Reproduce “Correcting Flows with Marginal Matching” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Meital Bojan, Nadav Bojan Sellam, Ev Zisselman, Aviv Tamar
Marginal Matching improves samples from a frozen flow-matching image generator by training an auxiliary corrector to move samples toward intermediate distributions traced by running the flow backward from real images and continuing the forward flow. On unconditional CIFAR-10, score correction lowers FID from 4.34 to 3.38, with evaluations rising from 40 to 43; a measured quality gain, not evidence of faster generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.