Merge branch 'main' into peft-lora-test-fixes

* Add community pipeline: sde_drag.py

* Update README.md

* Update README.md

Update example code and visual example

* Update sde_drag.py

Update code example.

.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

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 +1,297 @@
+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,
+        )
+
+
+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

@@ -264,6 +264,10 @@
       title: ControlNet
     - local: api/pipelines/controlnet_sdxl
       title: ControlNet with Stable Diffusion XL
+    - local: api/pipelines/controlnetxs
+      title: ControlNet-XS
+    - local: api/pipelines/controlnetxs_sdxl
+      title: ControlNet-XS with Stable Diffusion XL
     - local: api/pipelines/cycle_diffusion
       title: Cycle Diffusion
     - local: api/pipelines/dance_diffusion

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/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/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/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/push_to_hub.md

@@ -174,10 +174,4 @@ Set `private=True` in the [`~diffusers.utils.PushToHubMixin.push_to_hub`] functi
 controlnet.push_to_hub("my-controlnet-model-private", private=True)
 ```
 
-Private repositories are only visible to you, and other users won't be able to clone the repository and your repository won't appear in search results. Even if a user has the URL to your private repository, they'll receive a `404 - Sorry, we can't find the page you are looking for.`
-
-To load a model, scheduler, or pipeline from private or gated repositories, set `use_auth_token=True`:
-
-```py
-model = ControlNetModel.from_pretrained("your-namespace/my-controlnet-model-private", use_auth_token=True)
-```
+Private repositories are only visible to you, and other users won't be able to clone the repository and your repository won't appear in search results. Even if a user has the URL to your private repository, they'll receive a `404 - Sorry, we can't find the page you are looking for`. You must be [logged in](https://huggingface.co/docs/huggingface_hub/quick-start#login) to load a model from a private repository.

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

@@ -133,10 +133,10 @@ 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 +145,7 @@ 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
@@ -158,8 +157,10 @@ tags:
 base_model: {base_model}
 instance_prompt: {instance_prompt}
 license: openrail++
+widget:
+    - text: '{validation_prompt if validation_prompt else instance_prompt}'
 ---
-    """
+"""
 
     model_card = f"""
 # SDXL LoRA DreamBooth - {repo_id}
@@ -170,14 +171,6 @@ 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 +189,24 @@ 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 +673,12 @@ 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 +1182,7 @@ 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 +1613,20 @@ 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 +1742,7 @@ 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 +1754,13 @@ 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)
 

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) | - | [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
@@ -512,7 +514,6 @@ device = torch.device('cpu' if not has_cuda else 'cuda')
 pipe = DiffusionPipeline.from_pretrained(
     "CompVis/stable-diffusion-v1-4",
         safety_checker=None,
-    use_auth_token=True,
     custom_pipeline="imagic_stable_diffusion",
     scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
 ).to(device)
@@ -539,7 +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
@@ -552,7 +553,6 @@ device = th.device('cpu' if not has_cuda else 'cuda')
 
 pipe = DiffusionPipeline.from_pretrained(
     "CompVis/stable-diffusion-v1-4",
-    use_auth_token=True,
     custom_pipeline="seed_resize_stable_diffusion"
 ).to(device)
 
@@ -588,7 +588,6 @@ generator = th.Generator("cuda").manual_seed(0)
 
 pipe = DiffusionPipeline.from_pretrained(
     "CompVis/stable-diffusion-v1-4",
-    use_auth_token=True,
     custom_pipeline="/home/mark/open_source/diffusers/examples/community/"
 ).to(device)
 
@@ -607,7 +606,6 @@ image.save('./seed_resize/seed_resize_{w}_{h}_image.png'.format(w=width, h=heigh
 
 pipe_compare = DiffusionPipeline.from_pretrained(
     "CompVis/stable-diffusion-v1-4",
-    use_auth_token=True,
     custom_pipeline="/home/mark/open_source/diffusers/examples/community/"
 ).to(device)
 
@@ -670,14 +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"]
@@ -718,7 +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")
@@ -761,8 +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
@@ -840,8 +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")
 
@@ -864,16 +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)
 
@@ -940,8 +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,
@@ -1052,7 +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
@@ -1089,7 +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
@@ -1130,8 +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.
@@ -1173,7 +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.
@@ -1325,8 +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,
 )
 
@@ -1540,7 +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
 ```
 
@@ -1548,7 +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()'
 ```
@@ -1607,21 +1605,21 @@ 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
@@ -1642,8 +1640,8 @@ neg_prompt = "blur, low quality, carton, animate"
 
 pipe.to("cuda")
 images = pipe(
-    prompt                  = prompt 
-    , negative_prompt       = neg_prompt 
+    prompt                  = prompt
+    , negative_prompt       = neg_prompt
 ).images[0]
 
 pipe.to("cpu")
@@ -1651,7 +1649,7 @@ torch.cuda.empty_cache()
 images
 ```
 
-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)
 
 ## Example Images Mixing (with CoCa)
@@ -1703,7 +1701,7 @@ 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)
@@ -1732,7 +1730,7 @@ 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
 
@@ -1805,7 +1803,7 @@ 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
@@ -2014,7 +2012,7 @@ 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`
 
@@ -2025,7 +2023,7 @@ Reference Image
 ![reference_image](https://github.com/huggingface/diffusers/assets/34944964/449bdab6-e744-4fb2-9620-d4068d9a741b)
 
 
-Output Image 
+Output Image
 
 `prompt: A dog`
 
@@ -2106,7 +2104,7 @@ 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
@@ -2259,7 +2257,7 @@ 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
 ```
@@ -2295,7 +2293,7 @@ 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
 ```
@@ -2348,7 +2346,7 @@ 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*
@@ -2389,8 +2387,8 @@ 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
@@ -2541,7 +2539,7 @@ pipe = RegionalPromptingStableDiffusionPipeline.from_single_file(model_path, vae
 rp_args = {
     "mode":"rows",
     "div": "1;1;1"
-}  
+}
 
