论文概要
研究领域: CV 作者: Hongyu Liu, Chun Wang, Feng Gao, Xuanhua He, Yue Ma, Ziyu Wan, Yong Zhang, Xiaoming Wei, Qifeng Chen 发布时间: 2026-07-09 arXiv: 2607.08766
中文摘要
我们提出OPSD-V,一种用于后训练少步自回归(AR)视频扩散模型的在线策略自蒸馏范式。现有少步AR视频生成器可以低延迟生成长视频,但在长自回归展开过程中仍遭受误差累积和运动动力学减弱的问题。OPSD-V在保持原始少步推理路径的同时,减少了长程退化。核心思想是在训练期间引入真实长视频数据作为时间上下文,并提供密集的轨迹级监督。具体而言,学生遵循精确的推理时间展开,在其自身先前生成的KV缓存条件下生成每个片段。同时,教师在相同的学生访问的去噪状态下进行评估,但使用更干净的AR一致时间缓存,其中较早的历史可以被真实视频上下文替换。这提供了在线策略AR缓存动态下的密集去噪级校正目标,而无需改变采样器、去噪步数或推理时间缓存机制。我们将OPSD-V应用于代表性少步AR视频模型,包括Self-Forcing和LongLive。实验显示在视觉质量、运动动力学和VBenchLong分数上均有持续提升。一项涉及10名参与者比较20个视频对的用户研究表明,OPSD-V在整体偏好判断中优于基础模型66.0%(排除平局为82.5%)。
原文摘要
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 cleane…
— 自动采集于 2026-07-12
#论文 #arXiv #CV #小凯
