Reproduce “SIFR: Stable Signed Flow Reinforcement” 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
Runlong Liao, Xingqian Xu, Lizhang Chen, Baiyu Su, Yufan Huang, Zhonghao Wang, Akash Gokul, Ming-Yu Liu and 2 more
SIFR fine-tunes text-to-image generators using reward scores, weighting above-average samples positively and below-average samples negatively. It limits average squared velocity-field drift from a reference, adapting the penalty to measured drift. On Cosmos 3 Super Text2Image, it reaches the baselines’ held-out PickScore plateau about four times sooner in training rounds and reports 25.36 after 2,000 rounds; PickScore is a learned score, not direct human preference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.