 prompt ="""
 green hair twintail BREAK
@@ -2570,7 +2568,7 @@ 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
@@ -2628,7 +2626,7 @@ 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)
@@ -2647,7 +2645,7 @@ 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
@@ -2678,13 +2676,13 @@ 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)
 ```
@@ -2763,7 +2761,7 @@ 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,
@@ -2838,11 +2836,75 @@ 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).
@@ -2869,7 +2931,7 @@ 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(
@@ -2884,24 +2946,24 @@ 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
@@ -2919,9 +2981,48 @@ 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/checkpoint_merger.py

@@ -5,10 +5,11 @@ from typing import Dict, List, Union
 import safetensors.torch
 import torch
 from huggingface_hub import snapshot_download
+from huggingface_hub.utils import validate_hf_hub_args
 
 from diffusers import DiffusionPipeline, __version__
 from diffusers.schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
-from diffusers.utils import CONFIG_NAME, DIFFUSERS_CACHE, ONNX_WEIGHTS_NAME, WEIGHTS_NAME
+from diffusers.utils import CONFIG_NAME, ONNX_WEIGHTS_NAME, WEIGHTS_NAME
 
 
 class CheckpointMergerPipeline(DiffusionPipeline):
@@ -57,6 +58,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
         return (temp_dict, meta_keys)
 
     @torch.no_grad()
+    @validate_hf_hub_args
     def merge(self, pretrained_model_name_or_path_list: List[Union[str, os.PathLike]], **kwargs):
         """
         Returns a new pipeline object of the class 'DiffusionPipeline' with the merged checkpoints(weights) of the models passed
@@ -69,7 +71,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
             **kwargs:
                 Supports all the default DiffusionPipeline.get_config_dict kwargs viz..
 
-                cache_dir, resume_download, force_download, proxies, local_files_only, use_auth_token, revision, torch_dtype, device_map.
+                cache_dir, resume_download, force_download, proxies, local_files_only, token, revision, torch_dtype, device_map.
 
                 alpha - The interpolation parameter. Ranges from 0 to 1.  It affects the ratio in which the checkpoints are merged. A 0.8 alpha
                     would mean that the first model checkpoints would affect the final result far less than an alpha of 0.2
@@ -81,12 +83,12 @@ class CheckpointMergerPipeline(DiffusionPipeline):
 
         """
         # Default kwargs from DiffusionPipeline
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         torch_dtype = kwargs.pop("torch_dtype", None)
         device_map = kwargs.pop("device_map", None)
@@ -123,7 +125,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
                 force_download=force_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
             )
             config_dicts.append(config_dict)
@@ -159,7 +161,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
+                    token=token,
                     revision=revision,
                     allow_patterns=allow_patterns,
                     user_agent=user_agent,

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/community/stable_diffusion_tensorrt_img2img.py

@@ -28,6 +28,7 @@ import PIL.Image
 import tensorrt as trt
 import torch
 from huggingface_hub import snapshot_download
+from huggingface_hub.utils import validate_hf_hub_args
 from onnx import shape_inference
 from polygraphy import cuda
 from polygraphy.backend.common import bytes_from_path
@@ -50,7 +51,7 @@ from diffusers.pipelines.stable_diffusion import (
     StableDiffusionSafetyChecker,
 )
 from diffusers.schedulers import DDIMScheduler
-from diffusers.utils import DIFFUSERS_CACHE, logging
+from diffusers.utils import logging
 
 
 """
@@ -778,12 +779,13 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
             self.models["vae_encoder"] = make_VAEEncoder(self.vae, **models_args)
 
     @classmethod
+    @validate_hf_hub_args
     def set_cached_folder(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
 
         cls.cached_folder = (
@@ -795,7 +797,7 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
             )
         )

examples/community/stable_diffusion_tensorrt_inpaint.py

@@ -28,6 +28,7 @@ import PIL.Image
 import tensorrt as trt
 import torch
 from huggingface_hub import snapshot_download
+from huggingface_hub.utils import validate_hf_hub_args
 from onnx import shape_inference
 from polygraphy import cuda
 from polygraphy.backend.common import bytes_from_path
@@ -51,7 +52,7 @@ from diffusers.pipelines.stable_diffusion import (
 )
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint import prepare_mask_and_masked_image
 from diffusers.schedulers import DDIMScheduler
-from diffusers.utils import DIFFUSERS_CACHE, logging
+from diffusers.utils import logging
 
 
 """
@@ -779,12 +780,13 @@ class TensorRTStableDiffusionInpaintPipeline(StableDiffusionInpaintPipeline):
             self.models["vae_encoder"] = make_VAEEncoder(self.vae, **models_args)
 
     @classmethod
+    @validate_hf_hub_args
     def set_cached_folder(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
 
         cls.cached_folder = (
@@ -796,7 +798,7 @@ class TensorRTStableDiffusionInpaintPipeline(StableDiffusionInpaintPipeline):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
             )
         )

examples/community/stable_diffusion_tensorrt_txt2img.py

@@ -27,6 +27,7 @@ import onnx_graphsurgeon as gs
 import tensorrt as trt
 import torch
 from huggingface_hub import snapshot_download
+from huggingface_hub.utils import validate_hf_hub_args
 from onnx import shape_inference
 from polygraphy import cuda
 from polygraphy.backend.common import bytes_from_path
@@ -49,7 +50,7 @@ from diffusers.pipelines.stable_diffusion import (
     StableDiffusionSafetyChecker,
 )
 from diffusers.schedulers import DDIMScheduler
-from diffusers.utils import DIFFUSERS_CACHE, logging
+from diffusers.utils import logging
 
 
 """
@@ -691,12 +692,13 @@ class TensorRTStableDiffusionPipeline(StableDiffusionPipeline):
             self.models["vae"] = make_VAE(self.vae, **models_args)
 
     @classmethod
+    @validate_hf_hub_args
     def set_cached_folder(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
 
         cls.cached_folder = (
@@ -708,7 +710,7 @@ class TensorRTStableDiffusionPipeline(StableDiffusionPipeline):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
             )
         )

examples/consistency_distillation/train_lcm_distill_lora_sd_wds.py

@@ -423,7 +423,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
     )
     model_class = text_encoder_config.architectures[0]
 

examples/consistency_distillation/train_lcm_distill_lora_sdxl_wds.py

@@ -397,7 +397,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
     )
     model_class = text_encoder_config.architectures[0]
 

examples/consistency_distillation/train_lcm_distill_sd_wds.py

@@ -400,7 +400,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
     )
     model_class = text_encoder_config.architectures[0]
 

examples/consistency_distillation/train_lcm_distill_sdxl_wds.py

@@ -419,7 +419,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
     )
     model_class = text_encoder_config.architectures[0]
 

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
 
@@ -864,79 +858,19 @@ 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 +882,9 @@ 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 +944,10 @@ 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 +1190,7 @@ 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 +1313,19 @@ 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
 
@@ -1009,54 +1010,19 @@ 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)
 
     # 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 +1035,11 @@ 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 +1096,12 @@ 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 +1166,10 @@ 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 +1615,13 @@ 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/controlnet/train_controlnet_webdataset.py

@@ -420,7 +420,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
     )
     model_class = text_encoder_config.architectures[0]
 
@@ -975,7 +975,7 @@ def main(args):
         revision=args.revision,
     )
     unet = UNet2DConditionModel.from_pretrained(
-        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, use_auth_token=True
+        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
     )
 
     if args.controlnet_model_name_or_path:

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 -->
 
@@ -143,11 +145,11 @@ 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 +157,7 @@ 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 +167,7 @@ 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 +186,7 @@ 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 +197,7 @@ 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 +225,11 @@ 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 +248,9 @@ 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 +287,7 @@ 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 +321,5 @@ 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
@@ -479,62 +480,20 @@ 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())
+    unet.add_adapter(unet_lora_config)
 
     if args.enable_xformers_memory_efficient_attention:
         if is_xformers_available():
@@ -549,6 +508,8 @@ 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 +534,7 @@ 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 +661,8 @@ 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 +794,7 @@ 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 +831,15 @@ def main():
 
                         save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                         accelerator.save_state(save_path)
+
+                        unet_lora_state_dict = get_peft_model_state_dict(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 +896,13 @@ def main():
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         unet = unet.to(torch.float32)
-        unet.save_attn_procs(args.output_dir)
+
+        unet_lora_state_dict = get_peft_model_state_dict(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(

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
@@ -658,53 +658,20 @@ 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 +684,11 @@ 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 +759,13 @@ 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 +1097,7 @@ 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 +1193,21 @@ 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

@@ -72,7 +72,7 @@ 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"

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/commands/fp16_safetensors.py

@@ -19,6 +19,7 @@ Usage example:
 
 import glob
 import json
+import warnings
 from argparse import ArgumentParser, Namespace
 from importlib import import_module
 
@@ -32,12 +33,12 @@ from . import BaseDiffusersCLICommand
 
 
 def conversion_command_factory(args: Namespace):
-    return FP16SafetensorsCommand(
-        args.ckpt_id,
-        args.fp16,
-        args.use_safetensors,
-        args.use_auth_token,
-    )
+    if args.use_auth_token:
+        warnings.warn(
+            "The `--use_auth_token` flag is deprecated and will be removed in a future version. Authentication is now"
+            " handled automatically if user is logged in."
+        )
+    return FP16SafetensorsCommand(args.ckpt_id, args.fp16, args.use_safetensors)
 
 
 class FP16SafetensorsCommand(BaseDiffusersCLICommand):
@@ -62,7 +63,7 @@ class FP16SafetensorsCommand(BaseDiffusersCLICommand):
         )
         conversion_parser.set_defaults(func=conversion_command_factory)
 
-    def __init__(self, ckpt_id: str, fp16: bool, use_safetensors: bool, use_auth_token: bool):
+    def __init__(self, ckpt_id: str, fp16: bool, use_safetensors: bool):
         self.logger = logging.get_logger("diffusers-cli/fp16_safetensors")
         self.ckpt_id = ckpt_id
         self.local_ckpt_dir = f"/tmp/{ckpt_id}"
@@ -75,8 +76,6 @@ class FP16SafetensorsCommand(BaseDiffusersCLICommand):
                 "When `use_safetensors` and `fp16` both are False, then this command is of no use."
             )
 
-        self.use_auth_token = use_auth_token
-
     def run(self):
         if version.parse(huggingface_hub.__version__) < version.parse("0.9.0"):
             raise ImportError(
@@ -87,7 +86,7 @@ class FP16SafetensorsCommand(BaseDiffusersCLICommand):
             from huggingface_hub import create_commit
             from huggingface_hub._commit_api import CommitOperationAdd
 
-        model_index = hf_hub_download(repo_id=self.ckpt_id, filename="model_index.json", token=self.use_auth_token)
+        model_index = hf_hub_download(repo_id=self.ckpt_id, filename="model_index.json")
         with open(model_index, "r") as f:
             pipeline_class_name = json.load(f)["_class_name"]
         pipeline_class = getattr(import_module("diffusers"), pipeline_class_name)
@@ -96,7 +95,7 @@ class FP16SafetensorsCommand(BaseDiffusersCLICommand):
         # Load the appropriate pipeline. We could have use `DiffusionPipeline`
         # here, but just to avoid any rough edge cases.
         pipeline = pipeline_class.from_pretrained(
-            self.ckpt_id, torch_dtype=torch.float16 if self.fp16 else torch.float32, use_auth_token=self.use_auth_token
+            self.ckpt_id, torch_dtype=torch.float16 if self.fp16 else torch.float32
         )
         pipeline.save_pretrained(
             self.local_ckpt_dir,

src/diffusers/configuration_utils.py

@@ -27,12 +27,16 @@ from typing import Any, Dict, Tuple, Union
 
 import numpy as np
 from huggingface_hub import create_repo, hf_hub_download
-from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError
+from huggingface_hub.utils import (
+    EntryNotFoundError,
+    RepositoryNotFoundError,
+    RevisionNotFoundError,
+    validate_hf_hub_args,
+)
 from requests import HTTPError
 
 from . import __version__
 from .utils import (
-    DIFFUSERS_CACHE,
     HUGGINGFACE_CO_RESOLVE_ENDPOINT,
     DummyObject,
     deprecate,
@@ -275,6 +279,7 @@ class ConfigMixin:
         return cls.load_config(*args, **kwargs)
 
     @classmethod
+    @validate_hf_hub_args
     def load_config(
         cls,
         pretrained_model_name_or_path: Union[str, os.PathLike],
@@ -311,7 +316,7 @@ class ConfigMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -329,11 +334,11 @@ class ConfigMixin:
                 A dictionary of all the parameters stored in a JSON configuration file.
 
         """
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
         _ = kwargs.pop("mirror", None)
@@ -376,7 +381,7 @@ class ConfigMixin:
                     proxies=proxies,
                     resume_download=resume_download,
                     local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
+                    token=token,
                     user_agent=user_agent,
                     subfolder=subfolder,
                     revision=revision,
@@ -385,8 +390,7 @@ class ConfigMixin:
                 raise EnvironmentError(
                     f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier"
                     " listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a"
-                    " token having permission to this repo with `use_auth_token` or log in with `huggingface-cli"
-                    " login`."
+                    " token having permission to this repo with `token` or log in with `huggingface-cli login`."
                 )
             except RevisionNotFoundError:
                 raise EnvironmentError(

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/ip_adapter.py

@@ -15,11 +15,10 @@ import os
 from typing import Dict, Union
 
 import torch
+from huggingface_hub.utils import validate_hf_hub_args
 from safetensors import safe_open
 
 from ..utils import (
-    DIFFUSERS_CACHE,
-    HF_HUB_OFFLINE,
     _get_model_file,
     is_transformers_available,
     logging,
@@ -43,6 +42,7 @@ logger = logging.get_logger(__name__)
 class IPAdapterMixin:
     """Mixin for handling IP Adapters."""
 
+    @validate_hf_hub_args
     def load_ip_adapter(
         self,
         pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
@@ -77,7 +77,7 @@ class IPAdapterMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -88,12 +88,12 @@ class IPAdapterMixin:
         """
 
         # Load the main state dict first.
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
 
         user_agent = {
@@ -110,7 +110,7 @@ class IPAdapterMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 subfolder=subfolder,
                 user_agent=user_agent,

src/diffusers/loaders/lora.py

@@ -18,14 +18,13 @@ from typing import Callable, Dict, List, Optional, Union
 import safetensors
 import torch
 from huggingface_hub import model_info
+from huggingface_hub.utils import validate_hf_hub_args
 from packaging import version
 from torch import nn
 
 from .. import __version__
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
 from ..utils import (
-    DIFFUSERS_CACHE,
-    HF_HUB_OFFLINE,
     USE_PEFT_BACKEND,
     _get_model_file,
     convert_state_dict_to_diffusers,
@@ -132,6 +131,7 @@ class LoraLoaderMixin:
         )
 
     @classmethod
+    @validate_hf_hub_args
     def lora_state_dict(
         cls,
         pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
@@ -174,7 +174,7 @@ class LoraLoaderMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -195,12 +195,12 @@ class LoraLoaderMixin:
         """
         # Load the main state dict first which has the LoRA layers for either of
         # UNet and text encoder or both.
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         subfolder = kwargs.pop("subfolder", None)
         weight_name = kwargs.pop("weight_name", None)
@@ -239,7 +239,7 @@ class LoraLoaderMixin:
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        use_auth_token=use_auth_token,
+                        token=token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -265,7 +265,7 @@ class LoraLoaderMixin:
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
+                    token=token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,

src/diffusers/loaders/single_file.py

@@ -18,10 +18,9 @@ from pathlib import Path
 import requests
 import torch
 from huggingface_hub import hf_hub_download
+from huggingface_hub.utils import validate_hf_hub_args
 
 from ..utils import (
-    DIFFUSERS_CACHE,
-    HF_HUB_OFFLINE,
     deprecate,
     is_accelerate_available,
     is_omegaconf_available,
@@ -52,6 +51,7 @@ class FromSingleFileMixin:
         return cls.from_single_file(*args, **kwargs)
 
     @classmethod
+    @validate_hf_hub_args
     def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
         r"""
         Instantiate a [`DiffusionPipeline`] from pretrained pipeline weights saved in the `.ckpt` or `.safetensors`
@@ -81,7 +81,7 @@ class FromSingleFileMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -154,12 +154,12 @@ class FromSingleFileMixin:
 
         original_config_file = kwargs.pop("original_config_file", None)
         config_files = kwargs.pop("config_files", None)
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         extract_ema = kwargs.pop("extract_ema", False)
         image_size = kwargs.pop("image_size", None)
@@ -253,7 +253,7 @@ class FromSingleFileMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 force_download=force_download,
             )
@@ -293,6 +293,7 @@ class FromOriginalVAEMixin:
     """
 
     @classmethod
+    @validate_hf_hub_args
     def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
         r"""
         Instantiate a [`AutoencoderKL`] from pretrained ControlNet weights saved in the original `.ckpt` or
@@ -322,7 +323,7 @@ class FromOriginalVAEMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to True, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -379,12 +380,12 @@ class FromOriginalVAEMixin:
         )
 
         config_file = kwargs.pop("config_file", None)
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         image_size = kwargs.pop("image_size", None)
         scaling_factor = kwargs.pop("scaling_factor", None)
@@ -425,7 +426,7 @@ class FromOriginalVAEMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 force_download=force_download,
             )
@@ -490,6 +491,7 @@ class FromOriginalControlnetMixin:
     """
 
     @classmethod
+    @validate_hf_hub_args
     def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
         r"""
         Instantiate a [`ControlNetModel`] from pretrained ControlNet weights saved in the original `.ckpt` or
@@ -519,7 +521,7 @@ class FromOriginalControlnetMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to True, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -555,12 +557,12 @@ class FromOriginalControlnetMixin:
         from ..pipelines.stable_diffusion.convert_from_ckpt import download_controlnet_from_original_ckpt
 
         config_file = kwargs.pop("config_file", None)
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         num_in_channels = kwargs.pop("num_in_channels", None)
         use_linear_projection = kwargs.pop("use_linear_projection", None)
         revision = kwargs.pop("revision", None)
@@ -603,7 +605,7 @@ class FromOriginalControlnetMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 force_download=force_download,
             )

src/diffusers/loaders/textual_inversion.py

@@ -15,16 +15,10 @@ from typing import Dict, List, Optional, Union
 
 import safetensors
 import torch
+from huggingface_hub.utils import validate_hf_hub_args
 from torch import nn
 
-from ..utils import (
-    DIFFUSERS_CACHE,
-    HF_HUB_OFFLINE,
-    _get_model_file,
-    is_accelerate_available,
-    is_transformers_available,
-    logging,
-)
+from ..utils import _get_model_file, is_accelerate_available, is_transformers_available, logging
 
 
 if is_transformers_available():
@@ -39,13 +33,14 @@ TEXT_INVERSION_NAME = "learned_embeds.bin"
 TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
 
 
+@validate_hf_hub_args
 def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs):
