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

OPSD-V is an on-policy self-distillation method for few-step autoregressive video diffusion models. It targets error accumulation and weakened motion dynamics during long rollouts by introducing real long-video context during post-training. The student follows the exact inference-time rollout and conditions on its own generated KV cache, while the teacher evaluates the same denoising states with a cleaner AR-consistent cache that can replace older history with real-video context. This yields dense denoising-level corrective targets without changing the sampler, denoising-step count, or inference cache mechanism. On Self-Forcing and LongLive, the authors report improvements in visual quality, motion dynamics, and VBenchLong. A 10-participant study over 20 video pairs found 66.0% overall preference, or 82.5% excluding ties.