sygil-webui/scripts/stable_diffusion_pipeline.py

import inspect
from typing import Callable, List, Optional, Union
from pathlib import Path
from torchvision.transforms.functional import pil_to_tensor
import librosa
from PIL import Image
from torchvision.io import write_video
import numpy as np
import time
import json

import torch
from diffusers.configuration_utils import FrozenDict
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.utils import deprecate, logging
from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput

from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
from torch import nn

from sd_utils import RealESRGANModel

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


def get_timesteps_arr(audio_filepath, offset, duration, fps=30, margin=1.0, smooth=0.0):
    y, sr = librosa.load(audio_filepath, offset=offset, duration=duration)

    # librosa.stft hardcoded defaults...
    # n_fft defaults to 2048
    # hop length is win_length // 4
    # win_length defaults to n_fft
    D = librosa.stft(y, n_fft=2048, hop_length=2048 // 4, win_length=2048)

    # Extract percussive elements
    D_harmonic, D_percussive = librosa.decompose.hpss(D, margin=margin)
    y_percussive = librosa.istft(D_percussive, length=len(y))

    # Get normalized melspectrogram
    spec_raw = librosa.feature.melspectrogram(y=y_percussive, sr=sr)
    spec_max = np.amax(spec_raw, axis=0)
    spec_norm = (spec_max - np.min(spec_max)) / np.ptp(spec_max)

    # Resize cumsum of spec norm to our desired number of interpolation frames
    x_norm = np.linspace(0, spec_norm.shape[-1], spec_norm.shape[-1])
    y_norm = np.cumsum(spec_norm)
    y_norm /= y_norm[-1]
    x_resize = np.linspace(0, y_norm.shape[-1], int(duration*fps))

    T = np.interp(x_resize, x_norm, y_norm)

    # Apply smoothing
    return T * (1 - smooth) + np.linspace(0.0, 1.0, T.shape[0]) * smooth


def slerp(t, v0, v1, DOT_THRESHOLD=0.9995):
    """helper function to spherically interpolate two arrays v1 v2"""

    if not isinstance(v0, np.ndarray):
        inputs_are_torch = True
        input_device = v0.device
        v0 = v0.cpu().numpy()
        v1 = v1.cpu().numpy()

    dot = np.sum(v0 * v1 / (np.linalg.norm(v0) * np.linalg.norm(v1)))
    if np.abs(dot) > DOT_THRESHOLD:
        v2 = (1 - t) * v0 + t * v1
    else:
        theta_0 = np.arccos(dot)
        sin_theta_0 = np.sin(theta_0)
        theta_t = theta_0 * t
        sin_theta_t = np.sin(theta_t)
        s0 = np.sin(theta_0 - theta_t) / sin_theta_0
        s1 = sin_theta_t / sin_theta_0
        v2 = s0 * v0 + s1 * v1

    if inputs_are_torch:
        v2 = torch.from_numpy(v2).to(input_device)

    return v2


def make_video_pyav(
    frames_or_frame_dir: Union[str, Path, torch.Tensor],
    audio_filepath: Union[str, Path] = None,
    fps: int = 30,
    audio_offset: int = 0,
    audio_duration: int = 2,
    sr: int = 22050,
    output_filepath: Union[str, Path] = "output.mp4",
    glob_pattern: str = "*.png",
):
    """
    TODO - docstring here

    frames_or_frame_dir: (Union[str, Path, torch.Tensor]):
        Either a directory of images, or a tensor of shape (T, C, H, W) in range [0, 255].
    """

    # Torchvision write_video doesn't support pathlib paths
    output_filepath = str(output_filepath)

    if isinstance(frames_or_frame_dir, (str, Path)):
        frames = None
        for img in sorted(Path(frames_or_frame_dir).glob(glob_pattern)):
            frame = pil_to_tensor(Image.open(img)).unsqueeze(0)
            frames = frame if frames is None else torch.cat([frames, frame])
    else:
        frames = frames_or_frame_dir

