Immiscible Diffusion: Accelerating Diffusion Training with Noise Assignment

In this paper, we point out that suboptimal noise-data mapping leads to slow training of diffusion models.