[core] Move community AnimateDiff ControlNet to core (#8972)

* add animatediff controlnet to core

* make style; remove unused method

* fix copied from comment

* add tests

* changes to make tests work

* add utility function to load videos

* update docs

* update pipeline example

* make style

* update docs with example

* address review comments

* add latest freeinit test from #8969

* LoraLoaderMixin -> StableDiffusionLoraLoaderMixin

* fix docs

* Update src/diffusers/utils/loading_utils.py
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

* fix: variable out of scope

---------
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

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

@@ -25,6 +25,9 @@ The abstract of the paper is the following:
 | Pipeline | Tasks | Demo
 |---|---|:---:|
 | [AnimateDiffPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/animatediff/pipeline_animatediff.py) | *Text-to-Video Generation with AnimateDiff* |
+| [AnimateDiffControlNetPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/animatediff/pipeline_animatediff_controlnet.py) | *Controlled Video-to-Video Generation with AnimateDiff using ControlNet* |
+| [AnimateDiffSparseControlNetPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/animatediff/pipeline_animatediff_sparsectrl.py) | *Controlled Video-to-Video Generation with AnimateDiff using SparseCtrl* |
+| [AnimateDiffSDXLPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/animatediff/pipeline_animatediff_sdxl.py) | *Video-to-Video Generation with AnimateDiff* |
 | [AnimateDiffVideoToVideoPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/animatediff/pipeline_animatediff_video2video.py) | *Video-to-Video Generation with AnimateDiff* |
 
 ## Available checkpoints
@@ -100,6 +103,83 @@ AnimateDiff tends to work better with finetuned Stable Diffusion models. If you
 
 </Tip>
 
+### AnimateDiffControlNetPipeline
+
+AnimateDiff can also be used with ControlNets ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. With a ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide depth maps, the ControlNet model generates a video that'll preserve the spatial information from the depth maps. It is a more flexible and accurate way to control the video generation process.
+
+```python
+import torch
+from diffusers import AnimateDiffControlNetPipeline, AutoencoderKL, ControlNetModel, MotionAdapter, LCMScheduler
+from diffusers.utils import export_to_gif, load_video
+
+# Additionally, you will need a preprocess videos before they can be used with the ControlNet
+# HF maintains just the right package for it: `pip install controlnet_aux`
+from controlnet_aux.processor import ZoeDetector
+
+# Download controlnets from https://huggingface.co/lllyasviel/ControlNet-v1-1 to use .from_single_file
+# Download Diffusers-format controlnets, such as https://huggingface.co/lllyasviel/sd-controlnet-depth, to use .from_pretrained()
+controlnet = ControlNetModel.from_single_file("control_v11f1p_sd15_depth.pth", torch_dtype=torch.float16)
+
+# We use AnimateLCM for this example but one can use the original motion adapters as well (for example, https://huggingface.co/guoyww/animatediff-motion-adapter-v1-5-3)
+motion_adapter = MotionAdapter.from_pretrained("wangfuyun/AnimateLCM")
+
+vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16)
+pipe: AnimateDiffControlNetPipeline = AnimateDiffControlNetPipeline.from_pretrained(
+    "SG161222/Realistic_Vision_V5.1_noVAE",
+    motion_adapter=motion_adapter,
+    controlnet=controlnet,
+    vae=vae,
+).to(device="cuda", dtype=torch.float16)
+pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config, beta_schedule="linear")
+pipe.load_lora_weights("wangfuyun/AnimateLCM", weight_name="AnimateLCM_sd15_t2v_lora.safetensors", adapter_name="lcm-lora")
+pipe.set_adapters(["lcm-lora"], [0.8])
+
+depth_detector = ZoeDetector.from_pretrained("lllyasviel/Annotators").to("cuda")
+video = load_video("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-vid2vid-input-1.gif")
+conditioning_frames = []
+
+with pipe.progress_bar(total=len(video)) as progress_bar:
+    for frame in video:
+        conditioning_frames.append(depth_detector(frame))
+        progress_bar.update()
+
+prompt = "a panda, playing a guitar, sitting in a pink boat, in the ocean, mountains in background, realistic, high quality"
+negative_prompt = "bad quality, worst quality"
+
+video = pipe(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    num_frames=len(video),
+    num_inference_steps=10,
+    guidance_scale=2.0,
+    conditioning_frames=conditioning_frames,
+    generator=torch.Generator().manual_seed(42),
+).frames[0]
+
+export_to_gif(video, "animatediff_controlnet.gif", fps=8)
+```
+
+Here are some sample outputs:
+
+<table align="center">
+    <tr>
+      <th align="center">Source Video</th>
+      <th align="center">Output Video</th>
+    </tr>
+    <tr>
+        <td align="center">
+          raccoon playing a guitar
+          <br />
+          <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-vid2vid-input-1.gif" alt="racoon playing a guitar" />
+        </td>
+        <td align="center">
+          a panda, playing a guitar, sitting in a pink boat, in the ocean, mountains in background, realistic, high quality
+          <br/>
+          <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-controlnet-output.gif" alt="a panda, playing a guitar, sitting in a pink boat, in the ocean, mountains in background, realistic, high quality" />
+        </td>
+    </tr>
+</table>
+
 ### AnimateDiffSparseControlNetPipeline
 
