The authors addressed the problem of aligning few-step diffusion generators with external reward functions, such as text recognizers or human preference models.
Two classifier heads align few-step diffusion with external rewards.
DualHeads adapts a few-step generator using only black-box reward evaluations and reference samples, achieving task rewards comparable to 40-step RL methods.
Research Lab
Yuanzhi Zhu · Eleftherios Tsonis · Simo Alami Chehboune · Xi Wang · Stéphane Lathuilière · Vicky Kalogeiton
École Polytechnique · CNRS · Inria at Univ. Grenoble Alpes
Research Digest··3 min read
The authors propose DualHeads, which uses two lightweight binary classifiers on a frozen diffusion backbone to tilt a few-step generator toward higher rewards.
§