SC-Flow unifies velocity and endpoint prediction in rectified flow models

Self-Consistent Flow (SC-Flow) introduces a lightweight consistency loss to jointly train networks for both velocity and data endpoint prediction. This approach combines the training stability of endpoint prediction with the stable sampling dynamics of velocity prediction near the data manifold.