Add video img2img (#3900)

* Add image to image video

* Improve

* better naming

* make fix copies

* add docs

* finish tests

* trigger tests

* make style

* correct

* finish

* Fix more

* make style

* finish

docs/source/en/api/pipelines/text_to_video.mdx

@@ -37,9 +37,12 @@ Resources:
 | Pipeline | Tasks | Demo
 |---|---|:---:|
 | [TextToVideoSDPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py) | *Text-to-Video Generation* | [🤗 Spaces](https://huggingface.co/spaces/damo-vilab/modelscope-text-to-video-synthesis)
+| [VideoToVideoSDPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py) | *Text-Guided Video-to-Video Generation* | [(TODO)🤗 Spaces]()
 
 ## Usage example 
 
+### `text-to-video-ms-1.7b`
+
 Let's start by generating a short video with the default length of 16 frames (2s at 8 fps):
 
 ```python 
@@ -119,12 +122,72 @@ Here are some sample outputs:
     </tr>
 </table>
 
+### `cerspense/zeroscope_v2_576w` & `cerspense/zeroscope_v2_XL`
+
+Zeroscope are watermark-free model and have been trained on specific sizes such as `576x320` and `1024x576`.
+One should first generate a video using the lower resolution checkpoint [`cerspense/zeroscope_v2_576w`](https://huggingface.co/cerspense/zeroscope_v2_576w) with [`TextToVideoSDPipeline`],
+which can then be upscaled using [`VideoToVideoSDPipeline`] and [`cerspense/zeroscope_v2_XL`](https://huggingface.co/cerspense/zeroscope_v2_XL).
+
+
+```py
+import torch
+from diffusers import DiffusionPipeline
+from diffusers.utils import export_to_video
+
+pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_576w", torch_dtype=torch.float16)
+pipe.enable_model_cpu_offload()
+
+# memory optimization
+pipe.enable_vae_slicing()
+
+prompt = "Darth Vader surfing a wave"
+video_frames = pipe(prompt, num_frames=24).frames
+video_path = export_to_video(video_frames)
+video_path
+```
+
+Now the video can be upscaled:
+
+```py
+pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_XL", torch_dtype=torch.float16)
+pipe.vae.enable_slicing()
+pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+pipe.enable_model_cpu_offload()
+
+video = [Image.fromarray(frame).resize((1024, 576)) for frame in video_frames]
+
+video_frames = pipe(prompt, video=video, strength=0.6).frames
+video_path = export_to_video(video_frames)
+video_path
+```
+
+Here are some sample outputs: 
+
+<table>
+    <tr>
+        <td ><center>
+        Darth vader surfing in waves.
+        <br>
+        <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/darthvader_cerpense.gif"
+            alt="Darth vader surfing in waves."
+            style="width: 576px;" />
+        </center></td>
+    </tr>
+</table>
+
 ## Available checkpoints 
 
 * [damo-vilab/text-to-video-ms-1.7b](https://huggingface.co/damo-vilab/text-to-video-ms-1.7b/)
 * [damo-vilab/text-to-video-ms-1.7b-legacy](https://huggingface.co/damo-vilab/text-to-video-ms-1.7b-legacy)
+* [cerspense/zeroscope_v2_576w](https://huggingface.co/cerspense/zeroscope_v2_576w)
+* [cerspense/zeroscope_v2_XL](https://huggingface.co/cerspense/zeroscope_v2_XL)
 
 ## TextToVideoSDPipeline
 [[autodoc]] TextToVideoSDPipeline
 	- all
 	- __call__
+
+## VideoToVideoSDPipeline
+[[autodoc]] VideoToVideoSDPipeline
+	- all
+	- __call__

src/diffusers/__init__.py

@@ -173,6 +173,7 @@ else:
         VersatileDiffusionImageVariationPipeline,
         VersatileDiffusionPipeline,
         VersatileDiffusionTextToImagePipeline,
+        VideoToVideoSDPipeline,
         VQDiffusionPipeline,
     )
 

src/diffusers/models/autoencoder_kl.py

@@ -229,7 +229,12 @@ class AutoencoderKL(ModelMixin, ConfigMixin):
         if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
             return self.tiled_encode(x, return_dict=return_dict)
 
-        h = self.encoder(x)
+        if self.use_slicing and x.shape[0] > 1:
+            encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)]
+            h = torch.cat(encoded_slices)
+        else:
+            h = self.encoder(x)
+
         moments = self.quant_conv(h)
         posterior = DiagonalGaussianDistribution(moments)
 

src/diffusers/pipelines/__init__.py

@@ -89,7 +89,7 @@ else:
         StableUnCLIPPipeline,
     )
     from .stable_diffusion_safe import StableDiffusionPipelineSafe
-    from .text_to_video_synthesis import TextToVideoSDPipeline, TextToVideoZeroPipeline
+    from .text_to_video_synthesis import TextToVideoSDPipeline, TextToVideoZeroPipeline, VideoToVideoSDPipeline
     from .unclip import UnCLIPImageVariationPipeline, UnCLIPPipeline
     from .unidiffuser import ImageTextPipelineOutput, UniDiffuserModel, UniDiffuserPipeline, UniDiffuserTextDecoder
     from .versatile_diffusion import (

src/diffusers/pipelines/text_to_video_synthesis/__init__.py

@@ -28,5 +28,6 @@ try:
 except OptionalDependencyNotAvailable:
     from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
 else:
-    from .pipeline_text_to_video_synth import TextToVideoSDPipeline  # noqa: F401
