Abstract

Flow matching models learn a (possibly stochastic) mapping between source and target distributions. Common paradigms include diffusion models, score matching models, and continuous normalizing flows. In this talk I will discuss how incorporating ideas from optimal transport can lead to improved training and inference of flow matching models by reducing variance of the objective and creating “straighter” paths that can be integrated in fewer steps. I will then show how these methods can be used in the domains of image generation, modelling cell trajectories, and protein generation.