multimodalart/EchoMimic-zero

1import torch2import numpy as np3 4def get_alpha(alphas_cumprod, timestep):5    timestep_lt_zero_mask = torch.lt(timestep, 0).to(alphas_cumprod.dtype)6    normal_alpha = alphas_cumprod[torch.clip(timestep, 0)]7    one_alpha = torch.ones_like(normal_alpha).to(normal_alpha.dtype).to(normal_alpha.dtype) 8    return normal_alpha * (1 - timestep_lt_zero_mask) + one_alpha * timestep_lt_zero_mask9 10def psuedo_velocity_wrt_noisy_and_timestep(noisy_images, noisy_images_pre, alphas_cumprod, timestep, timestep_prev):11    alpha_prod_t = get_alpha(alphas_cumprod, timestep).view(-1, 1, 1, 1, 1).detach()12    beta_prod_t = 1 - alpha_prod_t13    alpha_prod_t_prev = get_alpha(alphas_cumprod, timestep_prev).view(-1, 1, 1, 1, 1).detach()14    beta_prod_t_prev = 1 - alpha_prod_t_prev15 16    a_s = (alpha_prod_t_prev ** (0.5)).to(noisy_images.dtype)17    a_t = (alpha_prod_t ** (0.5)).to(noisy_images.dtype)18    b_s = (beta_prod_t_prev ** (0.5)).to(noisy_images.dtype)19    b_t = (beta_prod_t ** (0.5)).to(noisy_images.dtype)20 21    psuedo_velocity = (noisy_images_pre - (22        a_s * a_t + b_s * b_t23    ) * noisy_images) / (24        b_s * a_t -  a_s * b_t25    )26 27    return psuedo_velocity28 29def origin_by_velocity_and_sample(velocity, noisy_images, alphas_cumprod, timestep):30    alpha_prod_t = get_alpha(alphas_cumprod, timestep).view(-1, 1, 1, 1, 1).detach()31    beta_prod_t = 1 - alpha_prod_t32    a_t = (alpha_prod_t ** (0.5)).to(noisy_images.dtype)33    b_t = (beta_prod_t ** (0.5)).to(noisy_images.dtype)34 35    pred_original_sample = a_t * noisy_images - b_t * velocity36    return pred_original_sample37