ReSFlow: One-Step Categorical Generation on Statistical Manifolds
October 12, 2026
ReSFlow enables one-step generation for discrete data by learning geodesic trajectories on statistical manifolds. By rectifying the flow to follow the intrinsic Riemannian geometry, the model avoids the computational costs of multi-step numerical integration.
HOW THIS AFFECTS YOU
●
researcherYou can achieve single-step inference for discrete generative models by utilizing rectified statistical flows.