OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators
2026-07-09 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors propose a new training method called OPSD-V to improve video diffusion models that generate videos step-by-step. Their method helps reduce errors and blurry motion that happen when making long videos by teaching the model using real long videos as a guide during training. They use two models at once: a student that generates video chunks as it normally would, and a teacher that has access to real video history, helping the student learn better. This approach improves video quality and motion without changing how the videos are generated at test time. Their tests show that OPSD-V helps models produce better videos that people prefer.
video diffusion modelsautoregressive modelsself-distillationpost-trainingtemporal contextdenoising stepsKV cachemotion dynamicsfew-step inferencelong-horizon degradation
Authors
Hongyu Liu, Chun Wang, Feng Gao, Xuanhua He, Yue Ma, Ziyu Wan, Yong Zhang, Xiaoming Wei, Qifeng Chen
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).