OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators

OPSD-V implements on-policy self-distillation to mitigate error accumulation in few-step autoregressive video diffusion models. The method provides dense trajectory-level supervision by conditioning student generation on previously generated KV caches while using a teacher model to supervise denoising states. This approach maintains motion dynamics during long-horizon rollouts without altering the inference path.