pipeline_pia.py

# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import inspect
import math
from dataclasses import dataclass
from typing import Any, Callable, Dict, List, Optional, Tuple, Union

import numpy as np
import PIL
import torch
import torch.fft as fft
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection

from ...image_processor import PipelineImageInput, VaeImageProcessor
from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel, UNetMotionModel
from ...models.lora import adjust_lora_scale_text_encoder
from ...models.unets.unet_motion_model import MotionAdapter
from ...schedulers import (
    DDIMScheduler,
    DPMSolverMultistepScheduler,
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
)
from ...utils import (
    USE_PEFT_BACKEND,
    BaseOutput,
    logging,
    replace_example_docstring,
    scale_lora_layers,
    unscale_lora_layers,
)
from ...utils.torch_utils import randn_tensor
from ..pipeline_utils import DiffusionPipeline


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from diffusers import (
        ...     EulerDiscreteScheduler,
        ...     MotionAdapter,
        ...     PIAPipeline,
        ... )
        >>> from diffusers.utils import export_to_gif, load_image
        >>> adapter = MotionAdapter.from_pretrained("../checkpoints/pia-diffusers")
        >>> pipe = PIAPipeline.from_pretrained("SG161222/Realistic_Vision_V6.0_B1_noVAE", motion_adapter=adapter)
        >>> pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
        >>> image = load_image(
        ...     "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/pix2pix/cat_6.png?download=true"
        ... )
        >>> image = image.resize((512, 512))
        >>> prompt = "cat in a hat"
        >>> negative_prompt = "wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg"
        >>> generator = torch.Generator("cpu").manual_seed(0)
        >>> output = pipe(image=image, prompt=prompt, negative_prompt=negative_prompt, generator=generator)
        >>> frames = output.frames[0]
        >>> export_to_gif(frames, "pia-animation.gif")
        ```
"""

RANGE_LIST = [
    [1.0, 0.9, 0.85, 0.85, 0.85, 0.8],  # 0 Small Motion
    [1.0, 0.8, 0.8, 0.8, 0.79, 0.78, 0.75],  # Moderate Motion
    [1.0, 0.8, 0.7, 0.7, 0.7, 0.7, 0.7, 0.7, 0.7, 0.7, 0.6, 0.5, 0.5],  # Large Motion
    [1.0, 0.9, 0.85, 0.85, 0.85, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.85, 0.85, 0.9, 1.0],  # Loop
    [1.0, 0.8, 0.8, 0.8, 0.79, 0.78, 0.75, 0.75, 0.75, 0.75, 0.75, 0.78, 0.79, 0.8, 0.8, 1.0],  # Loop
    [1.0, 0.8, 0.7, 0.7, 0.7, 0.7, 0.6, 0.5, 0.5, 0.6, 0.7, 0.7, 0.7, 0.7, 0.8, 1.0],  # Loop
    [0.5, 0.4, 0.4, 0.4, 0.35, 0.3],  # Style Transfer Candidate Small Motion
    [0.5, 0.4, 0.4, 0.4, 0.35, 0.35, 0.3, 0.25, 0.2],  # Style Transfer Moderate Motion
    [0.5, 0.2],  # Style Transfer Large Motion
]


# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
def tensor2vid(video: torch.Tensor, processor: "VaeImageProcessor", output_type: str = "np"):
    batch_size, channels, num_frames, height, width = video.shape
    outputs = []
    for batch_idx in range(batch_size):
        batch_vid = video[batch_idx].permute(1, 0, 2, 3)
        batch_output = processor.postprocess(batch_vid, output_type)

        outputs.append(batch_output)

    if output_type == "np":
        outputs = np.stack(outputs)

    elif output_type == "pt":
        outputs = torch.stack(outputs)

    elif not output_type == "pil":
        raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil]")

    return outputs


def prepare_mask_coef_by_statistics(num_frames: int, cond_frame: int, motion_scale: int):
    assert num_frames > 0, "video_length should be greater than 0"