 [SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion Models](https://arxiv.org/abs/2311.16933) for achieving controlled generation in text-to-video diffusion models by Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai.
@@ -762,6 +842,12 @@ pipe = AnimateDiffPipeline.from_pretrained("emilianJR/epiCRealism", motion_adapt
   - all
   - __call__
 
+## AnimateDiffControlNetPipeline
+
+[[autodoc]] AnimateDiffControlNetPipeline
+  - all
+  - __call__
+
 ## AnimateDiffSparseControlNetPipeline
 
 [[autodoc]] AnimateDiffSparseControlNetPipeline

src/diffusers/__init__.py

@@ -232,6 +232,7 @@ else:
             "AmusedImg2ImgPipeline",
             "AmusedInpaintPipeline",
             "AmusedPipeline",
+            "AnimateDiffControlNetPipeline",
             "AnimateDiffPipeline",
             "AnimateDiffSDXLPipeline",
             "AnimateDiffSparseControlNetPipeline",
@@ -652,6 +653,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AmusedImg2ImgPipeline,
             AmusedInpaintPipeline,
             AmusedPipeline,
+            AnimateDiffControlNetPipeline,
             AnimateDiffPipeline,
             AnimateDiffSDXLPipeline,
             AnimateDiffSparseControlNetPipeline,

src/diffusers/pipelines/__init__.py

@@ -118,6 +118,7 @@ else:
     _import_structure["amused"] = ["AmusedImg2ImgPipeline", "AmusedInpaintPipeline", "AmusedPipeline"]
     _import_structure["animatediff"] = [
         "AnimateDiffPipeline",
+        "AnimateDiffControlNetPipeline",
         "AnimateDiffSDXLPipeline",
         "AnimateDiffSparseControlNetPipeline",
         "AnimateDiffVideoToVideoPipeline",
@@ -419,6 +420,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     else:
         from .amused import AmusedImg2ImgPipeline, AmusedInpaintPipeline, AmusedPipeline
         from .animatediff import (
+            AnimateDiffControlNetPipeline,
             AnimateDiffPipeline,
             AnimateDiffSDXLPipeline,
             AnimateDiffSparseControlNetPipeline,

src/diffusers/pipelines/animatediff/__init__.py

@@ -22,6 +22,7 @@ except OptionalDependencyNotAvailable:
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_animatediff"] = ["AnimateDiffPipeline"]
+    _import_structure["pipeline_animatediff_controlnet"] = ["AnimateDiffControlNetPipeline"]
     _import_structure["pipeline_animatediff_sdxl"] = ["AnimateDiffSDXLPipeline"]
     _import_structure["pipeline_animatediff_sparsectrl"] = ["AnimateDiffSparseControlNetPipeline"]
     _import_structure["pipeline_animatediff_video2video"] = ["AnimateDiffVideoToVideoPipeline"]
@@ -35,6 +36,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
 
     else:
         from .pipeline_animatediff import AnimateDiffPipeline
+        from .pipeline_animatediff_controlnet import AnimateDiffControlNetPipeline
         from .pipeline_animatediff_sdxl import AnimateDiffSDXLPipeline
         from .pipeline_animatediff_sparsectrl import AnimateDiffSparseControlNetPipeline
         from .pipeline_animatediff_video2video import AnimateDiffVideoToVideoPipeline

src/diffusers/utils/__init__.py

@@ -93,7 +93,7 @@ from .import_utils import (
     is_xformers_available,
     requires_backends,
 )
-from .loading_utils import load_image
+from .loading_utils import load_image, load_video
 from .logging import get_logger
 from .outputs import BaseOutput
 from .peft_utils import (

