Reproduce “Smoother Flow Matching via Contrastive Trajectory Repulsion” 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
Ziqi Jiang, Zhenqi He, Long Chen
CoFlow changes intermediate Flow Matching paths during training by nudging each sample away from a nearby negative path, then trains on the adjusted path and its matching velocity target. On class-conditional ImageNet 256×256 with SiT-XL/2, it lowers FID from 38.50 to 29.65 versus standard Flow Matching at 20 sampling steps. This result is limited to the evaluated setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.