    # TCHW -> THWC
    frames = frames.permute(0, 2, 3, 1)

    if audio_filepath:
        # Read audio, convert to tensor
        audio, sr = librosa.load(audio_filepath, sr=sr, mono=True, offset=audio_offset, duration=audio_duration)
        audio_tensor = torch.tensor(audio).unsqueeze(0)

        write_video(
            output_filepath,
            frames,
            fps=fps,
            audio_array=audio_tensor,
            audio_fps=sr,
            audio_codec="aac",
            options={"crf": "10", "pix_fmt": "yuv420p"},
        )
    else:
        write_video(output_filepath, frames, fps=fps, options={"crf": "10", "pix_fmt": "yuv420p"})

    return output_filepath


class StableDiffusionWalkPipeline(DiffusionPipeline):
    r"""
    Pipeline for generating videos by interpolating  Stable Diffusion's latent space.
    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModel`]):
            Frozen text-encoder. Stable Diffusion uses the text portion of
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
        safety_checker ([`StableDiffusionSafetyChecker`]):
            Classification module that estimates whether generated images could be considered offensive or harmful.
            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
        feature_extractor ([`CLIPFeatureExtractor`]):
            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
    """

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
        safety_checker: StableDiffusionSafetyChecker,
        feature_extractor: CLIPFeatureExtractor,
    ):
        super().__init__()

        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
                " file"
            )
            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(scheduler.config)
            new_config["steps_offset"] = 1
            scheduler._internal_dict = FrozenDict(new_config)

        if safety_checker is None:
            logger.warn(
                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
            )

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

    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
        r"""
        Enable sliced attention computation.
        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
        in several steps. This is useful to save some memory in exchange for a small speed decrease.
        Args:
            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
                `attention_head_dim` must be a multiple of `slice_size`.
        """
        if slice_size == "auto":
            # half the attention head size is usually a good trade-off between
            # speed and memory
            slice_size = self.unet.config.attention_head_dim // 2
        self.unet.set_attention_slice(slice_size)

    def disable_attention_slicing(self):
        r"""
        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
        back to computing attention in one step.
        """
        # set slice_size = `None` to disable `attention slicing`
        self.enable_attention_slicing(None)

    @torch.no_grad()
    def __call__(
        self,
        prompt: Optional[Union[str, List[str]]] = None,
        height: int = 512,
        width: int = 512,
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[torch.Generator] = None,
        latents: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: Optional[int] = 1,
        text_embeddings: Optional[torch.FloatTensor] = None,
        **kwargs,
    ):
        r"""
        Function invoked when calling the pipeline for generation.
        Args:
            prompt (`str` or `List[str]`, *optional*, defaults to `None`):
                The prompt or prompts to guide the image generation. If not provided, `text_embeddings` is required.
            height (`int`, *optional*, defaults to 512):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to 512):
                The width in pixels of the generated image.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            guidance_scale (`float`, *optional*, defaults to 7.5):
                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 images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            generator (`torch.Generator`, *optional*):
                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
                deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            callback (`Callable`, *optional*):
                A function that 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.
            text_embeddings (`torch.FloatTensor`, *optional*, defaults to `None`):
                Pre-generated text embeddings to be used as inputs for image generation. Can be used in place of
                `prompt` to avoid re-computing the embeddings. If not provided, the embeddings will be generated from
                the supplied `prompt`.
        Returns:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
            When returning a tuple, the first element is a list with the generated images, and the second element is a
            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
            (nsfw) content, according to the `safety_checker`.
        """

        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_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 text_embeddings is None:
            if isinstance(prompt, str):
                batch_size = 1
            elif isinstance(prompt, list):
                batch_size = len(prompt)
            else:
                raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")

            # get prompt text embeddings
            text_inputs = self.tokenizer(
                prompt,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids

            if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
                removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
                print(
                    "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}"
                )
                text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
            text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
        else:
            batch_size = text_embeddings.shape[0]

        # duplicate text embeddings for each generation per prompt, using mps friendly method
        bs_embed, seq_len, _ = text_embeddings.shape
        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)

        # 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
        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""]
            elif 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

            max_length = self.tokenizer.model_max_length
            uncond_input = self.tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="pt",
            )
            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]

            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = uncond_embeddings.shape[1]
            uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1)
            uncond_embeddings = uncond_embeddings.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
            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])

        # get the initial random noise unless the user supplied it

        # Unlike in other pipelines, latents need to be generated in the target device
        # for 1-to-1 results reproducibility with the CompVis implementation.
        # However this currently doesn't work in `mps`.
        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
        latents_dtype = text_embeddings.dtype
        if latents is None:
            if self.device.type == "mps":
                # randn does not exist on mps
                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
                    self.device
                )
            else:
                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
        else:
            if latents.shape != latents_shape:
                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
            latents = latents.to(self.device)

