Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL
Junyi Wu, Weijian Luo, Haoyang Zheng, Ruizhe Zhang, Guang Lin, in NeurIPS 2026, 2026.
Abstract: Diff-Instruct with Diffused Reward (DIDR) aligns one-step generators by distributing reward guidance across diffusion noise levels. Derived from integral KL minimization, the framework uses a practical diffused reward estimator to improve preference alignment while preserving image quality, without requiring training data.