Reproduce “Scout-Refine-Correct: Teaching Flow Models to Refine Their Own Samples” 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
Sergey Karpukhin, Aibek Alanov
Scout-Refine-Correct (SRC) trains a flow-matching image generator to refine its own samples by previewing future sampling steps and conditioning the next step on a clean-image estimate. Training also corrects off-path rollout states. On class-conditional ImageNet-256, it reports FID 7.26 at NFE 550, versus 14.52 for the base model at 50 steps; this comparison uses different training budgets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.