-    cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+    cache_dir = kwargs.pop("cache_dir", None)
     force_download = kwargs.pop("force_download", False)
     resume_download = kwargs.pop("resume_download", False)
     proxies = kwargs.pop("proxies", None)
-    local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-    use_auth_token = kwargs.pop("use_auth_token", None)
+    local_files_only = kwargs.pop("local_files_only", None)
+    token = kwargs.pop("token", None)
     revision = kwargs.pop("revision", None)
     subfolder = kwargs.pop("subfolder", None)
     weight_name = kwargs.pop("weight_name", None)
@@ -79,7 +74,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        use_auth_token=use_auth_token,
+                        token=token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -100,7 +95,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
+                    token=token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,
@@ -267,6 +262,7 @@ class TextualInversionLoaderMixin:
 
         return all_tokens, all_embeddings
 
+    @validate_hf_hub_args
     def load_textual_inversion(
         self,
         pretrained_model_name_or_path: Union[str, List[str], Dict[str, torch.Tensor], List[Dict[str, torch.Tensor]]],
@@ -320,7 +316,7 @@ class TextualInversionLoaderMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):

src/diffusers/loaders/unet.py

@@ -19,13 +19,12 @@ from typing import Callable, Dict, List, Optional, Union
 import safetensors
 import torch
 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 (
-    DIFFUSERS_CACHE,
-    HF_HUB_OFFLINE,
     USE_PEFT_BACKEND,
     _get_model_file,
     delete_adapter_layers,
@@ -62,6 +61,7 @@ class UNet2DConditionLoadersMixin:
     text_encoder_name = TEXT_ENCODER_NAME
     unet_name = UNET_NAME
 
+    @validate_hf_hub_args
     def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
         r"""
         Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
@@ -95,7 +95,7 @@ class UNet2DConditionLoadersMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
@@ -130,12 +130,12 @@ class UNet2DConditionLoadersMixin:
         from ..models.attention_processor import CustomDiffusionAttnProcessor
         from ..models.lora import LoRACompatibleConv, LoRACompatibleLinear, LoRAConv2dLayer, LoRALinearLayer
 
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         subfolder = kwargs.pop("subfolder", None)
         weight_name = kwargs.pop("weight_name", None)
@@ -184,7 +184,7 @@ class UNet2DConditionLoadersMixin:
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        use_auth_token=use_auth_token,
+                        token=token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -204,7 +204,7 @@ class UNet2DConditionLoadersMixin:
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
+                    token=token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,
@@ -675,6 +675,9 @@ 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]
@@ -744,8 +747,32 @@ class UNet2DConditionLoadersMixin:
                     "norm.bias": state_dict["image_proj"]["norm.bias"],
                 }
             )
-
             image_projection.load_state_dict(image_proj_state_dict)
+            del image_proj_state_dict
+
+        elif "proj.3.weight" in state_dict["image_proj"]:
+            clip_embeddings_dim = state_dict["image_proj"]["proj.0.weight"].shape[0]
+            cross_attention_dim = state_dict["image_proj"]["proj.3.weight"].shape[0]
+
+            image_projection = MLPProjection(
+                cross_attention_dim=cross_attention_dim, image_embed_dim=clip_embeddings_dim
+            )
+            image_projection.to(dtype=self.dtype, device=self.device)
+
+            # load image projection layer weights
+            image_proj_state_dict = {}
+            image_proj_state_dict.update(
+                {
+                    "ff.net.0.proj.weight": state_dict["image_proj"]["proj.0.weight"],
+                    "ff.net.0.proj.bias": state_dict["image_proj"]["proj.0.bias"],
+                    "ff.net.2.weight": state_dict["image_proj"]["proj.2.weight"],
+                    "ff.net.2.bias": state_dict["image_proj"]["proj.2.bias"],
+                    "norm.weight": state_dict["image_proj"]["proj.3.weight"],
+                    "norm.bias": state_dict["image_proj"]["proj.3.bias"],
+                }
+            )
+            image_projection.load_state_dict(image_proj_state_dict)
+            del image_proj_state_dict
 
         else:
             # IP-Adapter Plus

src/diffusers/models/__init__.py

@@ -32,9 +32,10 @@ if is_torch_available():
     _import_structure["autoencoder_tiny"] = ["AutoencoderTiny"]
     _import_structure["consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["controlnet"] = ["ControlNetModel"]
+    _import_structure["controlnetxs"] = ["ControlNetXSModel"]
     _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
-    _import_structure["modeling_utils"] = ["ModelMixin"]
     _import_structure["embeddings"] = ["ImageProjection"]
+    _import_structure["modeling_utils"] = ["ModelMixin"]
     _import_structure["prior_transformer"] = ["PriorTransformer"]
     _import_structure["t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformer_2d"] = ["Transformer2DModel"]
@@ -63,6 +64,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .autoencoder_tiny import AutoencoderTiny
         from .consistency_decoder_vae import 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/controlnetxs.py

@@ -0,0 +1,977 @@
+# 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.
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import functional as F
+from torch.nn.modules.normalization import GroupNorm
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput, logging
+from .attention_processor import (
+    AttentionProcessor,
+)
+from .autoencoder_kl import AutoencoderKL
+from .lora import LoRACompatibleConv
+from .modeling_utils import ModelMixin
+from .unet_2d_blocks import (
+    CrossAttnDownBlock2D,
+    CrossAttnUpBlock2D,
+    DownBlock2D,
+    Downsample2D,
+    ResnetBlock2D,
+    Transformer2DModel,
+    UpBlock2D,
+    Upsample2D,
+)
+from .unet_2d_condition import UNet2DConditionModel
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class ControlNetXSOutput(BaseOutput):
+    """
+    The output of [`ControlNetXSModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The output of the `ControlNetXSModel`. Unlike `ControlNetOutput` this is NOT to be added to the base model
+            output, but is already the final output.
+    """
+
+    sample: torch.FloatTensor = None
+
+
+# copied from diffusers.models.controlnet.ControlNetConditioningEmbedding
+class ControlNetConditioningEmbedding(nn.Module):
+    """
+    Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
+    [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
+    training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
+    convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
+    (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
+    model) to encode image-space conditions ... into feature maps ..."
+    """
+
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        conditioning_channels: int = 3,
+        block_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
+    ):
+        super().__init__()
+
+        self.conv_in = nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
+
+        self.blocks = nn.ModuleList([])
+
+        for i in range(len(block_out_channels) - 1):
+            channel_in = block_out_channels[i]
+            channel_out = block_out_channels[i + 1]
+            self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
+            self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
+
+        self.conv_out = zero_module(
+            nn.Conv2d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
+        )
+
+    def forward(self, conditioning):
+        embedding = self.conv_in(conditioning)
+        embedding = F.silu(embedding)
+
+        for block in self.blocks:
+            embedding = block(embedding)
+            embedding = F.silu(embedding)
+
+        embedding = self.conv_out(embedding)
+
+        return embedding
+
+
+class ControlNetXSModel(ModelMixin, ConfigMixin):
+    r"""
+    A ControlNet-XS model
+
+    This model inherits from [`ModelMixin`] and [`ConfigMixin`]. Check the superclass documentation for it's generic
+    methods implemented for all models (such as downloading or saving).
+
+    Most of parameters for this model are passed into the [`UNet2DConditionModel`] it creates. Check the documentation
+    of [`UNet2DConditionModel`] for them.
+
+    Parameters:
+        conditioning_channels (`int`, defaults to 3):
+            Number of channels of conditioning input (e.g. an image)
+        controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+            The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+        conditioning_embedding_out_channels (`tuple[int]`, defaults to `(16, 32, 96, 256)`):
+            The tuple of output channel for each block in the `controlnet_cond_embedding` layer.
+        time_embedding_input_dim (`int`, defaults to 320):
+            Dimension of input into time embedding. Needs to be same as in the base model.
+        time_embedding_dim (`int`, defaults to 1280):
+            Dimension of output from time embedding. Needs to be same as in the base model.
+        learn_embedding (`bool`, defaults to `False`):
+            Whether to use time embedding of the control model. If yes, the time embedding is a linear interpolation of
+            the time embeddings of the control and base model with interpolation parameter `time_embedding_mix**3`.
+        time_embedding_mix (`float`, defaults to 1.0):
+            Linear interpolation parameter used if `learn_embedding` is `True`. A value of 1.0 means only the
+            control model's time embedding will be used. A value of 0.0 means only the base model's time embedding will be used.
+        base_model_channel_sizes (`Dict[str, List[Tuple[int]]]`):
+            Channel sizes of each subblock of base model. Use `gather_subblock_sizes` on your base model to compute it.
+    """
+
+    @classmethod
+    def init_original(cls, base_model: UNet2DConditionModel, is_sdxl=True):
+        """
+        Create a ControlNetXS model with the same parameters as in the original paper (https://github.com/vislearn/ControlNet-XS).
+
+        Parameters:
+            base_model (`UNet2DConditionModel`):
+                Base UNet model. Needs to be either StableDiffusion or StableDiffusion-XL.
+            is_sdxl (`bool`, defaults to `True`):
+                Whether passed `base_model` is a StableDiffusion-XL model.
+        """
+
+        def get_dim_attn_heads(base_model: UNet2DConditionModel, size_ratio: float, num_attn_heads: int):
+            """
+            Currently, diffusers can only set the dimension of attention heads (see https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 for why).
+            The original ControlNet-XS model, however, define the number of attention heads.
+            That's why compute the dimensions needed to get the correct number of attention heads.
+            """
+            block_out_channels = [int(size_ratio * c) for c in base_model.config.block_out_channels]
+            dim_attn_heads = [math.ceil(c / num_attn_heads) for c in block_out_channels]
+            return dim_attn_heads
+
+        if is_sdxl:
+            return ControlNetXSModel.from_unet(
+                base_model,
+                time_embedding_mix=0.95,
+                learn_embedding=True,
+                size_ratio=0.1,
+                conditioning_embedding_out_channels=(16, 32, 96, 256),
+                num_attention_heads=get_dim_attn_heads(base_model, 0.1, 64),
+            )
+        else:
+            return ControlNetXSModel.from_unet(
+                base_model,
+                time_embedding_mix=1.0,
+                learn_embedding=True,
+                size_ratio=0.0125,
+                conditioning_embedding_out_channels=(16, 32, 96, 256),
+                num_attention_heads=get_dim_attn_heads(base_model, 0.0125, 8),
+            )
+
+    @classmethod
+    def _gather_subblock_sizes(cls, unet: UNet2DConditionModel, base_or_control: str):
+        """To create correctly sized connections between base and control model, we need to know
+        the input and output channels of each subblock.
+
+        Parameters:
+            unet (`UNet2DConditionModel`):
+                Unet of which the subblock channels sizes are to be gathered.
+            base_or_control (`str`):
+                Needs to be either "base" or "control". If "base", decoder is also considered.
+        """
+        if base_or_control not in ["base", "control"]:
+            raise ValueError("`base_or_control` needs to be either `base` or `control`")
+
+        channel_sizes = {"down": [], "mid": [], "up": []}
+
+        # input convolution
+        channel_sizes["down"].append((unet.conv_in.in_channels, unet.conv_in.out_channels))
+
+        # encoder blocks
+        for module in unet.down_blocks:
+            if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
+                for r in module.resnets:
+                    channel_sizes["down"].append((r.in_channels, r.out_channels))
+                if module.downsamplers:
+                    channel_sizes["down"].append(
+                        (module.downsamplers[0].channels, module.downsamplers[0].out_channels)
+                    )
+            else:
+                raise ValueError(f"Encountered unknown module of type {type(module)} while creating ControlNet-XS.")
+
+        # middle block
+        channel_sizes["mid"].append((unet.mid_block.resnets[0].in_channels, unet.mid_block.resnets[0].out_channels))
+
+        # decoder blocks
+        if base_or_control == "base":
+            for module in unet.up_blocks:
+                if isinstance(module, (CrossAttnUpBlock2D, UpBlock2D)):
+                    for r in module.resnets:
+                        channel_sizes["up"].append((r.in_channels, r.out_channels))
+                else:
+                    raise ValueError(
+                        f"Encountered unknown module of type {type(module)} while creating ControlNet-XS."
+                    )
+
+        return channel_sizes
+
+    @register_to_config
+    def __init__(
+        self,
+        conditioning_channels: int = 3,
+        conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
+        controlnet_conditioning_channel_order: str = "rgb",
+        time_embedding_input_dim: int = 320,
+        time_embedding_dim: int = 1280,
+        time_embedding_mix: float = 1.0,
+        learn_embedding: bool = False,
+        base_model_channel_sizes: Dict[str, List[Tuple[int]]] = {
+            "down": [
+                (4, 320),
+                (320, 320),
+                (320, 320),
+                (320, 320),
+                (320, 640),
+                (640, 640),
+                (640, 640),
+                (640, 1280),
+                (1280, 1280),
+            ],
+            "mid": [(1280, 1280)],
+            "up": [
+                (2560, 1280),
+                (2560, 1280),
+                (1920, 1280),
+                (1920, 640),
+                (1280, 640),
+                (960, 640),
+                (960, 320),
+                (640, 320),
+                (640, 320),
+            ],
+        },
+        sample_size: Optional[int] = None,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        norm_num_groups: Optional[int] = 32,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = 8,
+        upcast_attention: bool = False,
+    ):
+        super().__init__()
+
+        # 1 - Create control unet
+        self.control_model = UNet2DConditionModel(
+            sample_size=sample_size,
+            down_block_types=down_block_types,
+            up_block_types=up_block_types,
+            block_out_channels=block_out_channels,
+            norm_num_groups=norm_num_groups,
+            cross_attention_dim=cross_attention_dim,
+            transformer_layers_per_block=transformer_layers_per_block,
+            attention_head_dim=num_attention_heads,
+            use_linear_projection=True,
+            upcast_attention=upcast_attention,
+            time_embedding_dim=time_embedding_dim,
+        )
+
+        # 2 - Do model surgery on control model
+        # 2.1 - Allow to use the same time information as the base model
+        adjust_time_dims(self.control_model, time_embedding_input_dim, time_embedding_dim)
+
+        # 2.2 - Allow for information infusion from base model
+
+        # We concat the output of each base encoder subblocks to the input of the next control encoder subblock
+        # (We ignore the 1st element, as it represents the `conv_in`.)
+        extra_input_channels = [input_channels for input_channels, _ in base_model_channel_sizes["down"][1:]]
+        it_extra_input_channels = iter(extra_input_channels)
+
+        for b, block in enumerate(self.control_model.down_blocks):
+            for r in range(len(block.resnets)):
+                increase_block_input_in_encoder_resnet(
+                    self.control_model, block_no=b, resnet_idx=r, by=next(it_extra_input_channels)
+                )
+
+            if block.downsamplers:
+                increase_block_input_in_encoder_downsampler(
+                    self.control_model, block_no=b, by=next(it_extra_input_channels)
+                )
+
+        increase_block_input_in_mid_resnet(self.control_model, by=extra_input_channels[-1])
+
+        # 2.3 - Make group norms work with modified channel sizes
+        adjust_group_norms(self.control_model)
+
+        # 3 - Gather Channel Sizes
+        self.ch_inout_ctrl = ControlNetXSModel._gather_subblock_sizes(self.control_model, base_or_control="control")
+        self.ch_inout_base = base_model_channel_sizes
+
+        # 4 - Build connections between base and control model
+        self.down_zero_convs_out = nn.ModuleList([])
+        self.down_zero_convs_in = nn.ModuleList([])
+        self.middle_block_out = nn.ModuleList([])
+        self.middle_block_in = nn.ModuleList([])
+        self.up_zero_convs_out = nn.ModuleList([])
+        self.up_zero_convs_in = nn.ModuleList([])
+
+        for ch_io_base in self.ch_inout_base["down"]:
+            self.down_zero_convs_in.append(self._make_zero_conv(in_channels=ch_io_base[1], out_channels=ch_io_base[1]))
+        for i in range(len(self.ch_inout_ctrl["down"])):
+            self.down_zero_convs_out.append(
+                self._make_zero_conv(self.ch_inout_ctrl["down"][i][1], self.ch_inout_base["down"][i][1])
+            )
+
+        self.middle_block_out = self._make_zero_conv(
+            self.ch_inout_ctrl["mid"][-1][1], self.ch_inout_base["mid"][-1][1]
+        )
+
+        self.up_zero_convs_out.append(
+            self._make_zero_conv(self.ch_inout_ctrl["down"][-1][1], self.ch_inout_base["mid"][-1][1])
+        )
+        for i in range(1, len(self.ch_inout_ctrl["down"])):
+            self.up_zero_convs_out.append(
+                self._make_zero_conv(self.ch_inout_ctrl["down"][-(i + 1)][1], self.ch_inout_base["up"][i - 1][1])
+            )
+
+        # 5 - Create conditioning hint embedding
+        self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
+            conditioning_embedding_channels=block_out_channels[0],
+            block_out_channels=conditioning_embedding_out_channels,
+            conditioning_channels=conditioning_channels,
+        )
+
+        # In the mininal implementation setting, we only need the control model up to the mid block
+        del self.control_model.up_blocks
+        del self.control_model.conv_norm_out
+        del self.control_model.conv_out
+
+    @classmethod
+    def from_unet(
+        cls,
+        unet: UNet2DConditionModel,
+        conditioning_channels: int = 3,
+        conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
+        controlnet_conditioning_channel_order: str = "rgb",
+        learn_embedding: bool = False,
+        time_embedding_mix: float = 1.0,
+        block_out_channels: Optional[Tuple[int]] = None,
+        size_ratio: Optional[float] = None,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = 8,
+        norm_num_groups: Optional[int] = None,
+    ):
+        r"""
+        Instantiate a [`ControlNetXSModel`] from [`UNet2DConditionModel`].
+
+        Parameters:
+            unet (`UNet2DConditionModel`):
+                The UNet model we want to control. The dimensions of the ControlNetXSModel will be adapted to it.
+            conditioning_channels (`int`, defaults to 3):
+                Number of channels of conditioning input (e.g. an image)
+            conditioning_embedding_out_channels (`tuple[int]`, defaults to `(16, 32, 96, 256)`):
+                The tuple of output channel for each block in the `controlnet_cond_embedding` layer.
+            controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+                The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+            learn_embedding (`bool`, defaults to `False`):
+                Wether to use time embedding of the control model. If yes, the time embedding is a linear interpolation
+                of the time embeddings of the control and base model with interpolation parameter
+                `time_embedding_mix**3`.
+            time_embedding_mix (`float`, defaults to 1.0):
+                Linear interpolation parameter used if `learn_embedding` is `True`.
+            block_out_channels (`Tuple[int]`, *optional*):
+                Down blocks output channels in control model. Either this or `size_ratio` must be given.
+            size_ratio (float, *optional*):
+                When given, block_out_channels is set to a relative fraction of the base model's block_out_channels.
+                Either this or `block_out_channels` must be given.
+            num_attention_heads (`Union[int, Tuple[int]]`, *optional*):
+                The dimension of the attention heads. The naming seems a bit confusing and it is, see https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 for why.
+            norm_num_groups (int, *optional*, defaults to `None`):
