adversarial-flow-distillation

Proposes Adversarial Flow Distillation (AFD) for distilling heterogeneous black-box video generation models into autoregressive students, using on-policy feedback and forward-process flow-matching updates.