embeds,
+            )
+
+            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:
+                    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_out" in diffusers_name:
+                    updated_state_dict[diffusers_name.replace("to_out", "to_out.0")] = value
+                else:
+                    updated_state_dict[diffusers_name] = value
+
+        image_projection.load_state_dict(updated_state_dict)
+        return image_projection
+
     def _load_ip_adapter_weights(self, state_dict):
         from ..models.attention_processor import (
             AttnProcessor,
@@ -675,6 +749,9 @@ class UNet2DConditionLoadersMixin:
         if "proj.weight" in state_dict["image_proj"]:
             # IP-Adapter
             num_image_text_embeds = 4
+        elif "proj.3.weight" in state_dict["image_proj"]:
+            # IP-Adapter Full Face
+            num_image_text_embeds = 257  # 256 CLIP tokens + 1 CLS token
         else:
             # IP-Adapter Plus
             num_image_text_embeds = state_dict["image_proj"]["latents"].shape[1]
@@ -721,79 +798,8 @@ class UNet2DConditionLoadersMixin:
 
         self.set_attn_processor(attn_procs)
 
-        # create image projection layers.
-        if "proj.weight" in state_dict["image_proj"]:
-            # IP-Adapter
-            clip_embeddings_dim = state_dict["image_proj"]["proj.weight"].shape[-1]
-            cross_attention_dim = state_dict["image_proj"]["proj.weight"].shape[0] // 4
-
-            image_projection = ImageProjection(
-                cross_attention_dim=cross_attention_dim,
-                image_embed_dim=clip_embeddings_dim,
-                num_image_text_embeds=num_image_text_embeds,
-            )
-            image_projection.to(dtype=self.dtype, device=self.device)
-
-            # load image projection layer weights
-            image_proj_state_dict = {}
-            image_proj_state_dict.update(
-                {
-                    "image_embeds.weight": state_dict["image_proj"]["proj.weight"],
-                    "image_embeds.bias": state_dict["image_proj"]["proj.bias"],
-                    "norm.weight": state_dict["image_proj"]["norm.weight"],
-                    "norm.bias": state_dict["image_proj"]["norm.bias"],
-                }
-            )
-
-            image_projection.load_state_dict(image_proj_state_dict)
-
-        else:
-            # IP-Adapter Plus
-            embed_dims = state_dict["image_proj"]["proj_in.weight"].shape[1]
-            output_dims = state_dict["image_proj"]["proj_out.weight"].shape[0]
-            hidden_dims = state_dict["image_proj"]["latents"].shape[2]
-            heads = state_dict["image_proj"]["layers.0.0.to_q.weight"].shape[0] // 64
-
-            image_projection = Resampler(
-                embed_dims=embed_dims,
-                output_dims=output_dims,
-                hidden_dims=hidden_dims,
-                heads=heads,
-                num_queries=num_image_text_embeds,
-            )
-
-            image_proj_state_dict = state_dict["image_proj"]
-
-            new_sd = OrderedDict()
-            for k, v in image_proj_state_dict.items():
-                if "0.to" in k:
-                    k = k.replace("0.to", "2.to")
-                elif "1.0.weight" in k:
-                    k = k.replace("1.0.weight", "3.0.weight")
-                elif "1.0.bias" in k:
-                    k = k.replace("1.0.bias", "3.0.bias")
-                elif "1.1.weight" in k:
-                    k = k.replace("1.1.weight", "3.1.net.0.proj.weight")
-                elif "1.3.weight" in k:
-                    k = k.replace("1.3.weight", "3.1.net.2.weight")
-
-                if "norm1" in k:
-                    new_sd[k.replace("0.norm1", "0")] = v
-                elif "norm2" in k:
-                    new_sd[k.replace("0.norm2", "1")] = v
-                elif "to_kv" in k:
-                    v_chunk = v.chunk(2, dim=0)
-                    new_sd[k.replace("to_kv", "to_k")] = v_chunk[0]
-                    new_sd[k.replace("to_kv", "to_v")] = v_chunk[1]
-                elif "to_out" in k:
-                    new_sd[k.replace("to_out", "to_out.0")] = v
-                else:
-                    new_sd[k] = v
-
-            image_projection.load_state_dict(new_sd)
-            del image_proj_state_dict
+        # convert IP-Adapter Image Projection layers to diffusers
+        image_projection = self._convert_ip_adapter_image_proj_to_diffusers(state_dict["image_proj"])
 
         self.encoder_hid_proj = image_projection.to(device=self.device, dtype=self.dtype)
         self.config.encoder_hid_dim_type = "ip_image_proj"
-
-    delete_adapter_layers

src/diffusers/models/__init__.py

@@ -26,12 +26,14 @@ _import_structure = {}
 
 if is_torch_available():
     _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
-    _import_structure["autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
-    _import_structure["autoencoder_kl"] = ["AutoencoderKL"]
-    _import_structure["autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
-    _import_structure["autoencoder_tiny"] = ["AutoencoderTiny"]
-    _import_structure["consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
+    _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
+    _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
+    _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["controlnet"] = ["ControlNetModel"]
+    _import_structure["controlnet_sparsectrl"] = ["SparseControlNetModel"]
+    _import_structure["controlnetxs"] = ["ControlNetXSModel"]
     _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
     _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
@@ -57,12 +59,15 @@ if is_flax_available():
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     if is_torch_available():
         from .adapter import MultiAdapter, T2IAdapter
-        from .autoencoder_asym_kl import AsymmetricAutoencoderKL
-        from .autoencoder_kl import AutoencoderKL
-        from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
-        from .autoencoder_tiny import AutoencoderTiny
-        from .consistency_decoder_vae import ConsistencyDecoderVAE
+        from .autoencoders import (
+            AsymmetricAutoencoderKL,
+            AutoencoderKL,
+            AutoencoderKLTemporalDecoder,
+            AutoencoderTiny,
+            ConsistencyDecoderVAE,
+        )
         from .controlnet import ControlNetModel
+        from .controlnetxs import ControlNetXSModel
         from .dual_transformer_2d import DualTransformer2DModel
         from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin

