On-Policy Adversarial Flow Distillation for Autoregressive Video Generation

Researchers introduced Adversarial Flow Distillation (AFD) to train autoregressive video models from black-box teachers using on-policy feedback.

AFD compares teacher and student videos on the same prompts, then learns a discrepancy signal with a prompt-paired Bradley-Terry discriminator.

That signal is converted into dense flow-matching updates on the student’s noised states, enabling practical distillation without teacher scores or full trajectories.

In tests on two causal autoregressive student families, AFD improved motion and physics quality while preserving overall video quality.

The method could transfer capabilities from proprietary or architecturally different video generators to efficient causal models.