+    from .pipeline_text_to_video_synth import TextToVideoSDPipeline
+    from .pipeline_text_to_video_synth_img2img import VideoToVideoSDPipeline  # noqa: F401
     from .pipeline_text_to_video_zero import TextToVideoZeroPipeline

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py

@@ -672,6 +672,9 @@ class TextToVideoSDPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lora
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
 
+        if output_type == "latent":
+            return TextToVideoSDPipelineOutput(frames=latents)
+
         video_tensor = self.decode_latents(latents)
 
         if output_type == "pt":

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py

@@ -0,0 +1,770 @@
+# 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
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet3DConditionModel
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+    is_accelerate_available,
+    is_accelerate_version,
+    logging,
+    randn_tensor,
+    replace_example_docstring,
+)
+from ..pipeline_utils import DiffusionPipeline
+from . import TextToVideoSDPipelineOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
+        >>> from diffusers.utils import export_to_video
+
+        >>> pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_576w", torch_dtype=torch.float16)
+        >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "spiderman running in the desert"
+        >>> video_frames = pipe(prompt, num_inference_steps=40, height=320, width=576, num_frames=24).frames
+        >>> # safe low-res video
+        >>> video_path = export_to_video(video_frames, output_video_path="./video_576_spiderman.mp4")
+
+        >>> # let's offload the text-to-image model
+        >>> pipe.to("cpu")
+
+        >>> # and load the image-to-image model
+        >>> pipe = DiffusionPipeline.from_pretrained(
+        ...     "cerspense/zeroscope_v2_XL", torch_dtype=torch.float16, revision="refs/pr/15"
+        ... )
+        >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+        >>> pipe.enable_model_cpu_offload()
+
+        >>> # The VAE consumes A LOT of memory, let's make sure we run it in sliced mode
+        >>> pipe.vae.enable_slicing()
+
+        >>> # now let's upscale it
+        >>> video = [Image.fromarray(frame).resize((1024, 576)) for frame in video_frames]
+
+        >>> # and denoise it
+        >>> video_frames = pipe(prompt, video=video, strength=0.6).frames
+        >>> video_path = export_to_video(video_frames, output_video_path="./video_1024_spiderman.mp4")
+        >>> video_path
+        ```
+"""
+
+
+def tensor2vid(video: torch.Tensor, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) -> List[np.ndarray]:
+    # This code is copied from https://github.com/modelscope/modelscope/blob/1509fdb973e5871f37148a4b5e5964cafd43e64d/modelscope/pipelines/multi_modal/text_to_video_synthesis_pipeline.py#L78
+    # reshape to ncfhw
+    mean = torch.tensor(mean, device=video.device).reshape(1, -1, 1, 1, 1)
+    std = torch.tensor(std, device=video.device).reshape(1, -1, 1, 1, 1)
+    # unnormalize back to [0,1]
+    video = video.mul_(std).add_(mean)
+    video.clamp_(0, 1)
+    # prepare the final outputs
+    i, c, f, h, w = video.shape
+    images = video.permute(2, 3, 0, 4, 1).reshape(
+        f, h, i * w, c
+    )  # 1st (frames, h, batch_size, w, c) 2nd (frames, h, batch_size * w, c)
+    images = images.unbind(dim=0)  # prepare a list of indvidual (consecutive frames)
+    images = [(image.cpu().numpy() * 255).astype("uint8") for image in images]  # f h w c
+    return images
+
+
+def preprocess_video(video):
+    supported_formats = (np.ndarray, torch.Tensor, PIL.Image.Image)
+
+    if isinstance(video, supported_formats):
+        video = [video]
+    elif not (isinstance(video, list) and all(isinstance(i, supported_formats) for i in video)):
+        raise ValueError(
+            f"Input is in incorrect format: {[type(i) for i in video]}. Currently, we only support {', '.join(supported_formats)}"
+        )
+
+    if isinstance(video[0], PIL.Image.Image):
+        video = [np.array(frame) for frame in video]
+
+    if isinstance(video[0], np.ndarray):
+        video = np.concatenate(video, axis=0) if video[0].ndim == 5 else np.stack(video, axis=0)
+
+        if video.dtype == np.uint8:
+            video = np.array(video).astype(np.float32) / 255.0
+
+        if video.ndim == 4:
+            video = video[None, ...]
+
+        video = torch.from_numpy(video.transpose(0, 4, 1, 2, 3))
+
+    elif isinstance(video[0], torch.Tensor):
+        video = torch.cat(video, axis=0) if video[0].ndim == 5 else torch.stack(video, axis=0)
+
+        # don't need any preprocess if the video is latents
+        channel = video.shape[1]
+        if channel == 4:
+            return video
+
+        # move channels before num_frames
+        video = video.permute(0, 2, 1, 3, 4)
+
+    # normalize video
+    video = 2.0 * video - 1.0
