utils/ddim_utils.py

import numpy as np
from typing import Union

import torch

from tqdm import tqdm
from diffusers import DDIMScheduler


# 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


def inverse_video(pipe, latents, num_steps):
    ddim_inv_scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
    ddim_inv_scheduler.set_timesteps(num_steps)

    ddim_inv_latent = ddim_inversion(
        pipe, ddim_inv_scheduler, video_latent=latents.to(pipe.device),
        num_inv_steps=num_steps, prompt="")[-1]
    return ddim_inv_latent