src/diffusers/models/attention_processor.py

@@ -113,12 +113,14 @@ class Attention(nn.Module):
     ):
         super().__init__()
         self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
         self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
         self.upcast_attention = upcast_attention
         self.upcast_softmax = upcast_softmax
         self.rescale_output_factor = rescale_output_factor
         self.residual_connection = residual_connection
         self.dropout = dropout
+        self.fused_projections = False
         self.out_dim = out_dim if out_dim is not None else query_dim
 
         # we make use of this private variable to know whether this class is loaded
@@ -180,6 +182,7 @@ class Attention(nn.Module):
         else:
             linear_cls = LoRACompatibleLinear
 
+        self.linear_cls = linear_cls
         self.to_q = linear_cls(query_dim, self.inner_dim, bias=bias)
 
         if not self.only_cross_attention:
@@ -692,6 +695,32 @@ class Attention(nn.Module):
 
         return encoder_hidden_states
 
+    @torch.no_grad()
+    def fuse_projections(self, fuse=True):
+        is_cross_attention = self.cross_attention_dim != self.query_dim
+        device = self.to_q.weight.data.device
+        dtype = self.to_q.weight.data.dtype
+
+        if not is_cross_attention:
+            # fetch weight matrices.
+            concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            # create a new single projection layer and copy over the weights.
+            self.to_qkv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype)
+            self.to_qkv.weight.copy_(concatenated_weights)
+
+        else:
+            concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            self.to_kv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype)
+            self.to_kv.weight.copy_(concatenated_weights)
+
+        self.fused_projections = fuse
+
 
 class AttnProcessor:
     r"""
@@ -1184,9 +1213,6 @@ class AttnProcessor2_0:
         scale: float = 1.0,
     ) -> torch.FloatTensor:
         residual = hidden_states
-
-        args = () if USE_PEFT_BACKEND else (scale,)
-
         if attn.spatial_norm is not None:
             hidden_states = attn.spatial_norm(hidden_states, temb)
 
@@ -1253,6 +1279,103 @@ class AttnProcessor2_0:
         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 fused projection layers. For self-attention modules, all projection matrices (i.e., query,
+    key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+    <Tip warning={true}>
+
+    This API is currently 🧪 experimental in nature and can change in future.
+
+    </Tip>
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "FusedAttnProcessor2_0 requires at least PyTorch 2.0, to use it. Please upgrade PyTorch to > 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        temb: Optional[torch.FloatTensor] = None,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        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)
+
+        args = () if USE_PEFT_BACKEND else (scale,)
+        if encoder_hidden_states is None:
+            qkv = attn.to_qkv(hidden_states, *args)
+            split_size = qkv.shape[-1] // 3
+            query, key, value = torch.split(qkv, split_size, dim=-1)
+        else:
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+            query = attn.to_q(hidden_states, *args)
+
+            kv = attn.to_kv(encoder_hidden_states, *args)
+            split_size = kv.shape[-1] // 2
+            key, value = torch.split(kv, split_size, dim=-1)
+
+        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)
+
+        # 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, *args)
+        # 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 CustomDiffusionXFormersAttnProcessor(nn.Module):
     r"""
     Processor for implementing memory efficient attention using xFormers for the Custom Diffusion method.
@@ -2251,6 +2374,7 @@ CROSS_ATTENTION_PROCESSORS = (
 AttentionProcessor = Union[
     AttnProcessor,
     AttnProcessor2_0,
+    FusedAttnProcessor2_0,
     XFormersAttnProcessor,
     SlicedAttnProcessor,
     AttnAddedKVProcessor,

src/diffusers/models/autoencoders/__init__.py

@@ -0,0 +1,5 @@
+from .autoencoder_asym_kl import AsymmetricAutoencoderKL
+from .autoencoder_kl import AutoencoderKL
+from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
+from .autoencoder_tiny import AutoencoderTiny
+from .consistency_decoder_vae import ConsistencyDecoderVAE

src/diffusers/models/autoencoders/autoencoder_asym_kl.py

@@ -16,10 +16,10 @@ from typing import Optional, Tuple, Union
 import torch
 import torch.nn as nn
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils.accelerate_utils import apply_forward_hook
-from .modeling_outputs import AutoencoderKLOutput
-from .modeling_utils import ModelMixin
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils.accelerate_utils import apply_forward_hook
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
 from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder, MaskConditionDecoder
 
 

src/diffusers/models/autoencoders/autoencoder_kl.py

@@ -16,18 +16,19 @@ from typing import Dict, Optional, Tuple, Union
 import torch
 import torch.nn as nn
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..loaders import FromOriginalVAEMixin
-from ..utils.accelerate_utils import apply_forward_hook
-from .attention_processor import (
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalVAEMixin
+from ...utils.accelerate_utils import apply_forward_hook
+from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
+    Attention,
     AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
 )
-from .modeling_outputs import AutoencoderKLOutput
-from .modeling_utils import ModelMixin
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
 from .vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
 
 
@@ -448,3 +449,41 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
             return (dec,)
 
         return DecoderOutput(sample=dec)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)