        # set timesteps
        self.scheduler.set_timesteps(num_inference_steps)

        # Some schedulers like PNDM have timesteps as arrays
        # It's more optimized to move all timesteps to correct device beforehand
        timesteps_tensor = self.scheduler.timesteps.to(self.device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma

        # 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

        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
            # 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=text_embeddings).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

            # call the callback, if provided
            if callback is not None and i % callback_steps == 0:
                callback(i, t, latents)

        latents = 1 / 0.18215 * latents
        image = self.vae.decode(latents).sample

        image = (image / 2 + 0.5).clamp(0, 1)

        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
        image = image.cpu().permute(0, 2, 3, 1).float().numpy()

        if self.safety_checker is not None:
            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
                self.device
            )
            image, has_nsfw_concept = self.safety_checker(
                images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
            )
        else:
            has_nsfw_concept = None

        if output_type == "pil":
            image = self.numpy_to_pil(image)

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

    def generate_inputs(self, prompt_a, prompt_b, seed_a, seed_b, noise_shape, T, batch_size):
        embeds_a = self.embed_text(prompt_a)
        embeds_b = self.embed_text(prompt_b)

        latents_a = self.init_noise(seed_a, noise_shape)
        latents_b = self.init_noise(seed_b, noise_shape)

        batch_idx = 0
        embeds_batch, noise_batch = None, None
        for i, t in enumerate(T):
            embeds = torch.lerp(embeds_a, embeds_b, t)
            noise = slerp(float(t), latents_a, latents_b)

            embeds_batch = embeds if embeds_batch is None else torch.cat([embeds_batch, embeds])
            noise_batch = noise if noise_batch is None else torch.cat([noise_batch, noise])
            batch_is_ready = embeds_batch.shape[0] == batch_size or i + 1 == T.shape[0]
            if not batch_is_ready:
                continue
            yield batch_idx, embeds_batch, noise_batch
            batch_idx += 1
            del embeds_batch, noise_batch
            torch.cuda.empty_cache()
            embeds_batch, noise_batch = None, None

    def make_clip_frames(
        self,
        prompt_a: str,
        prompt_b: str,
        seed_a: int,
        seed_b: int,
        num_interpolation_steps: int = 5,
        save_path: Union[str, Path] = "outputs/",
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        eta: float = 0.0,
        height: int = 512,
        width: int = 512,
        upsample: bool = False,
        batch_size: int = 1,
        image_file_ext: str = ".png",
        T: np.ndarray = None,
        skip: int = 0,
    ):
        save_path = Path(save_path)
        save_path.mkdir(parents=True, exist_ok=True)

        T = T if T is not None else np.linspace(0.0, 1.0, num_interpolation_steps)
        if T.shape[0] != num_interpolation_steps:
            raise ValueError(f"Unexpected T shape, got {T.shape}, expected dim 0 to be {num_interpolation_steps}")

        if upsample:
            if getattr(self, "upsampler", None) is None:
                self.upsampler = RealESRGANModel.from_pretrained("nateraw/real-esrgan")
            self.upsampler.to(self.device)

        batch_generator = self.generate_inputs(
            prompt_a,
            prompt_b,
            seed_a,
            seed_b,
            (1, self.unet.in_channels, height // 8, width // 8),
            T[skip:],
            batch_size,
        )

        frame_index = skip
        for _, embeds_batch, noise_batch in batch_generator:
            with torch.autocast("cuda"):
                outputs = self(
                    latents=noise_batch,
                    text_embeddings=embeds_batch,
                    height=height,
                    width=width,
                    guidance_scale=guidance_scale,
                    eta=eta,
                    num_inference_steps=num_inference_steps,
                    output_type="pil" if not upsample else "numpy",
                )["images"]

                for image in outputs:
                    frame_filepath = save_path / (f"frame%06d{image_file_ext}" % frame_index)
                    image = image if not upsample else self.upsampler(image)
                    image.save(frame_filepath)
                    frame_index += 1

    def walk(
        self,
        prompts: Optional[List[str]] = None,
        seeds: Optional[List[int]] = None,
        num_interpolation_steps: Optional[Union[int, List[int]]] = 5,  # int or list of int
        output_dir: Optional[str] = "./dreams",
        name: Optional[str] = None,
        image_file_ext: Optional[str] = ".png",
        fps: Optional[int] = 30,
        num_inference_steps: Optional[int] = 50,
        guidance_scale: Optional[float] = 7.5,
        eta: Optional[float] = 0.0,
        height: Optional[int] = 512,
        width: Optional[int] = 512,
        upsample: Optional[bool] = False,
        batch_size: Optional[int] = 1,
        resume: Optional[bool] = False,
        audio_filepath: str = None,
        audio_start_sec: Optional[Union[int, float]] = None,
        margin: Optional[float] = 1.0,
        smooth: Optional[float] = 0.0,
    ):
        """Generate a video from a sequence of prompts and seeds. Optionally, add audio to the
        video to interpolate to the intensity of the audio.

