Abstract

Over the several recent years, there has been a boom in development of flowmatching methods for generative modeling. One intriguing property pursued bythe community is the ability to learn flows with straight trajectories whichrealize the optimal transport (OT) displacements. Straightness is crucial forfast integration of the learned flow's paths. Unfortunately, most existing flowstraightening methods are based on non-trivial iterative procedures whichaccumulate the error during training or exploit heuristic minibatch OTapproximations. To address this issue, we develop a novel optimal flow matchingapproach which recovers the straight OT displacement for the quadratic cost injust one flow matching step.

 

Quick Read (beta)