    assert num_frames > cond_frame, "video_length should be greater than cond_frame"

    range_list = RANGE_LIST

    assert motion_scale < len(range_list), f"motion_scale type{motion_scale} not implemented"

    coef = range_list[motion_scale]
    coef = coef + ([coef[-1]] * (num_frames - len(coef)))

    order = [abs(i - cond_frame) for i in range(num_frames)]
    coef = [coef[order[i]] for i in range(num_frames)]

    return coef


def _get_freeinit_freq_filter(
    shape: Tuple[int, ...],
    device: Union[str, torch.dtype],
    filter_type: str,
    order: float,
    spatial_stop_frequency: float,
    temporal_stop_frequency: float,
) -> torch.Tensor:
    r"""Returns the FreeInit filter based on filter type and other input conditions."""

    time, height, width = shape[-3], shape[-2], shape[-1]
    mask = torch.zeros(shape)

    if spatial_stop_frequency == 0 or temporal_stop_frequency == 0:
        return mask

    if filter_type == "butterworth":

        def retrieve_mask(x):
            return 1 / (1 + (x / spatial_stop_frequency**2) ** order)
    elif filter_type == "gaussian":

        def retrieve_mask(x):
            return math.exp(-1 / (2 * spatial_stop_frequency**2) * x)
    elif filter_type == "ideal":

        def retrieve_mask(x):
            return 1 if x <= spatial_stop_frequency * 2 else 0
    else:
        raise NotImplementedError("`filter_type` must be one of gaussian, butterworth or ideal")

    for t in range(time):
        for h in range(height):
            for w in range(width):
                d_square = (
                    ((spatial_stop_frequency / temporal_stop_frequency) * (2 * t / time - 1)) ** 2
                    + (2 * h / height - 1) ** 2
                    + (2 * w / width - 1) ** 2
                )
                mask[..., t, h, w] = retrieve_mask(d_square)

    return mask.to(device)


def _freq_mix_3d(x: torch.Tensor, noise: torch.Tensor, LPF: torch.Tensor) -> torch.Tensor:
    r"""Noise reinitialization."""
    # FFT
    x_freq = fft.fftn(x, dim=(-3, -2, -1))
    x_freq = fft.fftshift(x_freq, dim=(-3, -2, -1))
    noise_freq = fft.fftn(noise, dim=(-3, -2, -1))
    noise_freq = fft.fftshift(noise_freq, dim=(-3, -2, -1))

    # frequency mix
    HPF = 1 - LPF
    x_freq_low = x_freq * LPF
    noise_freq_high = noise_freq * HPF
    x_freq_mixed = x_freq_low + noise_freq_high  # mix in freq domain

    # IFFT
    x_freq_mixed = fft.ifftshift(x_freq_mixed, dim=(-3, -2, -1))
    x_mixed = fft.ifftn(x_freq_mixed, dim=(-3, -2, -1)).real

    return x_mixed


@dataclass
class PIAPipelineOutput(BaseOutput):
    r"""
    Output class for PIAPipeline.

    Args:
        frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
        Nested list of length `batch_size` with denoised PIL image sequences of length `num_frames`,
        NumPy array of shape `(batch_size, num_frames, channels, height, width,
        Torch tensor of shape `(batch_size, num_frames, channels, height, width)`.
    """

    frames: Union[torch.Tensor, np.ndarray, List[List[PIL.Image.Image]]]


class PIAPipeline(
    DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin, FromSingleFileMixin
):
    r"""
    Pipeline for text-to-video generation.

