Reproduce “Dancing Diffusion: Learning Continuous Programs for Image Generation and Editing” 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
Cui Enhe, Jinyue Yang, Bojian Yin, Guoqi Li
Dancing Diffusion lets users redirect a generation run by saving its sampler state and changing conditions for the remaining denoising steps. A compiler produces continuous stage-specific conditions, while training feedback rewards requested changes and retained content. In SANA tests, prefix reuse reduced background LPIPS from 0.321 to 0.227 at branch 9, but preservation weakened with later branching and edits remained imperfect.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.