Reproduce “Test-Time Scaling of Text-to-Image Flow Matching Models with Noise-Rejuvenated Sequential Monte Carlo” 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
Wenyi Lian, Zheng Zhao, Thomas B. Schön
NR-SMC steers text-to-image flow-matching models toward rewards using sequential Monte Carlo without reward gradients or additional training. At resampling, it replaces source noise while holding an estimated endpoint fixed, then continues with the deterministic solver; look-ahead informs selection. On SD3, GenEval rises from 0.718 to 0.748 and ImageReward from 1.081 to 1.268, though endpoints are approximate and inference costs increase.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.