    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.
        text_encoder ([`CLIPTextModel`]):
            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
        tokenizer (`CLIPTokenizer`):
            A [`~transformers.CLIPTokenizer`] to tokenize text.
        unet ([`UNet2DConditionModel`]):
            A [`UNet2DConditionModel`] used to create a UNetMotionModel to denoise the encoded video latents.
        motion_adapter ([`MotionAdapter`]):
            A [`MotionAdapter`] to be used in combination with `unet` 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`].
    """

    model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
    _optional_components = ["feature_extractor", "image_encoder", "motion_adapter"]
    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: Union[UNet2DConditionModel, UNetMotionModel],
        scheduler: Union[
            DDIMScheduler,
            PNDMScheduler,
            LMSDiscreteScheduler,
            EulerDiscreteScheduler,
            EulerAncestralDiscreteScheduler,
            DPMSolverMultistepScheduler,
        ],
        motion_adapter: Optional[MotionAdapter] = None,
        feature_extractor: CLIPImageProcessor = None,
        image_encoder: CLIPVisionModelWithProjection = None,
    ):
        super().__init__()
        if isinstance(unet, UNet2DConditionModel):
            unet = UNetMotionModel.from_unet2d(unet, motion_adapter)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            motion_adapter=motion_adapter,
            scheduler=scheduler,
            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)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with num_images_per_prompt -> num_videos_per_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.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.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.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.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()

    @property
    def free_init_enabled(self):
        return hasattr(self, "_free_init_num_iters") and self._free_init_num_iters is not None

    def enable_free_init(
        self,
        num_iters: int = 3,
        use_fast_sampling: bool = False,
        method: str = "butterworth",
        order: int = 4,
        spatial_stop_frequency: float = 0.25,
        temporal_stop_frequency: float = 0.25,
        generator: Optional[torch.Generator] = None,
    ):
        """Enables the FreeInit mechanism as in https://arxiv.org/abs/2312.07537.

        This implementation has been adapted from the [official repository](https://github.com/TianxingWu/FreeInit).

        Args:
            num_iters (`int`, *optional*, defaults to `3`):
                Number of FreeInit noise re-initialization iterations.
            use_fast_sampling (`bool`, *optional*, defaults to `False`):
                Whether or not to speedup sampling procedure at the cost of probably lower quality results. Enables
                the "Coarse-to-Fine Sampling" strategy, as mentioned in the paper, if set to `True`.
            method (`str`, *optional*, defaults to `butterworth`):
                Must be one of `butterworth`, `ideal` or `gaussian` to use as the filtering method for the
                FreeInit low pass filter.
            order (`int`, *optional*, defaults to `4`):
                Order of the filter used in `butterworth` method. Larger values lead to `ideal` method behaviour
                whereas lower values lead to `gaussian` method behaviour.
            spatial_stop_frequency (`float`, *optional*, defaults to `0.25`):
                Normalized stop frequency for spatial dimensions. Must be between 0 to 1. Referred to as `d_s` in
                the original implementation.
            temporal_stop_frequency (`float`, *optional*, defaults to `0.25`):
                Normalized stop frequency for temporal dimensions. Must be between 0 to 1. Referred to as `d_t` in
                the original implementation.
            generator (`torch.Generator`, *optional*, defaults to `0.25`):
                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                FreeInit generation deterministic.
        """
        self._free_init_num_iters = num_iters
        self._free_init_use_fast_sampling = use_fast_sampling
        self._free_init_method = method
        self._free_init_order = order
        self._free_init_spatial_stop_frequency = spatial_stop_frequency
        self._free_init_temporal_stop_frequency = temporal_stop_frequency
        self._free_init_generator = generator

    def disable_free_init(self):
        """Disables the FreeInit mechanism if enabled."""
        self._free_init_num_iters = None

    # 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,
        height,
        width,
        negative_prompt=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        ip_adapter_image=None,
        ip_adapter_image_embeds=None,
        callback_on_step_end_tensor_inputs=None,
    ):
        if height % 8 != 0 or width % 8 != 0:
            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")

        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )

        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}."
                )

        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
            raise ValueError(
                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
            )

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
    def prepare_ip_adapter_image_embeds(
        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt
    ):
        if ip_adapter_image_embeds is None:
            if not isinstance(ip_adapter_image, list):
                ip_adapter_image = [ip_adapter_image]

