New OT-NFM framework enables one-step image generation by learning direct transport maps instead of time-dependent flows, eliminating the need for dozens of network evaluations.

Optimal Transport Neural Flow Matching (OT-NFM) uses neural flows to parameterize transport maps, achieving true one-step generation with a single forward pass instead of tens to hundreds of evaluations

Researchers identified and solved the mean collapse problem in naive flow-map training by proving that consistent optimal transport couplings are necessary for non-degenerate learning

Scalable minibatch and online coupling strategies enable practical implementation of optimal transport pairings for training

Experiments on MNIST and CIFAR-10 datasets show competitive sample quality while dramatically reducing inference cost to a single network evaluation