Reproduce “Getting Space Right: Agentic Rewards for Image Generation” 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
Xiaofeng Tan, Jiaxin Lin, Haobo Yuan, Tiancheng Shen, Zihan Su, Rui Chen, Hongsong Wang, Ming-Hsuan Yang
SpatialCritic judges spatial image prompts by breaking them into claims and gathering evidence from vision tools; curriculum training schedules generation from component skills toward compositions. On SpaRW-Eval, aligned 8B SpatialCritic reaches 95.1% accuracy across 5,572 preference pairs. On BAGEL, curriculum raises Flow-GRPO with SpatialCritic’s seven-benchmark mean from 83.7 to 86.5, but does not establish general OOD robustness; the prompt schedules differ.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.