            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
                raise ValueError(
                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
                )

            image_embeds = []
            for single_ip_adapter_image, image_proj_layer in zip(
                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
            ):
                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
                single_image_embeds, single_negative_image_embeds = self.encode_image(
                    single_ip_adapter_image, device, 1, output_hidden_state
                )
                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
                single_negative_image_embeds = torch.stack(
                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
                )

                if self.do_classifier_free_guidance:
                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
                    single_image_embeds = single_image_embeds.to(device)

                image_embeds.append(single_image_embeds)
        else:
            image_embeds = ip_adapter_image_embeds
        return image_embeds

    # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
    def prepare_latents(
        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
    ):
        shape = (
            batch_size,
            num_channels_latents,
            num_frames,
            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

    def prepare_masked_condition(
        self,
        image,
        batch_size,
        num_channels_latents,
        num_frames,
        height,
        width,
        dtype,
        device,
        generator,
        motion_scale=0,
    ):
        shape = (
            batch_size,
            num_channels_latents,
            num_frames,
            height // self.vae_scale_factor,
            width // self.vae_scale_factor,
        )
        _, _, _, scaled_height, scaled_width = shape

        image = self.image_processor.preprocess(image)
        image = image.to(device, dtype)

        if isinstance(generator, list):
            image_latent = [
                self.vae.encode(image[k : k + 1]).latent_dist.sample(generator[k]) for k in range(batch_size)
            ]
            image_latent = torch.cat(image_latent, dim=0)
        else:
            image_latent = self.vae.encode(image).latent_dist.sample(generator)

        image_latent = image_latent.to(device=device, dtype=dtype)
        image_latent = torch.nn.functional.interpolate(image_latent, size=[scaled_height, scaled_width])
        image_latent_padding = image_latent.clone() * self.vae.config.scaling_factor

        mask = torch.zeros((batch_size, 1, num_frames, scaled_height, scaled_width)).to(device=device, dtype=dtype)
        mask_coef = prepare_mask_coef_by_statistics(num_frames, 0, motion_scale)
        masked_image = torch.zeros(batch_size, 4, num_frames, scaled_height, scaled_width).to(
            device=device, dtype=self.unet.dtype
        )
        for f in range(num_frames):
            mask[:, :, f, :, :] = mask_coef[f]
            masked_image[:, :, f, :, :] = image_latent_padding.clone()

        mask = torch.cat([mask] * 2) if self.do_classifier_free_guidance else mask
        masked_image = torch.cat([masked_image] * 2) if self.do_classifier_free_guidance else masked_image

        return mask, masked_image

    def _denoise_loop(
        self,
        timesteps,
        num_inference_steps,
        do_classifier_free_guidance,
        guidance_scale,
        num_warmup_steps,
        prompt_embeds,
        negative_prompt_embeds,
        latents,
        mask,
        masked_image,
        cross_attention_kwargs,
        added_cond_kwargs,
        extra_step_kwargs,
        callback_on_step_end,
        callback_on_step_end_tensor_inputs,
    ):
        """Denoising loop for PIA."""
        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)
                latent_model_input = torch.cat([latent_model_input, mask, masked_image], dim=1)

                # predict the noise residual
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    cross_attention_kwargs=cross_attention_kwargs,
                    added_cond_kwargs=added_cond_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)

                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample

                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)

                # 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()

        return latents

    def _free_init_loop(
        self,
        height,
        width,
        num_frames,
        batch_size,
        num_videos_per_prompt,
        denoise_args,
        device,
    ):
        """Denoising loop for PIA using FreeInit noise reinitialization technique."""

