Reproduce “Learning to Sample Effective and Diverse Prompts for Text-to-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

Taeyoung Yun, Dinghuai Zhang, Jinkyoo Park, Ling Pan

PAG adapts text-to-image prompts by fine-tuning a language model to sample promising rewrites while leaving the image generator unchanged. It combines flow-function resets, reward-prioritized replay, and intermediate signals from partial prompts. On ChatGPT prompts, it reports best-of-16 reward 0.88 and prompt diversity 0.32, versus vanilla GFlowNets at 0.63 and 0.10; these are automatic-reward and text-embedding measures.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.