Reproduce “DualHeads: Reward-tilted few-step diffusion made simple” 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
Yuanzhi Zhu, Eleftherios Tsonis, Simo Alami Chehboune, Xi Wang, Stéphane Lathuilière, Vicky Kalogeiton
DualHeads adapts few-step image generators to black-box rewards using two classifiers trained on a frozen diffusion backbone: one learns reward-ranked samples, while the other distinguishes generated outputs from prompt-paired references. Training uses both signals; inference removes the classifiers. On SD3.5-M, the four-step model reaches 0.90 GenEval accuracy versus 0.63 for the 80-NFE base, showing reward gain rather than speedup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.