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -77,6 +77,21 @@ class AmusedPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class AnimateDiffControlNetPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class AnimateDiffPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

src/diffusers/utils/export_utils.py

@@ -9,10 +9,7 @@ import numpy as np
 import PIL.Image
 import PIL.ImageOps
 
-from .import_utils import (
-    BACKENDS_MAPPING,
-    is_opencv_available,
-)
+from .import_utils import BACKENDS_MAPPING, is_opencv_available
 from .logging import get_logger
 
 

src/diffusers/utils/loading_utils.py

 import os
-from typing import Callable, Union
+import tempfile
+from typing import Callable, List, Optional, Union
 
 import PIL.Image
 import PIL.ImageOps
 import requests
 
+from .import_utils import BACKENDS_MAPPING, is_opencv_available
+
 
 def load_image(
-    image: Union[str, PIL.Image.Image], convert_method: Callable[[PIL.Image.Image], PIL.Image.Image] = None
+    image: Union[str, PIL.Image.Image], convert_method: Optional[Callable[[PIL.Image.Image], PIL.Image.Image]] = None
 ) -> PIL.Image.Image:
     """
     Loads `image` to a PIL Image.
@@ -15,7 +18,7 @@ def load_image(
     Args:
         image (`str` or `PIL.Image.Image`):
             The image to convert to the PIL Image format.
-        convert_method (Callable[[PIL.Image.Image], PIL.Image.Image], optional):
+        convert_method (Callable[[PIL.Image.Image], PIL.Image.Image], *optional*):
             A conversion method to apply to the image after loading it. When set to `None` the image will be converted
             "RGB".
 
@@ -47,3 +50,73 @@ def load_image(
         image = image.convert("RGB")
 
     return image
+
+
+def load_video(
+    video: str,
+    convert_method: Optional[Callable[[List[PIL.Image.Image]], List[PIL.Image.Image]]] = None,
+) -> List[PIL.Image.Image]:
+    """
+    Loads `video` to a list of PIL Image.
+
+    Args:
+        video (`str`):
+            A URL or Path to a video to convert to a list of PIL Image format.
+        convert_method (Callable[[List[PIL.Image.Image]], List[PIL.Image.Image]], *optional*):
+            A conversion method to apply to the video after loading it. When set to `None` the images will be converted
+            to "RGB".
+
+    Returns:
+        `List[PIL.Image.Image]`:
+            The video as a list of PIL images.
+    """
+    is_url = video.startswith("http://") or video.startswith("https://")
+    is_file = os.path.isfile(video)
+    was_tempfile_created = False
+
+    if not (is_url or is_file):
+        raise ValueError(
+            f"Incorrect path or URL. URLs must start with `http://` or `https://`, and {video} is not a valid path."
+        )
+
+    if is_url:
+        video_data = requests.get(video, stream=True).raw
+        video_path = tempfile.NamedTemporaryFile(suffix=os.path.splitext(video)[1], delete=False).name
+        was_tempfile_created = True
+        with open(video_path, "wb") as f:
+            f.write(video_data.read())
+
+        video = video_path
+
+    pil_images = []
+    if video.endswith(".gif"):
+        gif = PIL.Image.open(video)
+        try:
+            while True:
+                pil_images.append(gif.copy())
+                gif.seek(gif.tell() + 1)
+        except EOFError:
+            pass
+
+    else:
+        if is_opencv_available():
+            import cv2
+        else:
+            raise ImportError(BACKENDS_MAPPING["opencv"][1].format("load_video"))
+
+        video_capture = cv2.VideoCapture(video)
+        success, frame = video_capture.read()
+        while success:
+            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            pil_images.append(PIL.Image.fromarray(frame))
+            success, frame = video_capture.read()
+
+        video_capture.release()
+
+    if was_tempfile_created:
+        os.remove(video_path)
+
+    if convert_method is not None:
+        pil_images = convert_method(pil_images)
+
+    return pil_images

