AI summaryⓘ
The authors address the challenge of teaching one video generator (student) to mimic another complex, black-box video generator (teacher) when only completed videos from the teacher are available, and the student must learn sequentially on its own outputs. They propose Adversarial Flow Distillation (AFD), which uses the teacher's videos and the student's rollouts with the same prompts to estimate differences and guide the student’s training in a way that doesn't require complex internal teacher data. Their method improves the student's ability to generate realistic motion and physics in videos while maintaining overall video quality. The authors show that AFD works across different student models and only needs access to clean teacher videos and student outputs.
Autoregressive video generatorsBlack-box distillationCausal student modelsAdversarial learningBradley-Terry discriminatorOn-policy trainingFlow-matchingVideo generationTeacher-student learningForward-process credit assignment
Authors
Yang Luo, Shengju Qian, Xiaohang Tang, Zirui Zhu, Yong Liu, Xin Wang, Yang You
Abstract
Autoregressive video generators are attractive for streaming, long-horizon, and interactive applications, but distilling strong black-box teachers into causal students remains difficult. The student must learn under its own rollout distribution, whereas practical teachers may expose only prompt-conditioned completed videos and may differ in architecture, capacity, temporal design, and sampling schedule. This interface makes supervised fine-tuning off-policy, score-based distillation inapplicable, and direct adversarial imitation too sparse for denoising-time credit assignment. We propose Adversarial Flow Distillation (AFD), an on-policy framework for heterogeneous black-box video distillation. AFD queries the teacher and rolls out the current student on the same prompts, trains a prompt-paired Bradley-Terry discriminator to estimate clean-sample teacher-student discrepancy, and converts the resulting on-policy advantage into forward-process flow-matching updates on the student's own noised states. Thus, AFD provides dense velocity-field supervision while requiring no teacher scores, latents, denoising trajectories, step alignment, or reverse-chain reinforcement learning. Experiments across two causal AR student families show that AFD consistently improves motion- and physics-sensitive generation while preserving general video quality, and ablations validate the importance of adaptive on-policy feedback and forward-process credit assignment. The method requires only clean teacher videos and student rollouts, providing a practical route for distilling proprietary or heterogeneous video generators into efficient autoregressive students.