+                The number of groups to use for the normalization of the control unet. If `None`,
+                `int(unet.config.norm_num_groups * size_ratio)` is taken.
+        """
+
+        # Check input
+        fixed_size = block_out_channels is not None
+        relative_size = size_ratio is not None
+        if not (fixed_size ^ relative_size):
+            raise ValueError(
+                "Pass exactly one of `block_out_channels` (for absolute sizing) or `control_model_ratio` (for relative sizing)."
+            )
+
+        # Create model
+        if block_out_channels is None:
+            block_out_channels = [int(size_ratio * c) for c in unet.config.block_out_channels]
+
+        # Check that attention heads and group norms match channel sizes
+        # - attention heads
+        def attn_heads_match_channel_sizes(attn_heads, channel_sizes):
+            if isinstance(attn_heads, (tuple, list)):
+                return all(c % a == 0 for a, c in zip(attn_heads, channel_sizes))
+            else:
+                return all(c % attn_heads == 0 for c in channel_sizes)
+
+        num_attention_heads = num_attention_heads or unet.config.attention_head_dim
+        if not attn_heads_match_channel_sizes(num_attention_heads, block_out_channels):
+            raise ValueError(
+                f"The dimension of attention heads ({num_attention_heads}) must divide `block_out_channels` ({block_out_channels}). If you didn't set `num_attention_heads` the default settings don't match your model. Set `num_attention_heads` manually."
+            )
+
+        # - group norms
+        def group_norms_match_channel_sizes(num_groups, channel_sizes):
+            return all(c % num_groups == 0 for c in channel_sizes)
+
+        if norm_num_groups is None:
+            if group_norms_match_channel_sizes(unet.config.norm_num_groups, block_out_channels):
+                norm_num_groups = unet.config.norm_num_groups
+            else:
+                norm_num_groups = min(block_out_channels)
+
+                if group_norms_match_channel_sizes(norm_num_groups, block_out_channels):
+                    print(
+                        f"`norm_num_groups` was set to `min(block_out_channels)` (={norm_num_groups}) so it divides all block_out_channels` ({block_out_channels}). Set it explicitly to remove this information."
+                    )
+                else:
+                    raise ValueError(
+                        f"`block_out_channels` ({block_out_channels}) don't match the base models `norm_num_groups` ({unet.config.norm_num_groups}). Setting `norm_num_groups` to `min(block_out_channels)` ({norm_num_groups}) didn't fix this. Pass `norm_num_groups` explicitly so it divides all block_out_channels."
+                    )
+
+        def get_time_emb_input_dim(unet: UNet2DConditionModel):
+            return unet.time_embedding.linear_1.in_features
+
+        def get_time_emb_dim(unet: UNet2DConditionModel):
+            return unet.time_embedding.linear_2.out_features
+
+        # Clone params from base unet if
+        #    (i)   it's required to build SD or SDXL, and
+        #    (ii)  it's not used for the time embedding (as time embedding of control model is never used), and
+        #    (iii) it's not set further below anyway
+        to_keep = [
+            "cross_attention_dim",
+            "down_block_types",
+            "sample_size",
+            "transformer_layers_per_block",
+            "up_block_types",
+            "upcast_attention",
+        ]
+        kwargs = {k: v for k, v in dict(unet.config).items() if k in to_keep}
+        kwargs.update(block_out_channels=block_out_channels)
+        kwargs.update(num_attention_heads=num_attention_heads)
+        kwargs.update(norm_num_groups=norm_num_groups)
+
+        # Add controlnetxs-specific params
+        kwargs.update(
+            conditioning_channels=conditioning_channels,
+            controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
+            time_embedding_input_dim=get_time_emb_input_dim(unet),
+            time_embedding_dim=get_time_emb_dim(unet),
+            time_embedding_mix=time_embedding_mix,
+            learn_embedding=learn_embedding,
+            base_model_channel_sizes=ControlNetXSModel._gather_subblock_sizes(unet, base_or_control="base"),
+            conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+        )
+
+        return cls(**kwargs)
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        return self.control_model.attn_processors
+
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        self.control_model.set_attn_processor(processor, _remove_lora)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        self.control_model.set_default_attn_processor()
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        self.control_model.set_attention_slice(slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (UNet2DConditionModel)):
+            if value:
+                module.enable_gradient_checkpointing()
+            else:
+                module.disable_gradient_checkpointing()
+
+    def forward(
+        self,
+        base_model: UNet2DConditionModel,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        controlnet_cond: torch.Tensor,
+        conditioning_scale: float = 1.0,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        return_dict: bool = True,
+    ) -> Union[ControlNetXSOutput, Tuple]:
+        """
+        The [`ControlNetModel`] forward method.
+
+        Args:
+            base_model (`UNet2DConditionModel`):
+                The base unet model we want to control.
+            sample (`torch.FloatTensor`):
+                The noisy input tensor.
+            timestep (`Union[torch.Tensor, float, int]`):
+                The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.Tensor`):
+                The encoder hidden states.
+            controlnet_cond (`torch.FloatTensor`):
+                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+            conditioning_scale (`float`, defaults to `1.0`):
+                How much the control model affects the base model outputs.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
+                Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
+                timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
+                embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            added_cond_kwargs (`dict`):
+                Additional conditions for the Stable Diffusion XL UNet.
+            cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
+                A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+            return_dict (`bool`, defaults to `True`):
+                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.controlnetxs.ControlNetXSOutput`] **or** `tuple`:
+                If `return_dict` is `True`, a [`~models.controlnetxs.ControlNetXSOutput`] is returned, otherwise a
+                tuple is returned where the first element is the sample tensor.
+        """
+        # check channel order
+        channel_order = self.config.controlnet_conditioning_channel_order
+
+        if channel_order == "rgb":
+            # in rgb order by default
+            ...
+        elif channel_order == "bgr":
+            controlnet_cond = torch.flip(controlnet_cond, dims=[1])
+        else:
+            raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
+
+        # scale control strength
+        n_connections = len(self.down_zero_convs_out) + 1 + len(self.up_zero_convs_out)
+        scale_list = torch.full((n_connections,), conditioning_scale)
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = base_model.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        if self.config.learn_embedding:
+            ctrl_temb = self.control_model.time_embedding(t_emb, timestep_cond)
+            base_temb = base_model.time_embedding(t_emb, timestep_cond)
+            interpolation_param = self.config.time_embedding_mix**0.3
+
+            temb = ctrl_temb * interpolation_param + base_temb * (1 - interpolation_param)
+        else:
+            temb = base_model.time_embedding(t_emb)
+
+        # added time & text embeddings
+        aug_emb = None
+
+        if base_model.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if base_model.config.class_embed_type == "timestep":
+                class_labels = base_model.time_proj(class_labels)
+
+            class_emb = base_model.class_embedding(class_labels).to(dtype=self.dtype)
+            temb = temb + class_emb
+
+        if base_model.config.addition_embed_type is not None:
+            if base_model.config.addition_embed_type == "text":
+                aug_emb = base_model.add_embedding(encoder_hidden_states)
+            elif base_model.config.addition_embed_type == "text_image":
+                raise NotImplementedError()
+            elif base_model.config.addition_embed_type == "text_time":
+                # SDXL - style
+                if "text_embeds" not in added_cond_kwargs:
+                    raise ValueError(
+                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                    )
+                text_embeds = added_cond_kwargs.get("text_embeds")
+                if "time_ids" not in added_cond_kwargs:
+                    raise ValueError(
+                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                    )
+                time_ids = added_cond_kwargs.get("time_ids")
+                time_embeds = base_model.add_time_proj(time_ids.flatten())
+                time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+                add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+                add_embeds = add_embeds.to(temb.dtype)
+                aug_emb = base_model.add_embedding(add_embeds)
+            elif base_model.config.addition_embed_type == "image":
+                raise NotImplementedError()
+            elif base_model.config.addition_embed_type == "image_hint":
+                raise NotImplementedError()
+
+        temb = temb + aug_emb if aug_emb is not None else temb
+
+        # text embeddings
+        cemb = encoder_hidden_states
+
+        # Preparation
+        guided_hint = self.controlnet_cond_embedding(controlnet_cond)
+
+        h_ctrl = h_base = sample
+        hs_base, hs_ctrl = [], []
+        it_down_convs_in, it_down_convs_out, it_dec_convs_in, it_up_convs_out = map(
+            iter, (self.down_zero_convs_in, self.down_zero_convs_out, self.up_zero_convs_in, self.up_zero_convs_out)
+        )
+        scales = iter(scale_list)
+
+        base_down_subblocks = to_sub_blocks(base_model.down_blocks)
+        ctrl_down_subblocks = to_sub_blocks(self.control_model.down_blocks)
+        base_mid_subblocks = to_sub_blocks([base_model.mid_block])
+        ctrl_mid_subblocks = to_sub_blocks([self.control_model.mid_block])
+        base_up_subblocks = to_sub_blocks(base_model.up_blocks)
+
+        # Cross Control
+        # 0 - conv in
+        h_base = base_model.conv_in(h_base)
+        h_ctrl = self.control_model.conv_in(h_ctrl)
+        if guided_hint is not None:
+            h_ctrl += guided_hint
+        h_base = h_base + next(it_down_convs_out)(h_ctrl) * next(scales)  # D - add ctrl -> base
+
+        hs_base.append(h_base)
+        hs_ctrl.append(h_ctrl)
+
+        # 1 - down
+        for m_base, m_ctrl in zip(base_down_subblocks, ctrl_down_subblocks):
+            h_ctrl = torch.cat([h_ctrl, next(it_down_convs_in)(h_base)], dim=1)  # A - concat base -> ctrl
+            h_base = m_base(h_base, temb, cemb, attention_mask, cross_attention_kwargs)  # B - apply base subblock
+            h_ctrl = m_ctrl(h_ctrl, temb, cemb, attention_mask, cross_attention_kwargs)  # C - apply ctrl subblock
+            h_base = h_base + next(it_down_convs_out)(h_ctrl) * next(scales)  # D - add ctrl -> base
+            hs_base.append(h_base)
+            hs_ctrl.append(h_ctrl)
+
+        # 2 - mid
+        h_ctrl = torch.cat([h_ctrl, next(it_down_convs_in)(h_base)], dim=1)  # A - concat base -> ctrl
+        for m_base, m_ctrl in zip(base_mid_subblocks, ctrl_mid_subblocks):
+            h_base = m_base(h_base, temb, cemb, attention_mask, cross_attention_kwargs)  # B - apply base subblock
+            h_ctrl = m_ctrl(h_ctrl, temb, cemb, attention_mask, cross_attention_kwargs)  # C - apply ctrl subblock
+        h_base = h_base + self.middle_block_out(h_ctrl) * next(scales)  # D - add ctrl -> base
+
+        # 3 - up
+        for i, m_base in enumerate(base_up_subblocks):
+            h_base = h_base + next(it_up_convs_out)(hs_ctrl.pop()) * next(scales)  # add info from ctrl encoder
+            h_base = torch.cat([h_base, hs_base.pop()], dim=1)  # concat info from base encoder+ctrl encoder
+            h_base = m_base(h_base, temb, cemb, attention_mask, cross_attention_kwargs)
+
+        h_base = base_model.conv_norm_out(h_base)
+        h_base = base_model.conv_act(h_base)
+        h_base = base_model.conv_out(h_base)
+
+        if not return_dict:
+            return h_base
+
+        return ControlNetXSOutput(sample=h_base)
+
+    def _make_zero_conv(self, in_channels, out_channels=None):
+        # keep running track of channels sizes
+        self.in_channels = in_channels
+        self.out_channels = out_channels or in_channels
+
+        return zero_module(nn.Conv2d(in_channels, out_channels, 1, padding=0))
+
+    @torch.no_grad()
+    def _check_if_vae_compatible(self, vae: AutoencoderKL):
+        condition_downscale_factor = 2 ** (len(self.config.conditioning_embedding_out_channels) - 1)
+        vae_downscale_factor = 2 ** (len(vae.config.block_out_channels) - 1)
+        compatible = condition_downscale_factor == vae_downscale_factor
+        return compatible, condition_downscale_factor, vae_downscale_factor
+
+
+class SubBlock(nn.ModuleList):
+    """A SubBlock is the largest piece of either base or control model, that is executed independently of the other model respectively.
+    Before each subblock, information is concatted from base to control. And after each subblock, information is added from control to base.
+    """
+
+    def __init__(self, ms, *args, **kwargs):
+        if not is_iterable(ms):
+            ms = [ms]
+        super().__init__(ms, *args, **kwargs)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        temb: torch.Tensor,
+        cemb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+    ):
+        """Iterate through children and pass correct information to each."""
+        for m in self:
+            if isinstance(m, ResnetBlock2D):
+                x = m(x, temb)
+            elif isinstance(m, Transformer2DModel):
+                x = m(x, cemb, attention_mask=attention_mask, cross_attention_kwargs=cross_attention_kwargs).sample
+            elif isinstance(m, Downsample2D):
+                x = m(x)
+            elif isinstance(m, Upsample2D):
+                x = m(x)
+            else:
+                raise ValueError(
+                    f"Type of m is {type(m)} but should be `ResnetBlock2D`, `Transformer2DModel`,  `Downsample2D` or `Upsample2D`"
+                )
+
+        return x
+
+
+def adjust_time_dims(unet: UNet2DConditionModel, in_dim: int, out_dim: int):
+    unet.time_embedding.linear_1 = nn.Linear(in_dim, out_dim)
+
+
+def increase_block_input_in_encoder_resnet(unet: UNet2DConditionModel, block_no, resnet_idx, by):
+    """Increase channels sizes to allow for additional concatted information from base model"""
+    r = unet.down_blocks[block_no].resnets[resnet_idx]
+    old_norm1, old_conv1 = r.norm1, r.conv1
+    # norm
+    norm_args = "num_groups num_channels eps affine".split(" ")
+    for a in norm_args:
+        assert hasattr(old_norm1, a)
+    norm_kwargs = {a: getattr(old_norm1, a) for a in norm_args}
+    norm_kwargs["num_channels"] += by  # surgery done here
+    # conv1
+    conv1_args = (
+        "in_channels out_channels kernel_size stride padding dilation groups bias padding_mode lora_layer".split(" ")
+    )
+    for a in conv1_args:
+        assert hasattr(old_conv1, a)
+    conv1_kwargs = {a: getattr(old_conv1, a) for a in conv1_args}
+    conv1_kwargs["bias"] = "bias" in conv1_kwargs  # as param, bias is a boolean, but as attr, it's a tensor.
+    conv1_kwargs["in_channels"] += by  # surgery done here
+    # conv_shortcut
+    # as we changed the input size of the block, the input and output sizes are likely different,
+    # therefore we need a conv_shortcut (simply adding won't work)
+    conv_shortcut_args_kwargs = {
+        "in_channels": conv1_kwargs["in_channels"],
+        "out_channels": conv1_kwargs["out_channels"],
+        # default arguments from resnet.__init__
+        "kernel_size": 1,
+        "stride": 1,
+        "padding": 0,
+        "bias": True,
+    }
+    # swap old with new modules
+    unet.down_blocks[block_no].resnets[resnet_idx].norm1 = GroupNorm(**norm_kwargs)
+    unet.down_blocks[block_no].resnets[resnet_idx].conv1 = LoRACompatibleConv(**conv1_kwargs)
+    unet.down_blocks[block_no].resnets[resnet_idx].conv_shortcut = LoRACompatibleConv(**conv_shortcut_args_kwargs)
+    unet.down_blocks[block_no].resnets[resnet_idx].in_channels += by  # surgery done here
+
+
+def increase_block_input_in_encoder_downsampler(unet: UNet2DConditionModel, block_no, by):
+    """Increase channels sizes to allow for additional concatted information from base model"""
+    old_down = unet.down_blocks[block_no].downsamplers[0].conv
+    # conv1
+    args = "in_channels out_channels kernel_size stride padding dilation groups bias padding_mode lora_layer".split(
+        " "
+    )
+    for a in args:
+        assert hasattr(old_down, a)
+    kwargs = {a: getattr(old_down, a) for a in args}
+    kwargs["bias"] = "bias" in kwargs  # as param, bias is a boolean, but as attr, it's a tensor.
+    kwargs["in_channels"] += by  # surgery done here
+    # swap old with new modules
+    unet.down_blocks[block_no].downsamplers[0].conv = LoRACompatibleConv(**kwargs)
+    unet.down_blocks[block_no].downsamplers[0].channels += by  # surgery done here
+
+
+def increase_block_input_in_mid_resnet(unet: UNet2DConditionModel, by):
+    """Increase channels sizes to allow for additional concatted information from base model"""
+    m = unet.mid_block.resnets[0]
+    old_norm1, old_conv1 = m.norm1, m.conv1
+    # norm
+    norm_args = "num_groups num_channels eps affine".split(" ")
+    for a in norm_args:
+        assert hasattr(old_norm1, a)
+    norm_kwargs = {a: getattr(old_norm1, a) for a in norm_args}
+    norm_kwargs["num_channels"] += by  # surgery done here
+    # conv1
+    conv1_args = (
+        "in_channels out_channels kernel_size stride padding dilation groups bias padding_mode lora_layer".split(" ")
+    )
+    for a in conv1_args:
+        assert hasattr(old_conv1, a)
+    conv1_kwargs = {a: getattr(old_conv1, a) for a in conv1_args}
+    conv1_kwargs["bias"] = "bias" in conv1_kwargs  # as param, bias is a boolean, but as attr, it's a tensor.
+    conv1_kwargs["in_channels"] += by  # surgery done here
+    # conv_shortcut
+    # as we changed the input size of the block, the input and output sizes are likely different,
+    # therefore we need a conv_shortcut (simply adding won't work)
+    conv_shortcut_args_kwargs = {
+        "in_channels": conv1_kwargs["in_channels"],
+        "out_channels": conv1_kwargs["out_channels"],
+        # default arguments from resnet.__init__
+        "kernel_size": 1,
+        "stride": 1,
+        "padding": 0,
+        "bias": True,
+    }
+    # swap old with new modules
+    unet.mid_block.resnets[0].norm1 = GroupNorm(**norm_kwargs)
+    unet.mid_block.resnets[0].conv1 = LoRACompatibleConv(**conv1_kwargs)
+    unet.mid_block.resnets[0].conv_shortcut = LoRACompatibleConv(**conv_shortcut_args_kwargs)
+    unet.mid_block.resnets[0].in_channels += by  # surgery done here
+
+
+def adjust_group_norms(unet: UNet2DConditionModel, max_num_group: int = 32):
+    def find_denominator(number, start):
+        if start >= number:
+            return number
+        while start != 0:
+            residual = number % start
+            if residual == 0:
+                return start
+            start -= 1
+
+    for block in [*unet.down_blocks, unet.mid_block]:
+        # resnets
+        for r in block.resnets:
+            if r.norm1.num_groups < max_num_group:
+                r.norm1.num_groups = find_denominator(r.norm1.num_channels, start=max_num_group)
+
+            if r.norm2.num_groups < max_num_group:
+                r.norm2.num_groups = find_denominator(r.norm2.num_channels, start=max_num_group)
+
+        # transformers
+        if hasattr(block, "attentions"):
+            for a in block.attentions:
+                if a.norm.num_groups < max_num_group:
+                    a.norm.num_groups = find_denominator(a.norm.num_channels, start=max_num_group)
+
+
+def is_iterable(o):
+    if isinstance(o, str):
+        return False
+    try:
+        iter(o)
+        return True
+    except TypeError:
+        return False
+
+
+def to_sub_blocks(blocks):
+    if not is_iterable(blocks):
+        blocks = [blocks]
+
+    sub_blocks = []
+
+    for b in blocks:
+        if hasattr(b, "resnets"):
+            if hasattr(b, "attentions") and b.attentions is not None:
+                for r, a in zip(b.resnets, b.attentions):
+                    sub_blocks.append([r, a])
+
+                num_resnets = len(b.resnets)
+                num_attns = len(b.attentions)
+
+                if num_resnets > num_attns:
+                    # we can have more resnets than attentions, so add each resnet as separate subblock
+                    for i in range(num_attns, num_resnets):
+                        sub_blocks.append([b.resnets[i]])
+            else:
+                for r in b.resnets:
+                    sub_blocks.append([r])
+
+        # upsamplers are part of the same subblock
+        if hasattr(b, "upsamplers") and b.upsamplers is not None:
+            for u in b.upsamplers:
+                sub_blocks[-1].extend([u])
+
+        # downsamplers are own subblock
+        if hasattr(b, "downsamplers") and b.downsamplers is not None:
+            for d in b.downsamplers:
+                sub_blocks.append([d])
+
+    return list(map(SubBlock, sub_blocks))
+
+
+def zero_module(module):