tests/pipelines/animatediff/test_animatediff_controlnet.py

+import unittest
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
+
+import diffusers
+from diffusers import (
+    AnimateDiffControlNetPipeline,
+    AutoencoderKL,
+    ControlNetModel,
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    LCMScheduler,
+    MotionAdapter,
+    StableDiffusionPipeline,
+    UNet2DConditionModel,
+    UNetMotionModel,
+)
+from diffusers.utils import logging
+from diffusers.utils.testing_utils import torch_device
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import (
+    IPAdapterTesterMixin,
+    PipelineFromPipeTesterMixin,
+    PipelineTesterMixin,
+    SDFunctionTesterMixin,
+)
+
+
+def to_np(tensor):
+    if isinstance(tensor, torch.Tensor):
+        tensor = tensor.detach().cpu().numpy()
+
+    return tensor
+
+
+class AnimateDiffControlNetPipelineFastTests(
+    IPAdapterTesterMixin, SDFunctionTesterMixin, PipelineTesterMixin, PipelineFromPipeTesterMixin, unittest.TestCase
+):
+    pipeline_class = AnimateDiffControlNetPipeline
+    params = TEXT_TO_IMAGE_PARAMS
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS.union({"conditioning_frames"})
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+
+    def get_dummy_components(self):
+        cross_attention_dim = 8
+        block_out_channels = (8, 8)
+
+        torch.manual_seed(0)
+        unet = UNet2DConditionModel(
+            block_out_channels=block_out_channels,
+            layers_per_block=2,
+            sample_size=8,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("CrossAttnDownBlock2D", "DownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            cross_attention_dim=cross_attention_dim,
+            norm_num_groups=2,
+        )
+        scheduler = DDIMScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            beta_schedule="linear",
+            clip_sample=False,
+        )
+        torch.manual_seed(0)
+        controlnet = ControlNetModel(
+            block_out_channels=block_out_channels,
+            layers_per_block=2,
+            in_channels=4,
+            down_block_types=("CrossAttnDownBlock2D", "DownBlock2D"),
+            cross_attention_dim=cross_attention_dim,
+            conditioning_embedding_out_channels=(8, 8),
+            norm_num_groups=1,
+        )
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=block_out_channels,
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+            norm_num_groups=2,
+        )
+        torch.manual_seed(0)
+        text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=cross_attention_dim,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+        )
+        text_encoder = CLIPTextModel(text_encoder_config)
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+        motion_adapter = MotionAdapter(
+            block_out_channels=block_out_channels,
+            motion_layers_per_block=2,
+            motion_norm_num_groups=2,
+            motion_num_attention_heads=4,
+        )
+
+        components = {
+            "unet": unet,
+            "controlnet": controlnet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "motion_adapter": motion_adapter,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "feature_extractor": None,
+            "image_encoder": None,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed: int = 0, num_frames: int = 2):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        video_height = 32
+        video_width = 32
+        conditioning_frames = [Image.new("RGB", (video_width, video_height))] * num_frames
+
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "conditioning_frames": conditioning_frames,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "num_frames": num_frames,
+            "guidance_scale": 7.5,
+            "output_type": "pt",
+        }
+        return inputs
+
+    def test_from_pipe_consistent_config(self):
+        assert self.original_pipeline_class == StableDiffusionPipeline
+        original_repo = "hf-internal-testing/tinier-stable-diffusion-pipe"
+        original_kwargs = {"requires_safety_checker": False}
+
+        # create original_pipeline_class(sd)
+        pipe_original = self.original_pipeline_class.from_pretrained(original_repo, **original_kwargs)
+
+        # original_pipeline_class(sd) -> pipeline_class
+        pipe_components = self.get_dummy_components()
+        pipe_additional_components = {}
+        for name, component in pipe_components.items():
+            if name not in pipe_original.components:
+                pipe_additional_components[name] = component
+
+        pipe = self.pipeline_class.from_pipe(pipe_original, **pipe_additional_components)
+
