OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators
Summary
OPSD-V improves few-step autoregressive video diffusion models by using real long-video data as temporal context during training, providing dense trajectory-level supervision that enhances visual quality and motion dynamics without altering inference mechanisms.
View Cached Full Text
Cached at: 07/10/26, 06:17 AM
Paper page - OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators
Source: https://huggingface.co/papers/2607.08766
Abstract
OPSD-V enhances few-step autoregressive video diffusion models by using real long-video data for temporal context during training, providing dense trajectory-level supervision that improves visual quality and motion dynamics without altering inference mechanisms.
We propose OPSD-V, anon-policy self-distillationparadigm 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 fromerror accumulationand weakenedmotion dynamicsduring 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 astemporal contextduring training and use it to providedense 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-visiteddenoising states, but uses a cleanerAR-consistent temporal cachein 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, andVBenchLongscores. Auser studywith 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).
View arXiv pageView PDFProject pageGitHub8Add to collection
Get this paper in your agent:
hf papers read 2607\.08766
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2607.08766 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2607.08766 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2607.08766 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models
This paper introduces D-OPSD, a novel training paradigm for step-distilled diffusion models that enables on-policy self-distillation during supervised fine-tuning. It allows models to learn new concepts or styles without compromising their efficient few-step inference capabilities.
Dynamic-in-Few-Step: Unifying Dynamic Computation and Few-Step Distillation for Efficient Video Generation
The paper proposes a post-training acceleration framework for video diffusion models that integrates dynamic structural sparsification with few-step distillation, achieving significant speedup while maintaining quality.
AAD-1: Asymmetric Adversarial Distillation for One-Step Autoregressive Video Generation
AAD-1 introduces asymmetric adversarial distillation with phased training to achieve one-step autoregressive video generation, outperforming prior methods on VBench.
Rollout-Marginal Distillation for Long-Horizon Autoregressive Video Generation
This paper introduces Rollout-Marginal Distillation (RMD) for autoregressive video diffusion, where each generated chunk is scored independently against a chunk teacher to avoid error-compounding from imperfect temporal context, followed by video-level DMD to restore coherence. RMD reportedly maintains high visual quality far beyond its training horizon and outperforms video-level DMD baselines.
StreamOPD: A Post-Training Recipe with Spatio-Temporal Cue Gating for Streaming Video Understanding
StreamOPD enhances streaming video understanding through a post-training recipe using on-policy distillation and spatio-temporal cue gating, achieving significant benchmark improvements without requiring inference-time memory.