+
+    return video
+
+
+class VideoToVideoSDPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation.
+
+    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 videos to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Same as Stable Diffusion 2.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet3DConditionModel`]): Conditional U-Net architecture to denoise the encoded video latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet3DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+    # 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 invoked, 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 to save a large amount of memory and to allow the processing of 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 invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded
+        to GPU only when their specific submodule has its `forward` method called. Note that offloading happens on a
+        submodule basis. Memory savings are higher than with `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            cpu_offload(cpu_offloaded_model, device)
+
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.vae, self.unet]:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    # 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,
+    ):
+        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.
+        """
+        # 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
+
+        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
+
+            prompt_embeds = self.text_encoder(
+                text_input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            prompt_embeds = prompt_embeds[0]
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.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=self.text_encoder.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)
+
+            # 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
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
+    def decode_latents(self, latents):
+        latents = 1 / self.vae.config.scaling_factor * latents
+
+        batch_size, channels, num_frames, height, width = latents.shape
+        latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+        image = self.vae.decode(latents).sample
+        video = (
+            image[None, :]
+            .reshape(
+                (
+                    batch_size,
+                    num_frames,
+                    -1,
+                )
+                + image.shape[2:]
+            )
+            .permute(0, 2, 1, 3, 4)
+        )
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        video = video.float()
+        return video
+
+    # 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
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.check_inputs
+    def check_inputs(
+        self, prompt, strength, callback_steps, negative_prompt=None, prompt_embeds=None, negative_prompt_embeds=None
+    ):
+        if strength < 0 or strength > 1:
+            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            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}."
+                )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+    def get_timesteps(self, num_inference_steps, strength, device):
+        # get the original timestep using init_timestep
+        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+        t_start = max(num_inference_steps - init_timestep, 0)
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+        return timesteps, num_inference_steps - t_start
+
+    def prepare_latents(self, video, timestep, batch_size, dtype, device, generator=None):
+        video = video.to(device=device, dtype=dtype)
+
+        # change from (b, c, f, h, w) -> (b * f, c, w, h)
+        bsz, channel, frames, width, height = video.shape
+        video = video.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+
+        if video.shape[1] == 4:
+            init_latents = video
+        else:
+            if isinstance(generator, list) and len(generator) != batch_size:
+                raise ValueError(
+                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                )
+
+            elif isinstance(generator, list):
+                init_latents = [
+                    self.vae.encode(video[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
+                ]
+                init_latents = torch.cat(init_latents, dim=0)
+            else:
+                init_latents = self.vae.encode(video).latent_dist.sample(generator)
+
+            init_latents = self.vae.config.scaling_factor * init_latents
+
+        if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+            raise ValueError(
+                f"Cannot duplicate `video` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+            )
+        else:
+            init_latents = torch.cat([init_latents], dim=0)
+
+        shape = init_latents.shape
