Tag
This paper proposes NormGuard, a norm-budget regularizer that suppresses reward-irrelevant velocity norm inflation during RL post-training of flow-matching generative models, improving perceptual image quality without sacrificing reward.