Flow Matching from Non-Gaussian Sources: A Theory of Efficient Sampling
The paper develops a nonasymptotic convergence theory for Euler discretization of flow matching with linear interpolation when the source distribution is non-Gaussian. Under assumptions including an everywhere positive continuously differentiable density, finite total variance, an L-Lipschitz score, and a mild target second-moment condition, it bounds total variation error by epsilon in O-tilde(L d TotVar_base / epsilon) iterations. The assumptions cover multivariate Student's t sources with more than two degrees of freedom and finite Gaussian mixtures with shared covariance, while allowing target distributions without smooth densities or bounded support.
It provides finite-step sampling guarantees for flow matching beyond Gaussian source assumptions, useful for researchers designing generative models with heavy-tailed or multimodal priors.
Comments & Questions (0)
No comments yet
Be the first to comment!