Connecting SPDE to SGMs

Abstract

This paper investigates a Stochastic Partial Differential Equation (SPDE) derived from the Fokker-Planck equation associated with Score-based Generative Models. We modify the standard Fokker-Planck equation to better represent practical SGMs and introduce noise to mitigate potential discretization issues. The primary goal is to prove the existence and uniqueness of solutions for this SPDE. This aspect requires careful consideration due to the time-dependent operator and unbounded domain. To overcome these hurdles, we employ a variational approach and introduce a novel space inspired by Ornstein-Uhlenbeck operators. By demonstrating that this space and its subspace satisfy the necessary assumptions, they establish the existence of a solution for the given SPDE.