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

2026-07-09

AI Summary

This work addresses the long-term degradation in few-step autoregressive video generation—caused by error accumulation and weakened motion dynamics—by introducing an in-policy self-distillation paradigm. Without altering the original few-step inference pathway, the method incorporates dense trajectory-level supervision derived from real long videos for the first time. It employs a teacher–student architecture to refine denoising targets and integrates KV caching with an autoregressive-consistent temporal cache replacement strategy to preserve dynamic coherence. Experiments demonstrate substantial improvements in visual quality, motion dynamics, and VBenchLong scores, achieving a user preference rate of 66.0% (82.5% when ties are excluded).

Abstract

We propose OPSD-V, an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models. Existing few-step AR video generators can produce long videos with low latency, but still suffer from error accumulation and weakened motion dynamics during long autoregressive rollout. OPSD-V reduces long-horizon degradation while preserving the original few-step inference path. The key idea is to introduce real long-video data as temporal context during training and use it to provide dense trajectory-level supervision. Specifically, the student follows the exact inference-time rollout, generating each chunk conditioned on its own previously generated KV cache. In parallel, the teacher is evaluated at the same student-visited denoising states, but uses a cleaner AR-consistent temporal cache in which older history can be replaced by real-video context. This provides dense denoising-level corrective targets under on-policy AR cache dynamics, without changing the sampler, number of denoising steps, or inference-time cache mechanism. We apply OPSD-V to representative few-step AR video models, including Self-Forcing and LongLive. Experiments show consistent improvements in visual quality, motion dynamics, and VBenchLong scores. A user study with 10 participants comparing 20 video pairs shows that OPSD-V is preferred over the base models in 66.0% of overall-preference judgments (82.5% excluding ties).

Problem

• autoregressive video generation
• error accumulation
• motion dynamics
• long-horizon degradation
• few-step diffusion

Innovation

• on-policy self-distillation
• few-step autoregressive video generation
• temporal context distillation
• KV cache consistency
• long-horizon video synthesis