animatediff/utils/util.py

import os
import imageio
import numpy as np
from typing import Union

import torch
import torchvision

from tqdm import tqdm
from einops import rearrange

import PIL.Image
import PIL.ImageOps
from packaging import version
from PIL import Image

def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=6, fps=8):
    videos = rearrange(videos, "b c t h w -> t b c h w")
    outputs = []
    for x in videos:
        x = torchvision.utils.make_grid(x, nrow=n_rows)
        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
        if rescale:
            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
        x = (x * 255).numpy().astype(np.uint8)
        outputs.append(x)

    os.makedirs(os.path.dirname(path), exist_ok=True)
    imageio.mimsave(path, outputs, fps=fps)


# DDIM Inversion
@torch.no_grad()
def init_prompt(prompt, pipeline):
    uncond_input = pipeline.tokenizer(
        [""], padding="max_length", max_length=pipeline.tokenizer.model_max_length,
        return_tensors="pt"
    )
    uncond_embeddings = pipeline.text_encoder(uncond_input.input_ids.to(pipeline.device))[0]
    text_input = pipeline.tokenizer(
        [prompt],
        padding="max_length",
        max_length=pipeline.tokenizer.model_max_length,
        truncation=True,
        return_tensors="pt",
    )
    text_embeddings = pipeline.text_encoder(text_input.input_ids.to(pipeline.device))[0]
    context = torch.cat([uncond_embeddings, text_embeddings])

    return context


def next_step(model_output: Union[torch.FloatTensor, np.ndarray], timestep: int,
              sample: Union[torch.FloatTensor, np.ndarray], ddim_scheduler):
    timestep, next_timestep = min(
        timestep - ddim_scheduler.config.num_train_timesteps // ddim_scheduler.num_inference_steps, 999), timestep
    alpha_prod_t = ddim_scheduler.alphas_cumprod[timestep] if timestep >= 0 else ddim_scheduler.final_alpha_cumprod
    alpha_prod_t_next = ddim_scheduler.alphas_cumprod[next_timestep]
    beta_prod_t = 1 - alpha_prod_t
    next_original_sample = (sample - beta_prod_t ** 0.5 * model_output) / alpha_prod_t ** 0.5
    next_sample_direction = (1 - alpha_prod_t_next) ** 0.5 * model_output
    next_sample = alpha_prod_t_next ** 0.5 * next_original_sample + next_sample_direction
    return next_sample


def get_noise_pred_single(latents, t, context, unet):
    noise_pred = unet(latents, t, encoder_hidden_states=context)["sample"]
    return noise_pred


@torch.no_grad()
def ddim_loop(pipeline, ddim_scheduler, latent, num_inv_steps, prompt):
    context = init_prompt(prompt, pipeline)
    uncond_embeddings, cond_embeddings = context.chunk(2)
    all_latent = [latent]
    latent = latent.clone().detach()
    for i in tqdm(range(num_inv_steps)):
        t = ddim_scheduler.timesteps[len(ddim_scheduler.timesteps) - i - 1]
        noise_pred = get_noise_pred_single(latent, t, cond_embeddings, pipeline.unet)
        latent = next_step(noise_pred, t, latent, ddim_scheduler)
        all_latent.append(latent)
    return all_latent


@torch.no_grad()
def ddim_inversion(pipeline, ddim_scheduler, video_latent, num_inv_steps, prompt=""):
    ddim_latents = ddim_loop(pipeline, ddim_scheduler, video_latent, num_inv_steps, prompt)
    return ddim_latents

if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"):
    PIL_INTERPOLATION = {
        "linear": PIL.Image.Resampling.BILINEAR,
        "bilinear": PIL.Image.Resampling.BILINEAR,
        "bicubic": PIL.Image.Resampling.BICUBIC,
        "lanczos": PIL.Image.Resampling.LANCZOS,
        "nearest": PIL.Image.Resampling.NEAREST,
    }
else:
    PIL_INTERPOLATION = {
        "linear": PIL.Image.LINEAR,
        "bilinear": PIL.Image.BILINEAR,
        "bicubic": PIL.Image.BICUBIC,
        "lanczos": PIL.Image.LANCZOS,
        "nearest": PIL.Image.NEAREST,
    }


def pt_to_pil(images):
    """
    Convert a torch image to a PIL image.
    """
    images = (images / 2 + 0.5).clamp(0, 1)
    images = images.cpu().permute(0, 2, 3, 1).float().numpy()
    images = numpy_to_pil(images)
    return images


def numpy_to_pil(images):
    """
    Convert a numpy image or a batch of images to a PIL image.
    """
    if images.ndim == 3:
        images = images[None, ...]
    images = (images * 255).round().astype("uint8")
    if images.shape[-1] == 1:
        # special case for grayscale (single channel) images
        pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
    else:
        pil_images = [Image.fromarray(image) for image in images]


def preprocess_image(image):
    if isinstance(image, torch.Tensor):
        return image
    elif isinstance(image, PIL.Image.Image):
        image = [image]

    if isinstance(image[0], PIL.Image.Image):
        w, h = image[0].size
        w, h = map(lambda x: x - x % 8, (w, h))  # resize to integer multiple of 8

        image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
        image = np.concatenate(image, axis=0)
        image = np.array(image).astype(np.float32) / 255.0
        image = image.transpose(0, 3, 1, 2)
        image = 2.0 * image - 1.0
        image = torch.from_numpy(image)
    elif isinstance(image[0], torch.Tensor):
        image = torch.cat(image, dim=0)
    return image