+    for p in module.parameters():
+        nn.init.zeros_(p)
+    return module

src/diffusers/models/embeddings.py

@@ -461,6 +461,18 @@ class ImageProjection(nn.Module):
         return image_embeds
 
 
+class MLPProjection(nn.Module):
+    def __init__(self, image_embed_dim=1024, cross_attention_dim=1024):
+        super().__init__()
+        from .attention import FeedForward
+
+        self.ff = FeedForward(image_embed_dim, cross_attention_dim, mult=1, activation_fn="gelu")
+        self.norm = nn.LayerNorm(cross_attention_dim)
+
+    def forward(self, image_embeds: torch.FloatTensor):
+        return self.norm(self.ff(image_embeds))
+
+
 class CombinedTimestepLabelEmbeddings(nn.Module):
     def __init__(self, num_classes, embedding_dim, class_dropout_prob=0.1):
         super().__init__()

src/diffusers/models/modeling_flax_utils.py

@@ -24,13 +24,17 @@ from flax.core.frozen_dict import FrozenDict, unfreeze
 from flax.serialization import from_bytes, to_bytes
 from flax.traverse_util import flatten_dict, unflatten_dict
 from huggingface_hub import create_repo, hf_hub_download
-from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError
+from huggingface_hub.utils import (
+    EntryNotFoundError,
+    RepositoryNotFoundError,
+    RevisionNotFoundError,
+    validate_hf_hub_args,
+)
 from requests import HTTPError
 
 from .. import __version__, is_torch_available
 from ..utils import (
     CONFIG_NAME,
-    DIFFUSERS_CACHE,
     FLAX_WEIGHTS_NAME,
     HUGGINGFACE_CO_RESOLVE_ENDPOINT,
     WEIGHTS_NAME,
@@ -197,6 +201,7 @@ class FlaxModelMixin(PushToHubMixin):
         raise NotImplementedError(f"init_weights method has to be implemented for {self}")
 
     @classmethod
+    @validate_hf_hub_args
     def from_pretrained(
         cls,
         pretrained_model_name_or_path: Union[str, os.PathLike],
@@ -288,13 +293,13 @@ class FlaxModelMixin(PushToHubMixin):
         ```
         """
         config = kwargs.pop("config", None)
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         from_pt = kwargs.pop("from_pt", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         subfolder = kwargs.pop("subfolder", None)
 
@@ -314,7 +319,7 @@ class FlaxModelMixin(PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 subfolder=subfolder,
                 **kwargs,
@@ -359,7 +364,7 @@ class FlaxModelMixin(PushToHubMixin):
                     proxies=proxies,
                     resume_download=resume_download,
                     local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
+                    token=token,
                     user_agent=user_agent,
                     subfolder=subfolder,
                     revision=revision,
@@ -369,7 +374,7 @@ class FlaxModelMixin(PushToHubMixin):
                 raise EnvironmentError(
                     f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier "
                     "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a "
-                    "token having permission to this repo with `use_auth_token` or log in with `huggingface-cli "
+                    "token having permission to this repo with `token` or log in with `huggingface-cli "
                     "login`."
                 )
             except RevisionNotFoundError:

src/diffusers/models/modeling_utils.py

@@ -25,14 +25,13 @@ from typing import Any, Callable, List, Optional, Tuple, Union
 import safetensors
 import torch
 from huggingface_hub import create_repo
+from huggingface_hub.utils import validate_hf_hub_args
 from torch import Tensor, nn
 
 from .. import __version__
 from ..utils import (
     CONFIG_NAME,
-    DIFFUSERS_CACHE,
     FLAX_WEIGHTS_NAME,
-    HF_HUB_OFFLINE,
     MIN_PEFT_VERSION,
     SAFETENSORS_WEIGHTS_NAME,
     WEIGHTS_NAME,
@@ -535,6 +534,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             )
 
     @classmethod
+    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
         r"""
         Instantiate a pretrained PyTorch model from a pretrained model configuration.
@@ -571,7 +571,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             local_files_only(`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -640,15 +640,15 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
         force_download = kwargs.pop("force_download", False)
         from_flax = kwargs.pop("from_flax", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         output_loading_info = kwargs.pop("output_loading_info", False)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         torch_dtype = kwargs.pop("torch_dtype", None)
         subfolder = kwargs.pop("subfolder", None)
@@ -718,7 +718,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             resume_download=resume_download,
             proxies=proxies,
             local_files_only=local_files_only,
-            use_auth_token=use_auth_token,
+            token=token,
             revision=revision,
             subfolder=subfolder,
             device_map=device_map,
@@ -740,7 +740,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 subfolder=subfolder,
                 user_agent=user_agent,
@@ -763,7 +763,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        use_auth_token=use_auth_token,
+                        token=token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -782,7 +782,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
+                    token=token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,

src/diffusers/pipelines/README.md

@@ -1,33 +1,33 @@
 # 🧨 Diffusers Pipelines
 
 Pipelines provide a simple way to run state-of-the-art diffusion models in inference.
-Most diffusion systems consist of multiple independently-trained models and highly adaptable scheduler 
+Most diffusion systems consist of multiple independently-trained models and highly adaptable scheduler
 components - all of which are needed to have a functioning end-to-end diffusion system.
 
 As an example, [Stable Diffusion](https://huggingface.co/blog/stable_diffusion) has three independently trained models:
 - [Autoencoder](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/models/vae.py#L392)
 - [Conditional Unet](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/models/unet_2d_condition.py#L12)
 - [CLIP text encoder](https://huggingface.co/docs/transformers/main/en/model_doc/clip#transformers.CLIPTextModel)
-- a scheduler component, [scheduler](https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_pndm.py), 
+- a scheduler component, [scheduler](https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_pndm.py),
 - a [CLIPImageProcessor](https://huggingface.co/docs/transformers/main/en/model_doc/clip#transformers.CLIPImageProcessor),
 - as well as a [safety checker](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/safety_checker.py).
-All of these components are necessary to run stable diffusion in inference even though they were trained 
+All of these components are necessary to run stable diffusion in inference even though they were trained
 or created independently from each other.
 
-To that end, we strive to offer all open-sourced, state-of-the-art diffusion system under a unified API. 
+To that end, we strive to offer all open-sourced, state-of-the-art diffusion system under a unified API.
 More specifically, we strive to provide pipelines that
 - 1. can load the officially published weights and yield 1-to-1 the same outputs as the original implementation according to the corresponding paper (*e.g.* [LDMTextToImagePipeline](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines/latent_diffusion), uses the officially released weights of [High-Resolution Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752)),
-- 2. have a simple user interface to run the model in inference (see the [Pipelines API](#pipelines-api) section), 
+- 2. have a simple user interface to run the model in inference (see the [Pipelines API](#pipelines-api) section),
 - 3. are easy to understand with code that is self-explanatory and can be read along-side the official paper (see [Pipelines summary](#pipelines-summary)),
 - 4. can easily be contributed by the community (see the [Contribution](#contribution) section).
 
-**Note** that pipelines do not (and should not) offer any training functionality. 
+**Note** that pipelines do not (and should not) offer any training functionality.
 If you are looking for *official* training examples, please have a look at [examples](https://github.com/huggingface/diffusers/tree/main/examples).
 
 
 ## Pipelines Summary
 
-The following table summarizes all officially supported pipelines, their corresponding paper, and if 
+The following table summarizes all officially supported pipelines, their corresponding paper, and if
 available a colab notebook to directly try them out.
 
 | Pipeline                                                                                                                      | Source                                                                                                                       | Tasks | Colab
@@ -35,35 +35,35 @@ available a colab notebook to directly try them out.
 | [dance diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/dance_diffusion)                 | [**Dance Diffusion**](https://github.com/Harmonai-org/sample-generator)                                                      | *Unconditional Audio Generation* |
 | [ddpm](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/ddpm)                                       | [**Denoising Diffusion Probabilistic Models**](https://arxiv.org/abs/2006.11239)                                             | *Unconditional Image Generation* |
 | [ddim](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/ddim)                                       | [**Denoising Diffusion Implicit Models**](https://arxiv.org/abs/2010.02502)                                                  | *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)
-| [latent_diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion)               | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752)                         | *Text-to-Image Generation* | 
-| [latent_diffusion_uncond](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion_uncond) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752)                         | *Unconditional Image Generation* | 
-| [pndm](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pndm)                                       | [**Pseudo Numerical Methods for Diffusion Models on Manifolds**](https://arxiv.org/abs/2202.09778)                           | *Unconditional Image Generation* | 
-| [score_sde_ve](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/score_sde_ve)                       | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | *Unconditional Image Generation* | 
-| [score_sde_vp](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/score_sde_vp)                       | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | *Unconditional Image Generation* | 
+| [latent_diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion)               | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752)                         | *Text-to-Image Generation* |
+| [latent_diffusion_uncond](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion_uncond) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752)                         | *Unconditional Image Generation* |
+| [pndm](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pndm)                                       | [**Pseudo Numerical Methods for Diffusion Models on Manifolds**](https://arxiv.org/abs/2202.09778)                           | *Unconditional Image Generation* |
+| [score_sde_ve](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/score_sde_ve)                       | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | *Unconditional Image Generation* |
+| [score_sde_vp](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/score_sde_vp)                       | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | *Unconditional Image Generation* |
 | [stable_diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion)               | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release)                                            | *Text-to-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/stable_diffusion.ipynb)
 | [stable_diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion)               | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release)                                            | *Image-to-Image Text-Guided Generation* | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb)
 | [stable_diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion)               | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release)                                            | *Text-Guided Image Inpainting* | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)
-| [stochastic_karras_ve](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stochastic_karras_ve)       | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364)                    | *Unconditional Image Generation* | 
+| [stochastic_karras_ve](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stochastic_karras_ve)       | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364)                    | *Unconditional Image Generation* |
 
-**Note**: Pipelines are simple examples of how to play around with the diffusion systems as described in the corresponding papers. 
+**Note**: Pipelines are simple examples of how to play around with the diffusion systems as described in the corresponding papers.
 However, most of them can be adapted to use different scheduler components or even different model components. Some pipeline examples are shown in the [Examples](#examples) below.
 
 ## Pipelines API
 
-Diffusion models often consist of multiple independently-trained models or other previously existing components. 
+Diffusion models often consist of multiple independently-trained models or other previously existing components.
 
 
-Each model has been trained independently on a different task and the scheduler can easily be swapped out and replaced with a different one. 
+Each model has been trained independently on a different task and the scheduler can easily be swapped out and replaced with a different one.
 During inference, we however want to be able to easily load all components and use them in inference - even if one component, *e.g.* CLIP's text encoder, originates from a different library, such as [Transformers](https://github.com/huggingface/transformers). To that end, all pipelines provide the following functionality:
 
 - [`from_pretrained` method](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L139) that accepts a Hugging Face Hub repository id, *e.g.* [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5) or a path to a local directory, *e.g.*
-"./stable-diffusion". To correctly retrieve which models and components should be loaded, one has to provide a `model_index.json` file, *e.g.* [runwayml/stable-diffusion-v1-5/model_index.json](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/model_index.json), which defines all components that should be 
+"./stable-diffusion". To correctly retrieve which models and components should be loaded, one has to provide a `model_index.json` file, *e.g.* [runwayml/stable-diffusion-v1-5/model_index.json](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/model_index.json), which defines all components that should be
 loaded into the pipelines. More specifically, for each model/component one needs to define the format `<name>: ["<library>", "<class name>"]`. `<name>` is the attribute name given to the loaded instance of `<class name>` which can be found in the library or pipeline folder called `"<library>"`.
-- [`save_pretrained`](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L90) that accepts a local path, *e.g.* `./stable-diffusion` under which all models/components of the pipeline will be saved. For each component/model a folder is created inside the local path that is named after the given attribute name, *e.g.* `./stable_diffusion/unet`. 
-In addition, a `model_index.json` file is created at the root of the local path, *e.g.* `./stable_diffusion/model_index.json` so that the complete pipeline can again be instantiated 
+- [`save_pretrained`](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L90) that accepts a local path, *e.g.* `./stable-diffusion` under which all models/components of the pipeline will be saved. For each component/model a folder is created inside the local path that is named after the given attribute name, *e.g.* `./stable_diffusion/unet`.
+In addition, a `model_index.json` file is created at the root of the local path, *e.g.* `./stable_diffusion/model_index.json` so that the complete pipeline can again be instantiated
 from the local path.
 - [`to`](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L118) which accepts a `string` or `torch.device` to move all models that are of type `torch.nn.Module` to the passed device. The behavior is fully analogous to [PyTorch's `to` method](https://pytorch.org/docs/stable/generated/torch.nn.Module.html#torch.nn.Module.to).
-- [`__call__`] method to use the pipeline in inference. `__call__` defines inference logic of the pipeline and should ideally encompass all aspects of it, from pre-processing to forwarding tensors to the different models and schedulers, as well as post-processing. The API of the `__call__` method can strongly vary from pipeline to pipeline. *E.g.* a text-to-image pipeline, such as [`StableDiffusionPipeline`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py) should accept among other things the text prompt to generate the image. A pure image generation pipeline, such as [DDPMPipeline](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines/ddpm) on the other hand can be run without providing any inputs. To better understand what inputs can be adapted for 
+- [`__call__`] method to use the pipeline in inference. `__call__` defines inference logic of the pipeline and should ideally encompass all aspects of it, from pre-processing to forwarding tensors to the different models and schedulers, as well as post-processing. The API of the `__call__` method can strongly vary from pipeline to pipeline. *E.g.* a text-to-image pipeline, such as [`StableDiffusionPipeline`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py) should accept among other things the text prompt to generate the image. A pure image generation pipeline, such as [DDPMPipeline](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines/ddpm) on the other hand can be run without providing any inputs. To better understand what inputs can be adapted for
 each pipeline, one should look directly into the respective pipeline.
 
 **Note**: All pipelines have PyTorch's autograd disabled by decorating the `__call__` method with a [`torch.no_grad`](https://pytorch.org/docs/stable/generated/torch.no_grad.html) decorator because pipelines should
@@ -71,12 +71,12 @@ not be used for training. If you want to store the gradients during the forward
 
 ## Contribution
 
-We are more than happy about any contribution to the officially supported pipelines 🤗. We aspire 
+We are more than happy about any contribution to the officially supported pipelines 🤗. We aspire
 all of our pipelines to be  **self-contained**, **easy-to-tweak**, **beginner-friendly** and for **one-purpose-only**.
 
-- **Self-contained**: A pipeline shall be as self-contained as possible. More specifically, this means that all functionality should be either directly defined in the pipeline file itself, should be inherited from (and only from) the [`DiffusionPipeline` class](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L56) or be directly attached to the model and scheduler components of the pipeline. 
-- **Easy-to-use**: Pipelines should be extremely easy to use - one should be able to load the pipeline and 
-use it for its designated task, *e.g.* text-to-image generation, in just a couple of lines of code. Most 
+- **Self-contained**: A pipeline shall be as self-contained as possible. More specifically, this means that all functionality should be either directly defined in the pipeline file itself, should be inherited from (and only from) the [`DiffusionPipeline` class](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L56) or be directly attached to the model and scheduler components of the pipeline.
+- **Easy-to-use**: Pipelines should be extremely easy to use - one should be able to load the pipeline and
+use it for its designated task, *e.g.* text-to-image generation, in just a couple of lines of code. Most
 logic including pre-processing, an unrolled diffusion loop, and post-processing should all happen inside the `__call__` method.
 - **Easy-to-tweak**: Certain pipelines will not be able to handle all use cases and tasks that you might like them to. If you want to use a certain pipeline for a specific use case that is not yet supported, you might have to copy the pipeline file and tweak the code to your needs. We try to make the pipeline code as readable as possible so that each part –from pre-processing to diffusing to post-processing– can easily be adapted. If you would like the community to benefit from your customized pipeline, we would love to see a contribution to our [community-examples](https://github.com/huggingface/diffusers/tree/main/examples/community). If you feel that an important pipeline should be part of the official pipelines but isn't, a contribution to the [official pipelines](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines) would be even better.
 - **One-purpose-only**: Pipelines should be used for one task and one task only. Even if two tasks are very similar from a modeling point of view, *e.g.* image2image translation and in-painting, pipelines shall be used for one task only to keep them *easy-to-tweak* and *readable*.
@@ -93,8 +93,8 @@ pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
 pipe = pipe.to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
-image = pipe(prompt).images[0]  
-    
+image = pipe(prompt).images[0]
+
 image.save("astronaut_rides_horse.png")
 ```
 

