Reproduce “ROAD: Modern Adversarial Training for Simple One-Step Distillation at Scale” 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

Xiaoou Li, Shan You, Tao Huang

ROAD trains a one-step text-to-image generator from a pretrained model using real images and an adversarial critic that reads frozen visual tokens and prompt tokens. It periodically refreshes the critic’s random feature mixing and uses a Jacobian penalty on its dense logits. On SDXL, it reports FID 11.03 versus 15.23 for SiDa; this result does not establish universal superiority.

Summary by alphaXiv.

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