OPSD-V
On-policy self-distillation for post-training few-step autoregressive video generators, improving long-horizon quality and motion dynamics while preserving the original sampler.
OPSD-V is an on-policy self-distillation framework for post-training few-step autoregressive video generators. It continues training the student on its inference-time rollout states while using a cleaner teacher context to reduce long-horizon degradation and strengthen motion dynamics.
The project targets long autoregressive video generation after few-step distillation, preserving the original sampler while improving visual stability and sustained motion.
Links: Project page · arXiv · GitHub