src/diffusers/pipelines/__init__.py

@@ -19,6 +19,7 @@ from ..utils import (
 _dummy_objects = {}
 _import_structure = {
     "controlnet": [],
+    "controlnet_xs": [],
     "latent_diffusion": [],
     "stable_diffusion": [],
     "stable_diffusion_xl": [],
@@ -93,6 +94,12 @@ else:
             "StableDiffusionXLControlNetPipeline",
         ]
     )
+    _import_structure["controlnet_xs"].extend(
+        [
+            "StableDiffusionControlNetXSPipeline",
+            "StableDiffusionXLControlNetXSPipeline",
+        ]
+    )
     _import_structure["deepfloyd_if"] = [
         "IFImg2ImgPipeline",
         "IFImg2ImgSuperResolutionPipeline",
@@ -347,6 +354,10 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionXLControlNetInpaintPipeline,
             StableDiffusionXLControlNetPipeline,
         )
+        from .controlnet_xs import (
+            StableDiffusionControlNetXSPipeline,
+            StableDiffusionXLControlNetXSPipeline,
+        )
         from .deepfloyd_if import (
             IFImg2ImgPipeline,
             IFImg2ImgSuperResolutionPipeline,

src/diffusers/pipelines/auto_pipeline.py

@@ -16,8 +16,9 @@
 import inspect
 from collections import OrderedDict
 
+from huggingface_hub.utils import validate_hf_hub_args
+
 from ..configuration_utils import ConfigMixin
-from ..utils import DIFFUSERS_CACHE
 from .controlnet import (
     StableDiffusionControlNetImg2ImgPipeline,
     StableDiffusionControlNetInpaintPipeline,
@@ -195,6 +196,7 @@ class AutoPipelineForText2Image(ConfigMixin):
         )
 
     @classmethod
+    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_or_path, **kwargs):
         r"""
         Instantiates a text-to-image Pytorch diffusion pipeline from pretrained pipeline weight.
@@ -246,7 +248,7 @@ class AutoPipelineForText2Image(ConfigMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -310,11 +312,11 @@ class AutoPipelineForText2Image(ConfigMixin):
         >>> image = pipeline(prompt).images[0]
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
 
@@ -323,7 +325,7 @@ class AutoPipelineForText2Image(ConfigMixin):
             "force_download": force_download,
             "resume_download": resume_download,
             "proxies": proxies,
-            "use_auth_token": use_auth_token,
+            "token": token,
             "local_files_only": local_files_only,
             "revision": revision,
         }
@@ -466,6 +468,7 @@ class AutoPipelineForImage2Image(ConfigMixin):
         )
 
     @classmethod
+    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_or_path, **kwargs):
         r"""
         Instantiates a image-to-image Pytorch diffusion pipeline from pretrained pipeline weight.
@@ -518,7 +521,7 @@ class AutoPipelineForImage2Image(ConfigMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -582,11 +585,11 @@ class AutoPipelineForImage2Image(ConfigMixin):
         >>> image = pipeline(prompt, image).images[0]
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
 
@@ -595,7 +598,7 @@ class AutoPipelineForImage2Image(ConfigMixin):
             "force_download": force_download,
             "resume_download": resume_download,
             "proxies": proxies,
-            "use_auth_token": use_auth_token,
+            "token": token,
             "local_files_only": local_files_only,
             "revision": revision,
         }
@@ -742,6 +745,7 @@ class AutoPipelineForInpainting(ConfigMixin):
         )
 
     @classmethod
+    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_or_path, **kwargs):
         r"""
         Instantiates a inpainting Pytorch diffusion pipeline from pretrained pipeline weight.
@@ -793,7 +797,7 @@ class AutoPipelineForInpainting(ConfigMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -857,11 +861,11 @@ class AutoPipelineForInpainting(ConfigMixin):
         >>> image = pipeline(prompt, image=init_image, mask_image=mask_image).images[0]
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
 
@@ -870,7 +874,7 @@ class AutoPipelineForInpainting(ConfigMixin):
             "force_download": force_download,
             "resume_download": resume_download,
             "proxies": proxies,
-            "use_auth_token": use_auth_token,
+            "token": token,
             "local_files_only": local_files_only,
             "revision": revision,
         }

src/diffusers/pipelines/controlnet/pipeline_controlnet_img2img.py

@@ -19,10 +19,10 @@ import numpy as np
 import PIL.Image
 import torch
 import torch.nn.functional as F
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ControlNetModel, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
@@ -130,7 +130,7 @@ def prepare_image(image):
 
 
 class StableDiffusionControlNetImg2ImgPipeline(
-    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for image-to-image generation using Stable Diffusion with ControlNet guidance.
@@ -140,7 +140,7 @@ class StableDiffusionControlNetImg2ImgPipeline(
 
     The pipeline also inherits the following loading methods:
         - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
-
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
@@ -166,7 +166,7 @@ class StableDiffusionControlNetImg2ImgPipeline(
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
     _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
 
@@ -180,6 +180,7 @@ class StableDiffusionControlNetImg2ImgPipeline(
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: CLIPVisionModelWithProjection = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -212,6 +213,7 @@ class StableDiffusionControlNetImg2ImgPipeline(
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
@@ -468,6 +470,31 @@ class StableDiffusionControlNetImg2ImgPipeline(
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:
@@ -861,6 +888,7 @@ class StableDiffusionControlNetImg2ImgPipeline(
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
@@ -922,6 +950,7 @@ class StableDiffusionControlNetImg2ImgPipeline(
             negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -1053,6 +1082,11 @@ class StableDiffusionControlNetImg2ImgPipeline(
         if self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
 
+        if ip_adapter_image is not None:
+            image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, device, num_images_per_prompt)
+            if self.do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 4. Prepare image
         image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
 
@@ -1111,7 +1145,10 @@ class StableDiffusionControlNetImg2ImgPipeline(
         # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
-        # 7.1 Create tensor stating which controlnets to keep
+        # 7.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
+        # 7.2 Create tensor stating which controlnets to keep
         controlnet_keep = []
         for i in range(len(timesteps)):
             keeps = [
@@ -1171,6 +1208,7 @@ class StableDiffusionControlNetImg2ImgPipeline(
                     cross_attention_kwargs=self.cross_attention_kwargs,
                     down_block_additional_residuals=down_block_res_samples,
                     mid_block_additional_residual=mid_block_res_sample,
+                    added_cond_kwargs=added_cond_kwargs,
                     return_dict=False,
                 )[0]
 

src/diffusers/pipelines/controlnet_xs/__init__.py

@@ -0,0 +1,68 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_flax_available,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_controlnet_xs"] = ["StableDiffusionControlNetXSPipeline"]
+    _import_structure["pipeline_controlnet_xs_sd_xl"] = ["StableDiffusionXLControlNetXSPipeline"]
+try:
+    if not (is_transformers_available() and is_flax_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_flax_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_flax_and_transformers_objects))
+else:
+    pass  # _import_structure["pipeline_flax_controlnet"] = ["FlaxStableDiffusionControlNetPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_controlnet_xs import StableDiffusionControlNetXSPipeline
+        from .pipeline_controlnet_xs_sd_xl import StableDiffusionXLControlNetXSPipeline
+
+    try:
+        if not (is_transformers_available() and is_flax_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_flax_and_transformers_objects import *  # noqa F403
+    else:
+        pass  # from .pipeline_flax_controlnet import FlaxStableDiffusionControlNetPipeline
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs.py

