Reproduce “Look Back to Move Forward: Learning to Revisit for Agentic Compositional 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
Zixu Wang, Xiaoran Zhao, Yahui Liu, Feng-ao Wang, Junhao Xiao, You Shu, Bing Liu, Yiting Liu and 3 more
GenRevisit plans multi-step text-to-image generation by choosing both an image from the interaction history and an edit instruction, while retaining earlier candidates. In matched-budget continuations from 200 unsolved teacher-trajectory prefixes, joint selection averages 55.2 GM, versus 52.4 for best-image-only and 50.2 for latest-image-only. This supports joint choice in that setup, not a guarantee for learned policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.