Scaling Online Optimal Transport Minibatches for Flow Matching
A scalable online entropic OT method stabilizes flow-matching training with larger source-target couplings.
Flow-matching stability depends on source-target pairing, and independent pairing adds substantial stochasticity. The paper introduces a scalable online entropic optimal-transport approximation that builds much larger couplings without computing full OT every iteration. It proves strict reductions in full stochastic-gradient variance and integration error versus independent pairing for both unregularized and entropic OT, with improvements growing as the minibatch grows. Single-cell gene interpolation and image-flow experiments show more stable training and better generation.
- Online entropic OT supports much larger source-target couplings.
- Proofs show lower gradient variance and integration error than independent pairing.
- Gains increase as the online coupling minibatch grows.
- Single-cell and image-flow experiments show stabler training and better generation.
From the original article
1 min read at huggingface.co
A central challenge in scaling flow models is maintaining stability in training and integration. In flow matching, this stability depends critically on how source samples are paired with data. Independent random pairing introduces substantial stochasticity, motivating minibatch optimal transport (OT) to construct online source-target couplings that reduce this randomness (Tong et al.
By Peter Chen, Sergey Levine · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.