+        # pipeline_class -> original_pipeline_class(sd)
+        original_pipe_additional_components = {}
+        for name, component in pipe_original.components.items():
+            if name not in pipe.components or not isinstance(component, pipe.components[name].__class__):
+                original_pipe_additional_components[name] = component
+
+        pipe_original_2 = self.original_pipeline_class.from_pipe(pipe, **original_pipe_additional_components)
+
+        # compare the config
+        original_config = {k: v for k, v in pipe_original.config.items() if not k.startswith("_")}
+        original_config_2 = {k: v for k, v in pipe_original_2.config.items() if not k.startswith("_")}
+        assert original_config_2 == original_config
+
+    def test_motion_unet_loading(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+
+        assert isinstance(pipe.unet, UNetMotionModel)
+
+    @unittest.skip("Attention slicing is not enabled in this pipeline")
+    def test_attention_slicing_forward_pass(self):
+        pass
+
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array(
+                [
+                    0.6604,
+                    0.4099,
+                    0.4928,
+                    0.5706,
+                    0.5096,
+                    0.5012,
+                    0.6051,
+                    0.5169,
+                    0.5021,
+                    0.4864,
+                    0.4261,
+                    0.5779,
+                    0.5822,
+                    0.4049,
+                    0.5253,
+                    0.6160,
+                    0.4150,
+                    0.5155,
+                ]
+            )
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.6051, 0.5169, 0.5021, 0.6160, 0.4150, 0.5155])
+        return super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
+    def test_inference_batch_single_identical(
+        self,
+        batch_size=2,
+        expected_max_diff=1e-4,
+        additional_params_copy_to_batched_inputs=["num_inference_steps"],
+    ):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for components in pipe.components.values():
+            if hasattr(components, "set_default_attn_processor"):
+                components.set_default_attn_processor()
+
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        inputs = self.get_dummy_inputs(torch_device)
+        # Reset generator in case it is has been used in self.get_dummy_inputs
+        inputs["generator"] = self.get_generator(0)
+
+        logger = logging.get_logger(pipe.__module__)
+        logger.setLevel(level=diffusers.logging.FATAL)
+
+        # batchify inputs
+        batched_inputs = {}
+        batched_inputs.update(inputs)
+
+        for name in self.batch_params:
+            if name not in inputs:
+                continue
+
+            value = inputs[name]
+            if name == "prompt":
+                len_prompt = len(value)
+                batched_inputs[name] = [value[: len_prompt // i] for i in range(1, batch_size + 1)]
+                batched_inputs[name][-1] = 100 * "very long"
+
+            else:
+                batched_inputs[name] = batch_size * [value]
+
+        if "generator" in inputs:
+            batched_inputs["generator"] = [self.get_generator(i) for i in range(batch_size)]
+
+        if "batch_size" in inputs:
+            batched_inputs["batch_size"] = batch_size
+
+        for arg in additional_params_copy_to_batched_inputs:
+            batched_inputs[arg] = inputs[arg]
+
+        output = pipe(**inputs)
+        output_batch = pipe(**batched_inputs)
+
+        assert output_batch[0].shape[0] == batch_size
+
+        max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
+        assert max_diff < expected_max_diff
+
+    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
+    def test_to_device(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+
+        pipe.to("cpu")
+        # pipeline creates a new motion UNet under the hood. So we need to check the device from pipe.components
+        model_devices = [
+            component.device.type for component in pipe.components.values() if hasattr(component, "device")
+        ]
+        self.assertTrue(all(device == "cpu" for device in model_devices))
+
+        output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
+        self.assertTrue(np.isnan(output_cpu).sum() == 0)
+
+        pipe.to("cuda")
+        model_devices = [
+            component.device.type for component in pipe.components.values() if hasattr(component, "device")
+        ]
+        self.assertTrue(all(device == "cuda" for device in model_devices))
+
+        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