        Args:
            prompts (Optional[List[str]], optional):
                list of text prompts. Defaults to None.
            seeds (Optional[List[int]], optional):
                list of random seeds corresponding to prompts. Defaults to None.
            num_interpolation_steps (Union[int, List[int]], *optional*):
                How many interpolation steps between each prompt. Defaults to None.
            output_dir (Optional[str], optional):
                Where to save the video. Defaults to './dreams'.
            name (Optional[str], optional):
                Name of the subdirectory of output_dir. Defaults to None.
            image_file_ext (Optional[str], *optional*, defaults to '.png'):
                The extension to use when writing video frames.
            fps (Optional[int], *optional*, defaults to 30):
                The frames per second in the resulting output videos.
            num_inference_steps (Optional[int], *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            guidance_scale (Optional[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 images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            eta (Optional[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.
            height (Optional[int], *optional*, defaults to 512):
                height of the images to generate.
            width (Optional[int], *optional*, defaults to 512):
                width of the images to generate.
            upsample (Optional[bool], *optional*, defaults to False):
                When True, upsamples images with realesrgan.
            batch_size (Optional[int], *optional*, defaults to 1):
                Number of images to generate at once.
            resume (Optional[bool], *optional*, defaults to False):
                When True, resumes from the last frame in the output directory based
                on available prompt config. Requires you to provide the `name` argument.
            audio_filepath (str, *optional*, defaults to None):
                Optional path to an audio file to influence the interpolation rate.
            audio_start_sec (Optional[Union[int, float]], *optional*, defaults to 0):
                Global start time of the provided audio_filepath.
            margin (Optional[float], *optional*, defaults to 1.0):
                Margin from librosa hpss to use for audio interpolation.
            smooth (Optional[float], *optional*, defaults to 0.0):
                Smoothness of the audio interpolation. 1.0 means linear interpolation.

        This function will create sub directories for each prompt and seed pair.

        For example, if you provide the following prompts and seeds:

        ```
        prompts = ['a dog', 'a cat', 'a bird']
        seeds = [1, 2, 3]
        num_interpolation_steps = 5
        output_dir = 'output_dir'
        name = 'name'
        fps = 5
        ```

        Then the following directories will be created:

        ```
        output_dir
        ├── name
        │   ├── name_000000
        │   │   ├── frame000000.png
        │   │   ├── ...
        │   │   ├── frame000004.png
        │   │   ├── name_000000.mp4
        │   ├── name_000001
        │   │   ├── frame000000.png
        │   │   ├── ...
        │   │   ├── frame000004.png
        │   │   ├── name_000001.mp4
        │   ├── ...
        │   ├── name.mp4
        |   |── prompt_config.json
        ```

        Returns:
            str: The resulting video filepath. This video includes all sub directories' video clips.
        """

        output_path = Path(output_dir)

        name = name or time.strftime("%Y%m%d-%H%M%S")
        save_path_root = output_path / name
        save_path_root.mkdir(parents=True, exist_ok=True)

        # Where the final video of all the clips combined will be saved
        output_filepath = save_path_root / f"{name}.mp4"

        # If using same number of interpolation steps between, we turn into list
        if not resume and isinstance(num_interpolation_steps, int):
            num_interpolation_steps = [num_interpolation_steps] * (len(prompts) - 1)