+        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+        # get latents
+        init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+        latents = init_latents
+
+        latents = latents[None, :].reshape((bsz, frames, latents.shape[1]) + latents.shape[2:]).permute(0, 2, 1, 3, 4)
+
+        return latents
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        video: Union[List[np.ndarray], torch.FloatTensor] = None,
+        strength: float = 0.6,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 15.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        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] = "np",
+        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,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the video generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            video: (`List[np.ndarray]` or `torch.FloatTensor`):
+                `video` frames or tensor representing a video batch, that will be used as the starting point for the
+                process. Can also accpet video latents as `image`, if passing latents directly, it will not be encoded
+                again.
+            strength (`float`, *optional*, defaults to 0.8):
+                Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+                will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+                denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+                be maximum and the denoising process will run for the full number of iterations specified in
+                `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate videos that are closely linked to the text `prompt`,
+                usually at the expense of lower video quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the video 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`).
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](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 video
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`. Latents should be of shape
+                `(batch_size, num_channel, num_frames, height, width)`.
+            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.
+            output_type (`str`, *optional*, defaults to `"np"`):
+                The output format of the generate video. Choose between `torch.FloatTensor` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.TextToVideoSDPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                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 will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.TextToVideoSDPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.TextToVideoSDPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated frames.
+        """
+        # 0. Default height and width to unet
+        num_images_per_prompt = 1
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(prompt, strength, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+        # 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 = 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,
+        )
+
+        # 4. Preprocess video
+        video = preprocess_video(video)
+
+        # 5. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+        # 5. Prepare latent variables
+        latents = self.prepare_latents(video, latent_timestep, batch_size, prompt_embeds.dtype, device, generator)
+
+        # 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)
+
+        # 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:
+            for i, t in enumerate(timesteps):
+                # 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
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                ).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)
+
+                # reshape latents
+                bsz, channel, frames, width, height = latents.shape
+                latents = latents.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+                noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+                # reshape latents back
+                latents = latents[None, :].reshape(bsz, frames, channel, width, height).permute(0, 2, 1, 3, 4)
+
+                # 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:
+                        callback(i, t, latents)
+
+        if output_type == "latent":
+            return TextToVideoSDPipelineOutput(frames=latents)
+
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.unet.to("cpu")
+
+        video_tensor = self.decode_latents(latents)
+
+        if output_type == "pt":
+            video = video_tensor
+        else:
+            video = tensor2vid(video_tensor)
+
+        # Offload last model to CPU
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.final_offload_hook.offload()
+
+        if not return_dict:
+            return (video,)
+
+        return TextToVideoSDPipelineOutput(frames=video)

