Tag
This paper proposes Discriminator-Guided Reinforcement Learning (DRL) to correct alignment issues in score- and flow-matching models by using a pretrained representation space discriminator as an optimal reward signal, significantly improving visual fidelity and semantic quality without human preferences.