Reproduce “StepX-OPD: Rethinking Multi-Teacher On-Policy Distillation in Diffusion Models” 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
Chenru Wang, Xuanyang Zhang, Yucheng Han, Wenbo Nie, Wei Cheng, Xianfang Zeng, Gang YU, Chi Zhang
StepX-OPD trains a four-step text-to-image student to integrate specialist capabilities and shorten sampling in one run. On-policy means supervision uses states the student generates; step distillation trains generation in fewer sampling steps. Its Average, the mean of GenEval, OCR, DeQA/5, and PickScore/26, exceeds tested four-step baselines on SD3.5-Medium and FLUX.1-dev, but not every metric or full-step model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.