Reverse-time diffusion processes in discrete time

Abstract

Generative AI relies on finding reverse models for linear discrete time forward diffusions with non-Gaussian initial states, but uses indirect methods for reversal as there is no theory for direct reversal in discrete time. We give sufficient conditions that guarantee the existence of a reverse time diffusion and give a method of meeting them. We also give a necessary and sufficient condition for the reverse model to be input-affine. Show more