@@ -0,0 +1,944 @@
+# 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.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ControlNetXSModel, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+    USE_PEFT_BACKEND,
+    deprecate,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> # !pip install opencv-python transformers accelerate
+        >>> from diffusers import StableDiffusionControlNetXSPipeline, ControlNetXSModel
+        >>> from diffusers.utils import load_image
+        >>> import numpy as np
+        >>> import torch
+
+        >>> import cv2
+        >>> from PIL import Image
+
+        >>> prompt = "aerial view, a futuristic research complex in a bright foggy jungle, hard lighting"
+        >>> negative_prompt = "low quality, bad quality, sketches"
+
+        >>> # download an image
+        >>> image = load_image(
+        ...    "https://hf.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png"
+        ... )
+
+        >>> # initialize the models and pipeline
+        >>> controlnet_conditioning_scale = 0.5
+        >>> controlnet = ControlNetXSModel.from_pretrained(
+        ...     "UmerHA/ConrolNetXS-SD2.1-canny", torch_dtype=torch.float16
+        ... )
+        >>> pipe = StableDiffusionControlNetXSPipeline.from_pretrained(
+        ...     "stabilityai/stable-diffusion-2-1", controlnet=controlnet, torch_dtype=torch.float16
+        ... )
+        >>> pipe.enable_model_cpu_offload()
+
+        >>> # get canny image
+        >>> image = np.array(image)
+        >>> image = cv2.Canny(image, 100, 200)
+        >>> image = image[:, :, None]
+        >>> image = np.concatenate([image, image, image], axis=2)
+        >>> canny_image = Image.fromarray(image)
+        >>> # generate image
+        >>> image = pipe(
+        ...     prompt, controlnet_conditioning_scale=controlnet_conditioning_scale, image=canny_image
+        ... ).images[0]
+        ```
+"""
+
+
+class StableDiffusionControlNetXSPipeline(
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion with ControlNet-XS guidance.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+        text_encoder ([`~transformers.CLIPTextModel`]):
+            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+        tokenizer ([`~transformers.CLIPTokenizer`]):
+            A `CLIPTokenizer` to tokenize text.
+        unet ([`UNet2DConditionModel`]):
+            A `UNet2DConditionModel` to denoise the encoded image latents.
+        controlnet ([`ControlNetXSModel`]):
+            Provides additional conditioning to the `unet` during the denoising process.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+            about a model's potential harms.
+        feature_extractor ([`~transformers.CLIPImageProcessor`]):
+            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+    """
+
+    model_cpu_offload_seq = "text_encoder->unet->vae>controlnet"
+    _optional_components = ["safety_checker", "feature_extractor"]
+    _exclude_from_cpu_offload = ["safety_checker"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        controlnet: ControlNetXSModel,
+        scheduler: KarrasDiffusionSchedulers,
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPImageProcessor,
+        requires_safety_checker: bool = True,
+    ):
+        super().__init__()
+
+        if safety_checker is None and requires_safety_checker:
+            logger.warning(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
+        vae_compatible, cnxs_condition_downsample_factor, vae_downsample_factor = controlnet._check_if_vae_compatible(
+            vae
+        )
+        if not vae_compatible:
+            raise ValueError(
+                f"The downsampling factors of the VAE ({vae_downsample_factor}) and the conditioning part of ControlNetXS model {cnxs_condition_downsample_factor} need to be equal. Consider building the ControlNetXS model with different `conditioning_block_sizes`."
+            )
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            controlnet=controlnet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+        self.control_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+        )
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+    def _encode_prompt(
+        self,
+        prompt,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+        **kwargs,
+    ):
+        deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+        deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+        prompt_embeds_tuple = self.encode_prompt(
+            prompt=prompt,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=lora_scale,
+            **kwargs,
+        )
+
+        # concatenate for backwards comp
+        prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+        clip_skip: Optional[int] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            lora_scale (`float`, *optional*):
+                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if not USE_PEFT_BACKEND:
+                adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+            else:
+                scale_lora_layers(self.text_encoder, lora_scale)
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            # textual inversion: procecss multi-vector tokens if necessary
+            if isinstance(self, TextualInversionLoaderMixin):
+                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=self.tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                text_input_ids, untruncated_ids
+            ):
+                removed_text = self.tokenizer.batch_decode(
+                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+                )
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+                )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = text_inputs.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            if clip_skip is None:
+                prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+                prompt_embeds = prompt_embeds[0]
+            else:
+                prompt_embeds = self.text_encoder(
+                    text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+                )
+                # Access the `hidden_states` first, that contains a tuple of
+                # all the hidden states from the encoder layers. Then index into
+                # the tuple to access the hidden states from the desired layer.
+                prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+                # We also need to apply the final LayerNorm here to not mess with the
+                # representations. The `last_hidden_states` that we typically use for
+                # obtaining the final prompt representations passes through the LayerNorm
+                # layer.
+                prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+        if self.text_encoder is not None:
+            prompt_embeds_dtype = self.text_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            # textual inversion: procecss multi-vector tokens if necessary
+            if isinstance(self, TextualInversionLoaderMixin):
+                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            negative_prompt_embeds = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            negative_prompt_embeds = negative_prompt_embeds[0]
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+            # Retrieve the original scale by scaling back the LoRA layers
+            unscale_lora_layers(self.text_encoder, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+    def run_safety_checker(self, image, device, dtype):
+        if self.safety_checker is None:
+            has_nsfw_concept = None
+        else:
+            if torch.is_tensor(image):
+                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+            else:
+                feature_extractor_input = self.image_processor.numpy_to_pil(image)
+            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+            )
+        return image, has_nsfw_concept
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+    def decode_latents(self, latents):
+        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+        latents = 1 / self.vae.config.scaling_factor * latents
+        image = self.vae.decode(latents, return_dict=False)[0]
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        image,
+        callback_steps,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        controlnet_conditioning_scale=1.0,
+        control_guidance_start=0.0,
+        control_guidance_end=1.0,
+    ):
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        # Check `image`
+        is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+            self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+        )
+        if (
+            isinstance(self.controlnet, ControlNetXSModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, ControlNetXSModel)
+        ):
+            self.check_image(image, prompt, prompt_embeds)
+        else:
+            assert False
+
+        # Check `controlnet_conditioning_scale`
+        if (
+            isinstance(self.controlnet, ControlNetXSModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, ControlNetXSModel)
+        ):
+            if not isinstance(controlnet_conditioning_scale, float):
+                raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+        else:
+            assert False
+
+        start, end = control_guidance_start, control_guidance_end
+        if start >= end:
+            raise ValueError(
+                f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+            )
+        if start < 0.0:
+            raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+        if end > 1.0:
+            raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+    def check_image(self, image, prompt, prompt_embeds):
+        image_is_pil = isinstance(image, PIL.Image.Image)
+        image_is_tensor = isinstance(image, torch.Tensor)
+        image_is_np = isinstance(image, np.ndarray)
+        image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+        image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+        image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+        if (
+            not image_is_pil
+            and not image_is_tensor
+            and not image_is_np
+            and not image_is_pil_list
+            and not image_is_tensor_list
+            and not image_is_np_list
+        ):
+            raise TypeError(
+                f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+            )
+
+        if image_is_pil:
+            image_batch_size = 1
+        else:
+            image_batch_size = len(image)
+
+        if prompt is not None and isinstance(prompt, str):
+            prompt_batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            prompt_batch_size = len(prompt)
+        elif prompt_embeds is not None:
+            prompt_batch_size = prompt_embeds.shape[0]
+
+        if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+            raise ValueError(
+                f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+            )
+
+    def prepare_image(
+        self,
+        image,
+        width,
+        height,
+        batch_size,
+        num_images_per_prompt,
+        device,
+        dtype,
+        do_classifier_free_guidance=False,
+    ):
+        image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+        image_batch_size = image.shape[0]
+
+        if image_batch_size == 1:
+            repeat_by = batch_size
+        else:
+            # image batch size is the same as prompt batch size
+            repeat_by = num_images_per_prompt
+
+        image = image.repeat_interleave(repeat_by, dim=0)
+
+        image = image.to(device=device, dtype=dtype)
+
+        if do_classifier_free_guidance:
+            image = torch.cat([image] * 2)
+
+        return image
+
+    # 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."
+            )
+
+        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
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        image: PipelineImageInput = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[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,
+        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,
+        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+        control_guidance_start: float = 0.0,
+        control_guidance_end: float = 1.0,
+        clip_skip: Optional[int] = None,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,
+                    `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+                specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
+                accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+                and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
+                `init`, images must be passed as a list such that each element of the list can be correctly batched for
+                input to a single ControlNet.
+            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.
+            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.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that calls every `callback_steps` steps during inference. The function is called with the
+                following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function is called. If not specified, the callback is called at
+                every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+                to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+                the corresponding scale as a list.
+            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+                The percentage of total steps at which the ControlNet starts applying.
+            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The percentage of total steps at which the ControlNet stops applying.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list with the generated images and the
+                second element is a list of `bool`s indicating whether the corresponding generated image contains
+                "not-safe-for-work" (nsfw) content.
+        """
+        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            image,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            controlnet_conditioning_scale,
+            control_guidance_start,
+            control_guidance_end,
+        )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+            clip_skip=clip_skip,
+        )
+        # For classifier free guidance, we need to do two forward passes.
+        # Here we concatenate the unconditional and text embeddings into a single batch
+        # to avoid doing two forward passes
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        # 4. Prepare image
+        if isinstance(controlnet, ControlNetXSModel):
+            image = self.prepare_image(
+                image=image,
+                width=width,
+                height=height,
+                batch_size=batch_size * num_images_per_prompt,
+                num_images_per_prompt=num_images_per_prompt,
+                device=device,
+                dtype=controlnet.dtype,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+            )
+            height, width = image.shape[-2:]
+        else:
+            assert False
+
+        # 5. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 6. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 7. 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)
+
+        # 8. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        is_unet_compiled = is_compiled_module(self.unet)
+        is_controlnet_compiled = is_compiled_module(self.controlnet)
+        is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # Relevant thread:
+                # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+                if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:
+                    torch._inductor.cudagraph_mark_step_begin()
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                dont_control = (
+                    i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end
+                )
+                if dont_control:
+                    noise_pred = self.unet(
+                        sample=latent_model_input,
+                        timestep=t,
+                        encoder_hidden_states=prompt_embeds,
+                        cross_attention_kwargs=cross_attention_kwargs,
+                        return_dict=True,
+                    ).sample
+                else:
+                    noise_pred = self.controlnet(
+                        base_model=self.unet,
+                        sample=latent_model_input,
+                        timestep=t,
+                        encoder_hidden_states=prompt_embeds,
+                        controlnet_cond=image,
+                        conditioning_scale=controlnet_conditioning_scale,
+                        cross_attention_kwargs=cross_attention_kwargs,
+                        return_dict=True,
+                    ).sample
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        # If we do sequential model offloading, let's offload unet and controlnet
+        # manually for max memory savings
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.unet.to("cpu")
+            self.controlnet.to("cpu")
+            torch.cuda.empty_cache()
+
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+                0
+            ]
+            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+        else:
+            image = latents
+            has_nsfw_concept = None
+
+        if has_nsfw_concept is None:
+            do_denormalize = [True] * image.shape[0]
+        else:
+            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

src/diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_img2img.py

@@ -20,11 +20,11 @@ from typing import Any, Callable, Dict, List, Optional, Union
 
 import PIL.Image
 import torch
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import LCMScheduler
 from ...utils import (
@@ -129,7 +129,7 @@ EXAMPLE_DOC_STRING = """
 
 
 class LatentConsistencyModelImg2ImgPipeline(
-    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for image-to-image generation using a latent consistency model.
@@ -142,6 +142,7 @@ class LatentConsistencyModelImg2ImgPipeline(
         - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
         - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
         - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
     Args:
         vae ([`AutoencoderKL`]):
@@ -166,7 +167,7 @@ class LatentConsistencyModelImg2ImgPipeline(
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
     _callback_tensor_inputs = ["latents", "denoised", "prompt_embeds", "w_embedding"]
 
@@ -179,6 +180,7 @@ class LatentConsistencyModelImg2ImgPipeline(
         scheduler: LCMScheduler,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -191,6 +193,7 @@ class LatentConsistencyModelImg2ImgPipeline(
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
 
         if safety_checker is None and requires_safety_checker:
@@ -449,6 +452,31 @@ class LatentConsistencyModelImg2ImgPipeline(
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:
@@ -647,6 +675,7 @@ class LatentConsistencyModelImg2ImgPipeline(
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
@@ -695,6 +724,8 @@ class LatentConsistencyModelImg2ImgPipeline(
             prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
                 provided, text embeddings are generated from the `prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -758,6 +789,12 @@ class LatentConsistencyModelImg2ImgPipeline(
         device = self._execution_device
         # do_classifier_free_guidance = guidance_scale > 1.0
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+
         # 3. Encode input prompt
         lora_scale = (
             self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
@@ -815,6 +852,9 @@ class LatentConsistencyModelImg2ImgPipeline(
         # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, None)
 
+        # 7.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
         # 8. LCM Multistep Sampling Loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         self._num_timesteps = len(timesteps)
@@ -829,6 +869,7 @@ class LatentConsistencyModelImg2ImgPipeline(
                     timestep_cond=w_embedding,
                     encoder_hidden_states=prompt_embeds,
                     cross_attention_kwargs=self.cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
                     return_dict=False,
                 )[0]
 

src/diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_text2img.py

@@ -19,11 +19,11 @@ import inspect
 from typing import Any, Callable, Dict, List, Optional, Union
 
 import torch
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
-from ...image_processor import VaeImageProcessor
-from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import LCMScheduler
 from ...utils import (
@@ -107,7 +107,7 @@ def retrieve_timesteps(
 
 
 class LatentConsistencyModelPipeline(
-    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for text-to-image generation using a latent consistency model.
@@ -120,6 +120,7 @@ class LatentConsistencyModelPipeline(
         - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
         - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
         - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
     Args:
         vae ([`AutoencoderKL`]):
@@ -144,7 +145,7 @@ class LatentConsistencyModelPipeline(
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
     _callback_tensor_inputs = ["latents", "denoised", "prompt_embeds", "w_embedding"]
 
@@ -157,6 +158,7 @@ class LatentConsistencyModelPipeline(
         scheduler: LCMScheduler,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -185,6 +187,7 @@ class LatentConsistencyModelPipeline(
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
@@ -433,6 +436,31 @@ class LatentConsistencyModelPipeline(
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:
@@ -581,6 +609,7 @@ class LatentConsistencyModelPipeline(
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
@@ -629,6 +658,8 @@ class LatentConsistencyModelPipeline(
             prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
                 provided, text embeddings are generated from the `prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -697,6 +728,12 @@ class LatentConsistencyModelPipeline(
         device = self._execution_device
         # do_classifier_free_guidance = guidance_scale > 1.0
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+
         # 3. Encode input prompt
         lora_scale = (
             self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
@@ -748,6 +785,9 @@ class LatentConsistencyModelPipeline(
         # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, None)
 
+        # 7.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
         # 8. LCM MultiStep Sampling Loop:
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         self._num_timesteps = len(timesteps)
@@ -762,6 +802,7 @@ class LatentConsistencyModelPipeline(
                     timestep_cond=w_embedding,
                     encoder_hidden_states=prompt_embeds,
                     cross_attention_kwargs=self.cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
                     return_dict=False,
                 )[0]
 

src/diffusers/pipelines/onnx_utils.py

@@ -22,6 +22,7 @@ from typing import Optional, Union
 
 import numpy as np
 from huggingface_hub import hf_hub_download
+from huggingface_hub.utils import validate_hf_hub_args
 
 from ..utils import ONNX_EXTERNAL_WEIGHTS_NAME, ONNX_WEIGHTS_NAME, is_onnx_available, logging
 
@@ -130,10 +131,11 @@ class OnnxRuntimeModel:
         self._save_pretrained(save_directory, **kwargs)
 
     @classmethod
+    @validate_hf_hub_args
     def _from_pretrained(
         cls,
         model_id: Union[str, Path],
-        use_auth_token: Optional[Union[bool, str, None]] = None,
+        token: Optional[Union[bool, str, None]] = None,
         revision: Optional[Union[str, None]] = None,
         force_download: bool = False,
         cache_dir: Optional[str] = None,
@@ -148,7 +150,7 @@ class OnnxRuntimeModel:
         Arguments:
             model_id (`str` or `Path`):
                 Directory from which to load
-            use_auth_token (`str` or `bool`):
+            token (`str` or `bool`):
                 Is needed to load models from a private or gated repository
             revision (`str`):
                 Revision is the specific model version to use. It can be a branch name, a tag name, or a commit id
@@ -179,7 +181,7 @@ class OnnxRuntimeModel:
             model_cache_path = hf_hub_download(
                 repo_id=model_id,
                 filename=model_file_name,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 cache_dir=cache_dir,
                 force_download=force_download,
@@ -190,11 +192,12 @@ class OnnxRuntimeModel:
         return cls(model=model, **kwargs)
 
     @classmethod
+    @validate_hf_hub_args
     def from_pretrained(
         cls,
         model_id: Union[str, Path],
         force_download: bool = True,
-        use_auth_token: Optional[str] = None,
+        token: Optional[str] = None,
         cache_dir: Optional[str] = None,
         **model_kwargs,
     ):
@@ -207,6 +210,6 @@ class OnnxRuntimeModel:
             revision=revision,
             cache_dir=cache_dir,
             force_download=force_download,
-            use_auth_token=use_auth_token,
+            token=token,
             **model_kwargs,
         )

src/diffusers/pipelines/pipeline_flax_utils.py

@@ -24,6 +24,7 @@ import numpy as np
 import PIL.Image
 from flax.core.frozen_dict import FrozenDict
 from huggingface_hub import create_repo, snapshot_download
+from huggingface_hub.utils import validate_hf_hub_args
 from PIL import Image
 from tqdm.auto import tqdm
 
@@ -32,7 +33,6 @@ from ..models.modeling_flax_utils import FLAX_WEIGHTS_NAME, FlaxModelMixin
 from ..schedulers.scheduling_utils_flax import SCHEDULER_CONFIG_NAME, FlaxSchedulerMixin
 from ..utils import (
     CONFIG_NAME,
-    DIFFUSERS_CACHE,
     BaseOutput,
     PushToHubMixin,
     http_user_agent,
@@ -227,6 +227,7 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
                 )
 
     @classmethod
+    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
         r"""
         Instantiate a Flax-based diffusion pipeline from pretrained pipeline weights.
@@ -264,7 +265,7 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -314,11 +315,11 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
         >>> dpm_params["scheduler"] = dpmpp_state
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         from_pt = kwargs.pop("from_pt", False)
         use_memory_efficient_attention = kwargs.pop("use_memory_efficient_attention", False)
@@ -334,7 +335,7 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
             )
             # make sure we only download sub-folders and `diffusers` filenames
@@ -365,7 +366,7 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 allow_patterns=allow_patterns,
                 ignore_patterns=ignore_patterns,

src/diffusers/pipelines/pipeline_utils.py

@@ -28,7 +28,14 @@ from typing import Any, Callable, Dict, List, Optional, Union
 import numpy as np
 import PIL.Image
 import torch
-from huggingface_hub import ModelCard, create_repo, hf_hub_download, model_info, snapshot_download
+from huggingface_hub import (
+    ModelCard,
+    create_repo,
+    hf_hub_download,
+    model_info,
+    snapshot_download,
+)
+from huggingface_hub.utils import validate_hf_hub_args
 from packaging import version
 from requests.exceptions import HTTPError
 from tqdm.auto import tqdm
@@ -40,8 +47,6 @@ from ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
 from ..utils import (
     CONFIG_NAME,
     DEPRECATED_REVISION_ARGS,
-    DIFFUSERS_CACHE,
-    HF_HUB_OFFLINE,
     SAFETENSORS_WEIGHTS_NAME,
     WEIGHTS_NAME,
     BaseOutput,
@@ -249,10 +254,11 @@ def variant_compatible_siblings(filenames, variant=None) -> Union[List[os.PathLi
     return usable_filenames, variant_filenames
 
 
-def warn_deprecated_model_variant(pretrained_model_name_or_path, use_auth_token, variant, revision, model_filenames):
+@validate_hf_hub_args
+def warn_deprecated_model_variant(pretrained_model_name_or_path, token, variant, revision, model_filenames):
     info = model_info(
         pretrained_model_name_or_path,
-        use_auth_token=use_auth_token,
+        token=token,
         revision=None,
     )
     filenames = {sibling.rfilename for sibling in info.siblings}
@@ -375,7 +381,6 @@ def _get_pipeline_class(
             custom_pipeline,
             module_file=file_name,
             class_name=class_name,
-            repo_id=repo_id,
             cache_dir=cache_dir,
             revision=revision,
         )
@@ -909,6 +914,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         return torch.float32
 
     @classmethod
+    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
         r"""
         Instantiate a PyTorch diffusion pipeline from pretrained pipeline weights.
@@ -976,7 +982,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -1056,12 +1062,12 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         >>> pipeline.scheduler = scheduler
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         from_flax = kwargs.pop("from_flax", False)
         torch_dtype = kwargs.pop("torch_dtype", None)
@@ -1094,7 +1100,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 force_download=force_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 from_flax=from_flax,
                 use_safetensors=use_safetensors,
@@ -1299,7 +1305,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 "force_download": force_download,
                 "proxies": proxies,
                 "local_files_only": local_files_only,
-                "use_auth_token": use_auth_token,
+                "token": token,
                 "revision": revision,
                 "torch_dtype": torch_dtype,
                 "custom_pipeline": custom_pipeline,
@@ -1529,6 +1535,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 cpu_offload(model, device, offload_buffers=offload_buffers)
 
     @classmethod
+    @validate_hf_hub_args
     def download(cls, pretrained_model_name, **kwargs) -> Union[str, os.PathLike]:
         r"""
         Download and cache a PyTorch diffusion pipeline from pretrained pipeline weights.
@@ -1576,7 +1583,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -1619,12 +1626,12 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         </Tip>
 
         """
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         from_flax = kwargs.pop("from_flax", False)
         custom_pipeline = kwargs.pop("custom_pipeline", None)
@@ -1646,11 +1653,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         model_info_call_error: Optional[Exception] = None
         if not local_files_only:
             try:
-                info = model_info(
-                    pretrained_model_name,
-                    use_auth_token=use_auth_token,
-                    revision=revision,
-                )
+                info = model_info(pretrained_model_name, token=token, revision=revision)
             except HTTPError as e:
                 logger.warn(f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache.")
                 local_files_only = True
@@ -1665,7 +1668,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 proxies=proxies,
                 force_download=force_download,
                 resume_download=resume_download,
-                use_auth_token=use_auth_token,
+                token=token,
             )
 
             config_dict = cls._dict_from_json_file(config_file)
@@ -1715,9 +1718,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             if revision in DEPRECATED_REVISION_ARGS and version.parse(
                 version.parse(__version__).base_version
             ) >= version.parse("0.22.0"):
-                warn_deprecated_model_variant(
-                    pretrained_model_name, use_auth_token, variant, revision, model_filenames
-                )
+                warn_deprecated_model_variant(pretrained_model_name, token, variant, revision, model_filenames)
 
             model_folder_names = {os.path.split(f)[0] for f in model_filenames if os.path.split(f)[0] in folder_names}
 
@@ -1859,7 +1860,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 allow_patterns=allow_patterns,
                 ignore_patterns=ignore_patterns,
@@ -1883,7 +1884,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                         "force_download": force_download,
                         "proxies": proxies,
                         "local_files_only": local_files_only,
-                        "use_auth_token": use_auth_token,
+                        "token": token,
                         "variant": variant,
                         "use_safetensors": use_safetensors,
                     }

src/diffusers/pipelines/spectrogram_diffusion/pipeline_spectrogram_diffusion.py

@@ -29,7 +29,7 @@ if is_onnx_available():
     from ..onnx_utils import OnnxRuntimeModel
 
 from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline
-from .continous_encoder import SpectrogramContEncoder
+from .continuous_encoder import SpectrogramContEncoder
 from .notes_encoder import SpectrogramNotesEncoder
 
 

src/diffusers/pipelines/stable_diffusion/README.md

@@ -6,13 +6,13 @@ Stable Diffusion was proposed in [Stable Diffusion Announcement](https://stabili
 
 The summary of the model is the following:
 
-*Stable Diffusion is a text-to-image model that will empower billions of people to create stunning art within seconds. It is a breakthrough in speed and quality meaning that it can run on consumer GPUs. You can see some of the amazing output that has been created by this model without pre or post-processing on this page. The model itself builds upon the work of the team at CompVis and Runway in their widely used latent diffusion model combined with insights from the conditional diffusion models by our lead generative AI developer Katherine Crowson, Dall-E 2 by Open AI, Imagen by Google Brain and many others. We are delighted that AI media generation is a cooperative field and hope it can continue this way to bring the gift of creativity to all.* 
+*Stable Diffusion is a text-to-image model that will empower billions of people to create stunning art within seconds. It is a breakthrough in speed and quality meaning that it can run on consumer GPUs. You can see some of the amazing output that has been created by this model without pre or post-processing on this page. The model itself builds upon the work of the team at CompVis and Runway in their widely used latent diffusion model combined with insights from the conditional diffusion models by our lead generative AI developer Katherine Crowson, Dall-E 2 by Open AI, Imagen by Google Brain and many others. We are delighted that AI media generation is a cooperative field and hope it can continue this way to bring the gift of creativity to all.*
 
 ## Tips:
 
 - Stable Diffusion has the same architecture as [Latent Diffusion](https://arxiv.org/abs/2112.10752) but uses a frozen CLIP Text Encoder instead of training the text encoder jointly with the diffusion model.
 - An in-detail explanation of the Stable Diffusion model can be found under [Stable Diffusion with 🧨 Diffusers](https://huggingface.co/blog/stable_diffusion).
-- If you don't want to rely on the Hugging Face Hub and having to pass a authentication token, you can 
+- If you don't want to rely on the Hugging Face Hub and having to pass a authentication token, you can
 download the weights with `git lfs install; git clone https://huggingface.co/runwayml/stable-diffusion-v1-5` and instead pass the local path to the cloned folder to `from_pretrained` as shown below.
 - Stable Diffusion can work with a variety of different samplers as is shown below.
 
@@ -28,7 +28,7 @@ download the weights with `git lfs install; git clone https://huggingface.co/run
 
 ### Using Stable Diffusion without being logged into the Hub.
 
-If you want to download the model weights using a single Python line, you need to be logged in via `huggingface-cli login`. 
+If you want to download the model weights using a single Python line, you need to be logged in via `huggingface-cli login`.
 
