ReSFlow rectifies categorical flows on statistical manifolds for one-step generation

jiqizhixin · x · 2026-11-22

Researchers from Inner Mongolia University, Shanghai University of Finance and Economics and CMU introduce ReSFlow (arXiv preprint). Discrete generators often spend many integration steps following curved paths between distributions; ReSFlow rectifies categorical flows on the statistical manifold, respecting its geometry so learned routes approach geodesics — changing the transport problem itself so a single sampling step suffices. Experiments span image, text and biological generation, reporting improved one-step quality plus geometric analysis of trajectories.

Original post →

More from Research

Research channel →