Importance-sample timesteps for a batch. :param batch_size: the number of timesteps. :param device: the torch device to save

raw latents -> [0, 1]

Args: x1: shape (bs, *dim), represents the target minibatch (data) x0: shape (bs, *dim), represents the source miniba

Compute training losses for a single timestep. :param model: the model to evaluate loss on. :param x_start: the [N x C x ...]

extract data

Unnormalize depth map from [-1, 1] to actual depth values. Args: depth: Torch tensor depth map in range [-1, 1]. q_min: Min q

[0, 1] -> raw latents

Update the reweighting using losses from a model. Call this method from each rank with a batch of timesteps and the correspon

SD3 & Meta Movie Gen show that val loss correlates well with human quality. They compute the loss in equidistant segments in (0, 1) t