        if not resume:
            audio_start_sec = audio_start_sec or 0

        # Save/reload prompt config
        prompt_config_path = save_path_root / "prompt_config.json"
        if not resume:
            prompt_config_path.write_text(
                json.dumps(
                    dict(
                        prompts=prompts,
                        seeds=seeds,
                        num_interpolation_steps=num_interpolation_steps,
                        fps=fps,
                        num_inference_steps=num_inference_steps,
                        guidance_scale=guidance_scale,
                        eta=eta,
                        upsample=upsample,
                        height=height,
                        width=width,
                        audio_filepath=audio_filepath,
                        audio_start_sec=audio_start_sec,
                    ),
                    indent=2,
                    sort_keys=False,
                )
            )
        else:
            data = json.load(open(prompt_config_path))
            prompts = data["prompts"]
            seeds = data["seeds"]
            num_interpolation_steps = data["num_interpolation_steps"]
            fps = data["fps"]
            num_inference_steps = data["num_inference_steps"]
            guidance_scale = data["guidance_scale"]
            eta = data["eta"]
            upsample = data["upsample"]
            height = data["height"]
            width = data["width"]
            audio_filepath = data["audio_filepath"]
            audio_start_sec = data["audio_start_sec"]

        for i, (prompt_a, prompt_b, seed_a, seed_b, num_step) in enumerate(
            zip(prompts, prompts[1:], seeds, seeds[1:], num_interpolation_steps)
        ):
            # {name}_000000 / {name}_000001 / ...
            save_path = save_path_root / f"{name}_{i:06d}"

            # Where the individual clips will be saved
            step_output_filepath = save_path / f"{name}_{i:06d}.mp4"

            # Determine if we need to resume from a previous run
            skip = 0
            if resume:
                if step_output_filepath.exists():
                    print(f"Skipping {save_path} because frames already exist")
                    continue

                existing_frames = sorted(save_path.glob(f"*{image_file_ext}"))
                if existing_frames:
                    skip = int(existing_frames[-1].stem[-6:]) + 1
                    if skip + 1 >= num_step:
                        print(f"Skipping {save_path} because frames already exist")
                        continue
                    print(f"Resuming {save_path.name} from frame {skip}")

            audio_offset = audio_start_sec + sum(num_interpolation_steps[:i]) / fps
            audio_duration = num_step / fps

            self.make_clip_frames(
                prompt_a,
                prompt_b,
                seed_a,
                seed_b,
                num_interpolation_steps=num_step,
                save_path=save_path,
                num_inference_steps=num_inference_steps,
                guidance_scale=guidance_scale,
                eta=eta,
                height=height,
                width=width,
                upsample=upsample,
                batch_size=batch_size,
                skip=skip,
                T=get_timesteps_arr(
                    audio_filepath,
                    offset=audio_offset,
                    duration=audio_duration,
                    fps=fps,
                    margin=margin,
                    smooth=smooth,
                )
                if audio_filepath
                else None,
            )
            make_video_pyav(
                save_path,
                audio_filepath=audio_filepath,
                fps=fps,
                output_filepath=step_output_filepath,
                glob_pattern=f"*{image_file_ext}",
                audio_offset=audio_offset,
                audio_duration=audio_duration,
                sr=44100,
            )

        return make_video_pyav(
            save_path_root,
            audio_filepath=audio_filepath,
            fps=fps,
            audio_offset=audio_start_sec,
            audio_duration=sum(num_interpolation_steps) / fps,
            output_filepath=output_filepath,
            glob_pattern=f"**/*{image_file_ext}",
            sr=44100,
        )

    def embed_text(self, text):
        """Helper to embed some text"""
        with torch.autocast("cuda"):
            text_input = self.tokenizer(
                text,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            with torch.no_grad():
                embed = self.text_encoder(text_input.input_ids.to(self.device))[0]
        return embed

    def init_noise(self, seed, noise_shape):
        """Helper to initialize noise"""
        # randn does not exist on mps, so we create noise on CPU here and move it to the device after initialization
        if self.device.type == "mps":
            noise = torch.randn(
                noise_shape,
                device='cpu',
                generator=torch.Generator(device='cpu').manual_seed(seed),
            ).to(self.device)
        else:
            noise = torch.randn(
                noise_shape,
                device=self.device,
                generator=torch.Generator(device=self.device).manual_seed(seed),
            )
        return noise

    @classmethod
    def from_pretrained(cls, *args, tiled=False, **kwargs):
        """Same as diffusers `from_pretrained` but with tiled option, which makes images tilable"""
        if tiled:

            def patch_conv(**patch):
                cls = nn.Conv2d
                init = cls.__init__

                def __init__(self, *args, **kwargs):
                    return init(self, *args, **kwargs, **patch)

                cls.__init__ = __init__

            patch_conv(padding_mode="circular")

        pipeline = super().from_pretrained(*args, **kwargs)
        pipeline.tiled = tiled
        return pipeline