Scalable Online Optimal Transport for Flow Matching
October 12, 2026
This work introduces a scalable online entropic optimal transport approach to improve training stability in flow matching models. The method constructs better source-target couplings for large, high-dimensional datasets without the computational cost of full minibatch OT.
HOW THIS AFFECTS YOU
●
researcherYou can improve the stability and convergence of flow models when scaling to larger datasets.