Reproduce “Few-Step Generation via Data-Space Iteration” 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
Shanchuan Lin, Yansong Peng, Fu-Yun Wang, Haoqi Fan
The authors introduce data-space iteration for few-step image generation, passing a draft between calls while retaining the original noise input instead of following a preset generation-time noise schedule. The method keeps DMD self-rollout training, which still uses noise time. On guided class-conditional ImageNet 256×256 at four NFE, it reports FID 2.29 versus 2.37 for the best of three flow-map schedules.
Key insights
- Changing the rollout space removes flow-time discretization from the sampling path, though the number of evaluations and refinement-index conditioning remain design choices.
- On the reported ImageNet experiments, data-space iteration generally improved FID over uniform-schedule flow-renoise and was competitive with schedule-searched flow-map generation.
- FID gains did not imply improvement on every metric: data-space DMD had higher recall but lower precision than DMD in the reported four-evaluation comparisons.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.