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -782,6 +782,21 @@ class VersatileDiffusionTextToImagePipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class VideoToVideoSDPipeline(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 VQDiffusionPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/test_pipelines_common.py

@@ -652,11 +652,11 @@ class PipelineTesterMixin:
         pipe.set_progress_bar_config(disable=None)
 
         inputs = self.get_dummy_inputs(torch_device)
-        output_without_offload = pipe(**inputs)[0]
+        output_without_offload = pipe(**inputs)[0].cpu()
 
         pipe.enable_xformers_memory_efficient_attention()
         inputs = self.get_dummy_inputs(torch_device)
-        output_with_offload = pipe(**inputs)[0]
+        output_with_offload = pipe(**inputs)[0].cpu()
 
         if test_max_difference:
             max_diff = np.abs(output_with_offload - output_without_offload).max()

tests/pipelines/text_to_video/test_video_to_video.py

@@ -0,0 +1,195 @@
+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# 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 random
+import unittest
+
+import numpy as np
+import torch
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
+
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    UNet3DConditionModel,
+    VideoToVideoSDPipeline,
+)
+from diffusers.utils import floats_tensor, is_xformers_available, skip_mps
+from diffusers.utils.testing_utils import enable_full_determinism, slow, torch_device
+
+from ..pipeline_params import (
+    TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS,
+    TEXT_GUIDED_IMAGE_VARIATION_PARAMS,
+)
+from ..test_pipelines_common import PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+@skip_mps
+class VideoToVideoSDPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = VideoToVideoSDPipeline
+    params = TEXT_GUIDED_IMAGE_VARIATION_PARAMS.union({"video"}) - {"image", "width", "height"}
+    batch_params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS.union({"video"}) - {"image"}
+    required_optional_params = PipelineTesterMixin.required_optional_params - {"latents"}
+    test_attention_slicing = False
+
+    # No `output_type`.
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback",
+            "callback_steps",
+        ]
+    )
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        unet = UNet3DConditionModel(
+            block_out_channels=(32, 64, 64, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("CrossAttnDownBlock3D", "CrossAttnDownBlock3D", "CrossAttnDownBlock3D", "DownBlock3D"),
+            up_block_types=("UpBlock3D", "CrossAttnUpBlock3D", "CrossAttnUpBlock3D", "CrossAttnUpBlock3D"),
+            cross_attention_dim=32,
+            attention_head_dim=4,
+        )
+        scheduler = DDIMScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            beta_schedule="scaled_linear",
+            clip_sample=False,
+            set_alpha_to_one=False,
+        )
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+            sample_size=128,
+        )
+        torch.manual_seed(0)
+        text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            hidden_act="gelu",
+            projection_dim=512,
+        )
+        text_encoder = CLIPTextModel(text_encoder_config)
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        components = {
+            "unet": unet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        # 3 frames
+        video = floats_tensor((1, 3, 3, 32, 32), rng=random.Random(seed)).to(device)
+
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "video": video,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "output_type": "pt",
+        }
+        return inputs
+
+    def test_text_to_video_default_case(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        sd_pipe = VideoToVideoSDPipeline(**components)
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["output_type"] = "np"
+        frames = sd_pipe(**inputs).frames
+        image_slice = frames[0][-3:, -3:, -1]
+
+        assert frames[0].shape == (32, 32, 3)
+        expected_slice = np.array([106, 117, 113, 174, 137, 112, 148, 151, 131])
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_attention_forwardGenerator_pass(self):
+        self._test_xformers_attention_forwardGenerator_pass(test_mean_pixel_difference=False, expected_max_diff=5e-3)
+
+    # (todo): sayakpaul
+    @unittest.skip(reason="Batching needs to be properly figured out first for this pipeline.")
+    def test_inference_batch_consistent(self):
+        pass
+
+    # (todo): sayakpaul
+    @unittest.skip(reason="Batching needs to be properly figured out first for this pipeline.")
+    def test_inference_batch_single_identical(self):
+        pass
+
+    @unittest.skip(reason="`num_images_per_prompt` argument is not supported for this pipeline.")
+    def test_num_images_per_prompt(self):
+        pass
+
+    def test_progress_bar(self):
+        return super().test_progress_bar()
+
+
+@slow
+@skip_mps
+class VideoToVideoSDPipelineSlowTests(unittest.TestCase):
+    def test_two_step_model(self):
+        pipe = VideoToVideoSDPipeline.from_pretrained("cerspense/zeroscope_v2_XL", torch_dtype=torch.float16)
+        pipe.enable_model_cpu_offload()
+
+        # 10 frames
+        generator = torch.Generator(device="cpu").manual_seed(0)
+        video = torch.randn((1, 10, 3, 1024, 576), generator=generator)
+        video = video.to("cuda")
+
+        prompt = "Spiderman is surfing"
+
+        video_frames = pipe(prompt, video=video, generator=generator, num_inference_steps=3, output_type="pt").frames
+
+        expected_array = np.array([-1.0458984, -1.1279297, -0.9663086, -0.91503906, -0.75097656])
+        assert np.abs(video_frames.cpu().numpy()[0, 0, 0, 0, -5:] - expected_array).sum() < 1e-2