        latents = denoise_args.get("latents")
        prompt_embeds = denoise_args.get("prompt_embeds")
        timesteps = denoise_args.get("timesteps")
        num_inference_steps = denoise_args.get("num_inference_steps")

        latent_shape = (
            batch_size * num_videos_per_prompt,
            4,
            num_frames,
            height // self.vae_scale_factor,
            width // self.vae_scale_factor,
        )
        free_init_filter_shape = (
            1,
            4,
            num_frames,
            height // self.vae_scale_factor,
            width // self.vae_scale_factor,
        )
        free_init_freq_filter = _get_freeinit_freq_filter(
            shape=free_init_filter_shape,
            device=device,
            filter_type=self._free_init_method,
            order=self._free_init_order,
            spatial_stop_frequency=self._free_init_spatial_stop_frequency,
            temporal_stop_frequency=self._free_init_temporal_stop_frequency,
        )

        with self.progress_bar(total=self._free_init_num_iters) as free_init_progress_bar:
            for i in range(self._free_init_num_iters):
                # For the first FreeInit iteration, the original latent is used without modification.
                # Subsequent iterations apply the noise reinitialization technique.
                if i == 0:
                    initial_noise = latents.detach().clone()
                else:
                    current_diffuse_timestep = (
                        self.scheduler.config.num_train_timesteps - 1
                    )  # diffuse to t=999 noise level
                    diffuse_timesteps = torch.full((batch_size,), current_diffuse_timestep).long()
                    z_T = self.scheduler.add_noise(
                        original_samples=latents, noise=initial_noise, timesteps=diffuse_timesteps.to(device)
                    ).to(dtype=torch.float32)
                    z_rand = randn_tensor(
                        shape=latent_shape,
                        generator=self._free_init_generator,
                        device=device,
                        dtype=torch.float32,
                    )
                    latents = _freq_mix_3d(z_T, z_rand, LPF=free_init_freq_filter)
                    latents = latents.to(prompt_embeds.dtype)

                # Coarse-to-Fine Sampling for faster inference (can lead to lower quality)
                if self._free_init_use_fast_sampling:
                    current_num_inference_steps = int(num_inference_steps / self._free_init_num_iters * (i + 1))
                    self.scheduler.set_timesteps(current_num_inference_steps, device=device)
                    timesteps = self.scheduler.timesteps
                    denoise_args.update({"timesteps": timesteps, "num_inference_steps": current_num_inference_steps})

                num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
                denoise_args.update({"latents": latents, "num_warmup_steps": num_warmup_steps})
                latents = self._denoise_loop(**denoise_args)

                free_init_progress_bar.update()

        return latents

    # 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 :]
        if hasattr(self.scheduler, "set_begin_index"):
            self.scheduler.set_begin_index(t_start * self.scheduler.order)

        return timesteps, num_inference_steps - t_start

    def _retrieve_video_frames(self, latents, output_type, return_dict):
        """Helper function to handle latents to output conversion."""
        if output_type == "latent":
            return PIAPipelineOutput(frames=latents)

        video_tensor = self.decode_latents(latents)
        video = tensor2vid(video_tensor, self.image_processor, output_type=output_type)

        if not return_dict:
            return (video,)

        return PIAPipelineOutput(frames=video)

    @property
    def guidance_scale(self):
        return self._guidance_scale

    @property
    def clip_skip(self):
        return self._clip_skip

    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
    # corresponds to doing no classifier free guidance.
    @property
    def do_classifier_free_guidance(self):
        return self._guidance_scale > 1

    @property
    def cross_attention_kwargs(self):
        return self._cross_attention_kwargs

    @property
    def num_timesteps(self):
        return self._num_timesteps

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        image: PipelineImageInput,
        prompt: Union[str, List[str]] = None,
        strength: float = 1.0,
        num_frames: Optional[int] = 16,
        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_videos_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,
        ip_adapter_image: Optional[PipelineImageInput] = None,
        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
        motion_scale: int = 0,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        clip_skip: Optional[int] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
    ):
        r"""
        The call function to the pipeline for generation.