 ```python
 from diffusers import DiffusionPipeline
@@ -61,8 +61,8 @@ pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
 pipe = pipe.to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
-image = pipe(prompt).images[0]  
-    
+image = pipe(prompt).images[0]
+
 image.save("astronaut_rides_horse.png")
 ```
 
@@ -75,13 +75,13 @@ from diffusers import StableDiffusionPipeline, DDIMScheduler
 scheduler =  DDIMScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler")
 
 pipe = StableDiffusionPipeline.from_pretrained(
-    "runwayml/stable-diffusion-v1-5", 
+    "runwayml/stable-diffusion-v1-5",
     scheduler=scheduler,
 ).to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
-image = pipe(prompt).images[0]  
-    
+image = pipe(prompt).images[0]
+
 image.save("astronaut_rides_horse.png")
 ```
 
@@ -94,13 +94,13 @@ from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler
 lms = LMSDiscreteScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler")
 
 pipe = StableDiffusionPipeline.from_pretrained(
-    "runwayml/stable-diffusion-v1-5", 
+    "runwayml/stable-diffusion-v1-5",
     scheduler=lms,
 ).to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
-image = pipe(prompt).images[0]  
-    
+image = pipe(prompt).images[0]
+
 image.save("astronaut_rides_horse.png")
 ```
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py

@@ -18,11 +18,11 @@ from typing import Callable, Dict, List, Optional, Union
 import numpy as np
 import PIL.Image
 import torch
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import PIL_INTERPOLATION, deprecate, logging
 from ...utils.torch_utils import randn_tensor
@@ -72,7 +72,9 @@ def retrieve_latents(
         raise AttributeError("Could not access latents of provided encoder_output")
 
 
-class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+class StableDiffusionInstructPix2PixPipeline(
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin
+):
     r"""
     Pipeline for pixel-level image editing by following text instructions (based on Stable Diffusion).
 
@@ -83,6 +85,7 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
         - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
         - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
         - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
     Args:
         vae ([`AutoencoderKL`]):
@@ -105,7 +108,7 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
     _callback_tensor_inputs = ["latents", "prompt_embeds", "image_latents"]
 
@@ -118,6 +121,7 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -146,6 +150,7 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
@@ -166,6 +171,7 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
@@ -213,6 +219,8 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
             negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -293,6 +301,16 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
         self._guidance_scale = guidance_scale
         self._image_guidance_scale = image_guidance_scale
 
+        device = self._execution_device
+
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if self.do_classifier_free_guidance:
+                image_embeds = torch.cat([image_embeds, negative_image_embeds, negative_image_embeds])
+
         if image is None:
             raise ValueError("`image` input cannot be undefined.")
 
@@ -367,6 +385,9 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
         # 8. 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)
 
+        # 8.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
         # 9. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         self._num_timesteps = len(timesteps)
@@ -383,7 +404,11 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
 
                 # predict the noise residual
                 noise_pred = self.unet(
-                    scaled_latent_model_input, t, encoder_hidden_states=prompt_embeds, return_dict=False
+                    scaled_latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
                 )[0]
 
                 # Hack:
@@ -598,11 +623,36 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline, TextualInversion
             # For classifier free guidance, we need to do two forward passes.
             # Here we concatenate the unconditional and text embeddings into a single batch
             # to avoid doing two forward passes
-            # pix2pix has two  negative embeddings, and unlike in other pipelines latents are ordered [prompt_embeds, negative_prompt_embeds, negative_prompt_embeds]
+            # pix2pix has two negative embeddings, and unlike in other pipelines latents are ordered [prompt_embeds, negative_prompt_embeds, negative_prompt_embeds]
             prompt_embeds = torch.cat([prompt_embeds, negative_prompt_embeds, negative_prompt_embeds])
 
         return prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_ldm3d.py

@@ -19,11 +19,11 @@ from typing import Any, Callable, Dict, List, Optional, Union
 import numpy as np
 import PIL.Image
 import torch
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
-from ...image_processor import VaeImageProcessorLDM3D
-from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...image_processor import PipelineImageInput, VaeImageProcessorLDM3D
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -82,7 +82,7 @@ class LDM3DPipelineOutput(BaseOutput):
 
 
 class StableDiffusionLDM3DPipeline(
-    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for text-to-image and 3D generation using LDM3D.
@@ -95,6 +95,7 @@ class StableDiffusionLDM3DPipeline(
         - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
         - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
         - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
     Args:
         vae ([`AutoencoderKL`]):
@@ -117,7 +118,7 @@ class StableDiffusionLDM3DPipeline(
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
 
     def __init__(
@@ -129,6 +130,7 @@ class StableDiffusionLDM3DPipeline(
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection],
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -157,6 +159,7 @@ class StableDiffusionLDM3DPipeline(
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessorLDM3D(vae_scale_factor=self.vae_scale_factor)
@@ -410,6 +413,31 @@ class StableDiffusionLDM3DPipeline(
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:
             has_nsfw_concept = None
@@ -529,6 +557,7 @@ class StableDiffusionLDM3DPipeline(
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -573,6 +602,8 @@ class StableDiffusionLDM3DPipeline(
             negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -622,6 +653,14 @@ class StableDiffusionLDM3DPipeline(
         # corresponds to doing no classifier free guidance.
         do_classifier_free_guidance = guidance_scale > 1.0
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 3. Encode input prompt
         prompt_embeds, negative_prompt_embeds = self.encode_prompt(
             prompt,
@@ -659,6 +698,9 @@ class StableDiffusionLDM3DPipeline(
         # 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)
 
+        # 6.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
         # 7. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         with self.progress_bar(total=num_inference_steps) as progress_bar:
@@ -673,6 +715,7 @@ class StableDiffusionLDM3DPipeline(
                     t,
                     encoder_hidden_states=prompt_embeds,
                     cross_attention_kwargs=cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
                     return_dict=False,
                 )[0]
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_panorama.py

@@ -16,11 +16,11 @@ import inspect
 from typing import Any, Callable, Dict, List, Optional, Union
 
 import torch
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
-from ...image_processor import VaeImageProcessor
-from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import DDIMScheduler
 from ...utils import (
@@ -59,13 +59,19 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin):
     r"""
     Pipeline for text-to-image generation using MultiDiffusion.
 
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
@@ -87,7 +93,7 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
 
     def __init__(
@@ -99,6 +105,7 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
         scheduler: DDIMScheduler,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -127,6 +134,7 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
@@ -363,6 +371,31 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:
@@ -529,6 +562,7 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -578,6 +612,8 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
             negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -632,6 +668,14 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
         # corresponds to doing no classifier free guidance.
         do_classifier_free_guidance = guidance_scale > 1.0
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 3. Encode input prompt
         text_encoder_lora_scale = (
             cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
@@ -681,6 +725,9 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
         # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 7.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
         # 8. Denoising loop
         # Each denoising step also includes refinement of the latents with respect to the
         # views.
@@ -743,6 +790,7 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
                         t,
                         encoder_hidden_states=prompt_embeds_input,
                         cross_attention_kwargs=cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
                     ).sample
 
                     # perform guidance

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_sag.py

@@ -17,11 +17,11 @@ from typing import Any, Callable, Dict, List, Optional, Union
 
 import torch
 import torch.nn.functional as F
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
-from ...image_processor import VaeImageProcessor
-from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -98,13 +98,17 @@ class CrossAttnStoreProcessor:
 
 
 # Modified to get self-attention guidance scale in this paper (https://arxiv.org/pdf/2210.00939.pdf) as an input
-class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin):
+class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
@@ -126,7 +130,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
 
     def __init__(
@@ -138,6 +142,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -150,6 +155,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
@@ -386,6 +392,31 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:
@@ -519,6 +550,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -565,6 +597,8 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
             negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -618,6 +652,14 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         # `sag_scale = 0` means no self-attention guidance
         do_self_attention_guidance = sag_scale > 0.0
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 3. Encode input prompt
         prompt_embeds, negative_prompt_embeds = self.encode_prompt(
             prompt,
@@ -655,6 +697,10 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         # 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)
 
+        # 6.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+        added_uncond_kwargs = {"image_embeds": negative_image_embeds} if ip_adapter_image is not None else None
+
         # 7. Denoising loop
         store_processor = CrossAttnStoreProcessor()
         self.unet.mid_block.attentions[0].transformer_blocks[0].attn1.processor = store_processor
@@ -680,6 +726,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
                         t,
                         encoder_hidden_states=prompt_embeds,
                         cross_attention_kwargs=cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
                     ).sample
 
                     # perform guidance
@@ -703,7 +750,12 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
                             )
                             uncond_emb, _ = prompt_embeds.chunk(2)
                             # forward and give guidance
-                            degraded_pred = self.unet(degraded_latents, t, encoder_hidden_states=uncond_emb).sample
+                            degraded_pred = self.unet(
+                                degraded_latents,
+                                t,
+                                encoder_hidden_states=uncond_emb,
+                                added_cond_kwargs=added_uncond_kwargs,
+                            ).sample
                             noise_pred += sag_scale * (noise_pred_uncond - degraded_pred)
                         else:
                             # DDIM-like prediction of x0
@@ -715,7 +767,12 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
                                 pred_x0, cond_attn, map_size, t, self.pred_epsilon(latents, noise_pred, t)
                             )
                             # forward and give guidance
-                            degraded_pred = self.unet(degraded_latents, t, encoder_hidden_states=prompt_embeds).sample
+                            degraded_pred = self.unet(
+                                degraded_latents,
+                                t,
+                                encoder_hidden_states=prompt_embeds,
+                                added_cond_kwargs=added_cond_kwargs,
+                            ).sample
                             noise_pred += sag_scale * (noise_pred - degraded_pred)
 
                     # compute the previous noisy sample x_t -> x_t-1

src/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py

@@ -5,10 +5,12 @@ from typing import Callable, List, Optional, Union
 import numpy as np
 import torch
 from packaging import version
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from ...configuration_utils import FrozenDict
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...image_processor import PipelineImageInput
+from ...loaders import IPAdapterMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import deprecate, logging
 from ...utils.torch_utils import randn_tensor
@@ -20,13 +22,16 @@ from .safety_checker import SafeStableDiffusionSafetyChecker
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class StableDiffusionPipelineSafe(DiffusionPipeline):
+class StableDiffusionPipelineSafe(DiffusionPipeline, IPAdapterMixin):
     r"""
     Pipeline based on the [`StableDiffusionPipeline`] for text-to-image generation using Safe Latent Diffusion.
 
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
@@ -48,7 +53,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
 
     def __init__(
         self,
@@ -59,6 +64,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: SafeStableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -140,6 +146,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self._safety_text_concept = safety_concept
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
@@ -467,6 +474,31 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
                 noise_guidance = noise_guidance - noise_guidance_safety
         return noise_guidance, safety_momentum
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     @torch.no_grad()
     def __call__(
         self,
@@ -480,6 +512,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         eta: float = 0.0,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -521,6 +554,8 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
                 Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
                 generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                 tensor is generated by sampling using the supplied random `generator`.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -588,6 +623,17 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         if not enable_safety_guidance:
             warnings.warn("Safety checker disabled!")
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if do_classifier_free_guidance:
+                if enable_safety_guidance:
+                    image_embeds = torch.cat([negative_image_embeds, image_embeds, image_embeds])
+                else:
+                    image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 3. Encode input prompt
         prompt_embeds = self._encode_prompt(
             prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt, enable_safety_guidance
@@ -613,6 +659,9 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         # 6. Prepare extra step kwargs.
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 6.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
         safety_momentum = None
 
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
@@ -627,7 +676,9 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
                 # predict the noise residual
-                noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=prompt_embeds).sample
+                noise_pred = self.unet(
+                    latent_model_input, t, encoder_hidden_states=prompt_embeds, added_cond_kwargs=added_cond_kwargs
+                ).sample
 
                 # perform guidance
                 if do_classifier_free_guidance:

src/diffusers/schedulers/scheduling_deis_multistep.py

@@ -734,7 +734,16 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
             schedule_timesteps = self.timesteps.to(original_samples.device)
             timesteps = timesteps.to(original_samples.device)
 
-        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+        step_indices = []
+        for timestep in timesteps:
+            index_candidates = (schedule_timesteps == timestep).nonzero()
+            if len(index_candidates) == 0:
+                step_index = len(schedule_timesteps) - 1
+            elif len(index_candidates) > 1:
+                step_index = index_candidates[1].item()
+            else:
+                step_index = index_candidates[0].item()
+            step_indices.append(step_index)
 
         sigma = sigmas[step_indices].flatten()
         while len(sigma.shape) < len(original_samples.shape):

src/diffusers/schedulers/scheduling_dpmsolver_multistep.py

@@ -896,7 +896,16 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             schedule_timesteps = self.timesteps.to(original_samples.device)
             timesteps = timesteps.to(original_samples.device)
 
-        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+        step_indices = []
+        for timestep in timesteps:
+            index_candidates = (schedule_timesteps == timestep).nonzero()
+            if len(index_candidates) == 0:
+                step_index = len(schedule_timesteps) - 1
+            elif len(index_candidates) > 1:
+                step_index = index_candidates[1].item()
+            else:
+                step_index = index_candidates[0].item()
+            step_indices.append(step_index)
 
         sigma = sigmas[step_indices].flatten()
         while len(sigma.shape) < len(original_samples.shape):