Convergence of score-based generative modeling

Score-based generative models and diffusion probabilistic models have exhibited exceptional performance in various applications. A natural question that arises is whether the distribution generated by the model is closely aligned with the given data distribution. In this talk, we will explore an upper bound of the Wasserstein distance between these two distributions. Based on the theory of optimal transport, we guarantee that the framework can approximate data distributions in the space of probability measures equipped with the Wasserstein distance.

This talk is based on joint work with Ying Fan and Kangwook Lee (UW-Madison).