💡 The paper proposes a method called On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators, or OPSD-V, which aims to improve the quality of videos generated by few-step autoregressive video diffusion models. The problem with existing models is that they can produce long videos with low latency, but the quality of the video degrades over time due to error accumulation and weakened motion dynamics.

To address this issue, OPSD-V introduces real long-video data as temporal context during training, providing dense trajectory-level supervision to improve visual quality and motion dynamics. The method works by having a student model follow the exact inference-time rollout, generating each chunk of the video conditioned on its own previously generated cache. In parallel, a teacher model is evaluated at the same denoising states, but uses a cleaner temporal cache that can be replaced by real-video context. This provides corrective targets under on-policy cache dynamics, without changing the inference mechanism.

The results show that OPSD-V consistently improves the visual quality, motion dynamics, and VBenchLong scores of the generated videos. The method is applied to representative few-step autoregressive video models, including Self-Forcing and LongLive, and the experiments demonstrate significant improvements. A user study with 10 participants also shows that OPSD-V is preferred over the base models in 66 percent of overall-preference judgments, and 82.5 percent excluding ties. Overall, the paper contributes a novel method for improving the quality of videos generated by few-step autoregressive video diffusion models, without altering the inference mechanism.