Abstract
Diffusion models have revolutionized generative AI. Conceptually, these methods define a mechanism for transporting a noise distribution to a data distribution. Recent work has extended this framework to define transports between arbitrary distributions, significantly broadening the potential of these diffusion models. However, existing methods often fail to approximate the optimal transport between these distributions. In this presentation, we will show how current methodologies can be modified to obtain Schrödinger bridges—an entropy-regularized variant of the optimal dynamic transport.