+        self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
+
+    def test_to_dtype(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+
+        # pipeline creates a new motion UNet under the hood. So we need to check the dtype from pipe.components
+        model_dtypes = [component.dtype for component in pipe.components.values() if hasattr(component, "dtype")]
+        self.assertTrue(all(dtype == torch.float32 for dtype in model_dtypes))
+
+        pipe.to(dtype=torch.float16)
+        model_dtypes = [component.dtype for component in pipe.components.values() if hasattr(component, "dtype")]
+        self.assertTrue(all(dtype == torch.float16 for dtype in model_dtypes))
+
+    def test_prompt_embeds(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        inputs.pop("prompt")
+        inputs["prompt_embeds"] = torch.randn((1, 4, pipe.text_encoder.config.hidden_size), device=torch_device)
+        pipe(**inputs)
+
+    def test_free_init(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffControlNetPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        inputs_normal = self.get_dummy_inputs(torch_device)
+        frames_normal = pipe(**inputs_normal).frames[0]
+
+        pipe.enable_free_init(
+            num_iters=2,
+            use_fast_sampling=True,
+            method="butterworth",
+            order=4,
+            spatial_stop_frequency=0.25,
+            temporal_stop_frequency=0.25,
+        )
+        inputs_enable_free_init = self.get_dummy_inputs(torch_device)
+        frames_enable_free_init = pipe(**inputs_enable_free_init).frames[0]
+
+        pipe.disable_free_init()
+        inputs_disable_free_init = self.get_dummy_inputs(torch_device)
+        frames_disable_free_init = pipe(**inputs_disable_free_init).frames[0]
+
+        sum_enabled = np.abs(to_np(frames_normal) - to_np(frames_enable_free_init)).sum()
+        max_diff_disabled = np.abs(to_np(frames_normal) - to_np(frames_disable_free_init)).max()
+        self.assertGreater(
+            sum_enabled, 1e1, "Enabling of FreeInit should lead to results different from the default pipeline results"
+        )
+        self.assertLess(
+            max_diff_disabled,
+            1e-4,
+            "Disabling of FreeInit should lead to results similar to the default pipeline results",
+        )
+
+    def test_free_init_with_schedulers(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffControlNetPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        inputs_normal = self.get_dummy_inputs(torch_device)
+        frames_normal = pipe(**inputs_normal).frames[0]
+
+        schedulers_to_test = [
+            DPMSolverMultistepScheduler.from_config(
+                components["scheduler"].config,
+                timestep_spacing="linspace",
+                beta_schedule="linear",
+                algorithm_type="dpmsolver++",
+                steps_offset=1,
+                clip_sample=False,
+            ),
+            LCMScheduler.from_config(
+                components["scheduler"].config,
+                timestep_spacing="linspace",
+                beta_schedule="linear",
+                steps_offset=1,
+                clip_sample=False,
+            ),
+        ]
+        components.pop("scheduler")
+
+        for scheduler in schedulers_to_test:
+            components["scheduler"] = scheduler
+            pipe: AnimateDiffControlNetPipeline = self.pipeline_class(**components)
+            pipe.set_progress_bar_config(disable=None)
+            pipe.to(torch_device)
+
+            pipe.enable_free_init(num_iters=2, use_fast_sampling=False)
+
+            inputs = self.get_dummy_inputs(torch_device)
+            frames_enable_free_init = pipe(**inputs).frames[0]
+            sum_enabled = np.abs(to_np(frames_normal) - to_np(frames_enable_free_init)).sum()
+
+            self.assertGreater(
+                sum_enabled,
+                1e1,
+                "Enabling of FreeInit should lead to results different from the default pipeline results",
+            )
+
+    def test_vae_slicing(self, video_count=2):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["prompt"] = [inputs["prompt"]] * video_count
+        inputs["conditioning_frames"] = [inputs["conditioning_frames"]] * video_count
+        output_1 = pipe(**inputs)
+
+        # make sure sliced vae decode yields the same result
+        pipe.enable_vae_slicing()
+        inputs = self.get_dummy_inputs(device)
+        inputs["prompt"] = [inputs["prompt"]] * video_count
+        inputs["conditioning_frames"] = [inputs["conditioning_frames"]] * video_count
+        output_2 = pipe(**inputs)
+
+        assert np.abs(output_2[0].flatten() - output_1[0].flatten()).max() < 1e-2