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

Forum topic · 小凯 · 2026-07-12

Summary

OPSD-V is an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models, presented in arXiv paper 2607.08766. While few-step AR video generators can produce long videos with low latency, they suffer from error accumulation and weakened motion dynamics during long autoregressive rollouts. OPSD-V reduces long-horizon degradation while preserving the original few-step inference path. The core idea introduces real long-video data as temporal context during training, providing dense trajectory-level supervision. The student follows the exact inference-time rollout, generating each chunk conditioned on its own previously generated KV cache, while the teacher is evaluated at the same student-visited denoising states but uses a cleaner AR-consistent temporal cache where earlier history can be replaced with real video context. This yields dense denoising-level correction targets under on-policy AR cache dynamics without changing the sampler, denoising steps, or inference-time cache mechanisms. Applied to representative models including Self-Forcing and LongLive, OPSD-V consistently improves visual quality, motion dynamics, and VBenchLong scores. A user study with 10 participants comparing 20 video pairs found OPSD-V preferred over the base model in 66.0% of overall preference judgments (82.5% excluding ties).

Paper Overview

Field: Computer Vision Authors: Hongyu Liu, Chun Wang, Feng Gao, Xuanhua He, Yue Ma, Ziyu Wan, Yong Zhang, Xiaoming Wei, Qifeng Chen Published: 2026-07-09 arXiv: 2607.08766

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 the earlier history can be replaced with real video context. This provides dense denoising-level correction targets under the on-policy AR cache dynamics, without changing the sampler, the number of denoising steps, or the inference-time cache mechanism.

Results

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 involving 10 participants comparing 20 video pairs demonstrates that OPSD-V outperforms the base model in 66.0% of overall preference judgments (82.5% excluding ties).

--- *Collected automatically on 2026-07-12.*