OPSD-V Enhances Autoregressive Video Generation with On-Policy Self-Distillation

What Changed

Few-step autoregressive (AR) video diffusion models offer low-latency generation of long videos but commonly suffer from error accumulation and degradation of motion dynamics over extended sequences. A new approach, OPSD-V (On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators), addresses these limitations by introducing an on-policy self-distillation paradigm during post-training. The core innovation lies in using real long-video data as temporal context during training to provide dense trajectory-level supervision.

Unlike prior methods that might modify the inference mechanism, OPSD-V preserves the original few-step inference path, including the sampler, number of denoising steps, and inference-time cache mechanism. This ensures that the improvements are achieved without introducing additional computational overhead or complexity during deployment. The method has been applied to existing few-step AR video models, such as Self-Forcing and LongLive, demonstrating consistent enhancements in visual quality and motion dynamics.

Technical Details

OPSD-V operates on a student-teacher framework during the post-training phase. The student model executes the exact inference-time rollout, generating each video chunk by conditioning on its own previously generated KV (Key-Value) cache. This mimics the real-world inference scenario where the model relies solely on its self-generated history.

In parallel, a teacher model is evaluated at the same denoising states visited by the student. However, the teacher utilizes a 'cleaner' AR-consistent temporal cache. This cleaner cache allows older history to be replaced by real-video context, effectively providing a more accurate and stable temporal reference. By comparing the student's output with the teacher's output, OPSD-V provides dense denoising-level corrective targets. This supervision occurs under the on-policy AR cache dynamics, meaning the corrective signals are generated in the context of how the student actually performs inference.

This self-distillation process ensures that the student learns to correct its own errors and maintain motion consistency over long horizons, without needing to change its inference-time architecture or parameters. The introduction of real long-video data as temporal context is crucial for providing the rich, accurate historical information that the teacher uses to guide the student. This effectively trains the student to produce outputs that are more consistent with real video data, even when relying on its own generated history.

Benchmark Analysis

Experiments conducted with OPSD-V applied to Self-Forcing and LongLive models showed consistent improvements across various metrics. The method led to enhanced visual quality and improved motion dynamics. Specifically, OPSD-V demonstrated improved VBenchLong scores, indicating better performance on benchmarks designed to evaluate long-duration video generation.

A user study involving 10 participants was conducted to compare 20 video pairs generated by OPSD-V enhanced models against their base counterparts. The results indicated a strong preference for OPSD-V. Participants preferred OPSD-V over the base models in 66.0% of overall-preference judgments. When ties were excluded, this preference rose to 82.5%, highlighting a significant perceived improvement in the quality of the generated videos.

Developer Implications

For developers working with autoregressive video generation, OPSD-V offers a valuable post-training technique to enhance model performance without requiring fundamental architectural changes or increased inference costs. The ability to mitigate error accumulation and improve motion dynamics in long video sequences is critical for applications requiring high-fidelity and temporally consistent video content.

Integrating OPSD-V into existing pipelines for models like Self-Forcing and LongLive could lead to immediate improvements in the quality of generated videos. The method's focus on preserving the original few-step inference path means that developers can leverage these enhancements without needing to re-engineer their deployment strategies or incur higher computational demands at inference time. This makes OPSD-V a practical and efficient solution for improving the robustness and visual appeal of AR video generators.

Bottom Line

OPSD-V represents a significant advancement in improving the performance of few-step autoregressive video diffusion models. By employing an on-policy self-distillation framework that uses real long-video data for dense temporal supervision, OPSD-V effectively addresses the long-standing issues of error accumulation and degraded motion dynamics in long video rollouts. The method's key strength lies in its ability to deliver these improvements while maintaining the original, efficient inference path. This makes OPSD-V a highly practical and impactful technique for developers aiming to generate higher quality, more consistent long-form video content with existing AR models.