CAR Guidance: Staying On-Manifold under Compositional Rewards
We introduce a plug-and-play module that corrects off-manifold drift when guiding flow models with multiple rewards at inference time.
Xuehui Yu