        Args:
            image (`PipelineImageInput`):
                The input image to be used for video generation.
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
            strength (`float`, *optional*, defaults to 1.0): Indicates extent to transform the reference `image`. Must be between 0 and 1.
            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
                The height in pixels of the generated video.
            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
                The width in pixels of the generated video.
            num_frames (`int`, *optional*, defaults to 16):
                The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds
                amounts to 2 seconds of video.
            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):
                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`).
            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 video
                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`. 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 (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.
            ip_adapter_image: (`PipelineImageInput`, *optional*):
                Optional image input to work with IP Adapters.
            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
                Pre-generated image embeddings for IP-Adapter. If not
                provided, embeddings are computed from the `ip_adapter_image` input argument.
            motion_scale: (`int`, *optional*, defaults to 0):
                Parameter that controls the amount and type of motion that is added to the image. Increasing the value increases the amount of motion, while specific
                ranges of values control the type of motion that is added. Must be between 0 and 8.
                Set between 0-2 to only increase the amount of motion.
                Set between 3-5 to create looping motion.
                Set between 6-8 to perform motion with image style transfer.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generated video. Choose between `torch.FloatTensor`, `PIL.Image` or
                `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead
                of a plain tuple.
            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).
            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.
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
                `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeine class.

        Examples:

        Returns:
            [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] or `tuple`:
                If `return_dict` is `True`, [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] is
                returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
        """
        # 0. Default height and width to unet
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        num_videos_per_prompt = 1

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            height,
            width,
            negative_prompt,
            prompt_embeds,
            negative_prompt_embeds,
            ip_adapter_image,
            ip_adapter_image_embeds,
            callback_on_step_end_tensor_inputs,
        )

        self._guidance_scale = guidance_scale
        self._clip_skip = clip_skip
        self._cross_attention_kwargs = cross_attention_kwargs

        # 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

        # 3. Encode input prompt
        text_encoder_lora_scale = (
            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
        )
        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
            prompt,
            device,
            num_videos_per_prompt,
            self.do_classifier_free_guidance,
            negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            lora_scale=text_encoder_lora_scale,
            clip_skip=self.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 self.do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])

        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
            image_embeds = self.prepare_ip_adapter_image_embeds(
                ip_adapter_image, ip_adapter_image_embeds, device, batch_size * num_videos_per_prompt
            )

        # 4. 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_videos_per_prompt)
        self._num_timesteps = len(timesteps)

        # 5. Prepare latent variables
        latents = self.prepare_latents(
            batch_size * num_videos_per_prompt,
            4,
            num_frames,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents=latents,
        )
        mask, masked_image = self.prepare_masked_condition(
            image,
            batch_size * num_videos_per_prompt,
            4,
            num_frames=num_frames,
            height=height,
            width=width,
            dtype=self.unet.dtype,
            device=device,
            generator=generator,
            motion_scale=motion_scale,
        )
        if strength < 1.0:
            noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype)
            latents = self.scheduler.add_noise(masked_image[0], noise, latent_timestep)

        # 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. 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
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        denoise_args = {
            "timesteps": timesteps,
            "num_inference_steps": num_inference_steps,
            "do_classifier_free_guidance": self.do_classifier_free_guidance,
            "guidance_scale": guidance_scale,
            "num_warmup_steps": num_warmup_steps,
            "prompt_embeds": prompt_embeds,
            "negative_prompt_embeds": negative_prompt_embeds,
            "latents": latents,
            "mask": mask,
            "masked_image": masked_image,
            "cross_attention_kwargs": self.cross_attention_kwargs,
            "added_cond_kwargs": added_cond_kwargs,
            "extra_step_kwargs": extra_step_kwargs,
            "callback_on_step_end": callback_on_step_end,
            "callback_on_step_end_tensor_inputs": callback_on_step_end_tensor_inputs,
        }

        if self.free_init_enabled:
            latents = self._free_init_loop(
                height=height,
                width=width,
                num_frames=num_frames,
                batch_size=batch_size,
                num_videos_per_prompt=num_videos_per_prompt,
                denoise_args=denoise_args,
                device=device,
            )
        else:
            latents = self._denoise_loop(**denoise_args)

        video = self._retrieve_video_frames(latents, output_type, return_dict)

        # 9. Offload all models
        self.maybe_free_model_hooks()

        return video