SAGA：面向自回归视频生成的稳定加速引导
英文标题：SAGA: Stable Acceleration Guidance for Autoregressive Video Generation
自回归视频扩散模型在生成长时序视频时，因重复使用生成潜变量作为因果上下文，会放大时序误差，导致闪烁、运动抖动和结构漂移。本文从频谱运动学角度提出 SAGA，一种免训练的稳定加速引导方法，通过有限窗口 Slepian 投影的加速度域频谱引导与结构化噪声初始化，在抑制短期时序相关性的同时保留长期运动结构。在 Self-Forcing 模型上，SAGA 将时序质量从 97.30 提升至 97.91，图像质量从 69.60 提升至 70.51，频谱分析和人类偏好研究也验证了其降低时序不稳定性、保持视觉保真度的能力。该方法可直接应用于现有的分块自回归扩散模型，无需修改主干网络，为长视频生成的质量提升提供了实用的零成本方案。
2.6 KB
2.1 KB
640 B
612 B
380 B
9.1 KB
22.8 KB
24.2 KB
34.7 KB
中文
[提交于 2026 年 7 月 9 日]
SAGA：面向自回归视频生成的稳定加速引导
查看论文《SAGA: Stable Acceleration Guidance for Autoregressive Video Generation》的 PDF，作者：Thanh-Nhan Vo 等 4 人
摘要：自回归视频扩散模型能实现高效的流式传输和长时域视频生成，但反复重用已生成的潜变量作为因果上下文会放大时序误差，导致闪烁、运动抖动和结构漂移。本文从频谱运动学角度研究这一失效模式，并识别出离散潜变量加速度是揭示不稳定高频时序扰动的有效信号。为此，我们提出 SAGA，一种面向自回归视频生成的免训练稳定加速引导（Stable Acceleration Guidance for Autoregressive video generation）方法。SAGA 集成了基于有限窗口 Slepian 投影的加速度域频谱引导目标，以及一种结构化自回归噪声初始化策略，该策略能抑制短期时序相关性，同时保留长期运动结构。无需重新训练或修改主干网络，SAGA 可直接应用于现有的分块自回归扩散模型，这是当前高质量生成的主流设置。大量实验表明，SAGA 在多种自回归扩散模型上持续提升了时序质量。在 Self-Forcing 模型上，SAGA 将时序质量从 97.30 提升至 97.91，图像质量从 69.60 提升至 70.51。此外，频谱分析和人类偏好研究证明，SAGA 在保持视觉保真度的同时减少了时序不稳定性。
提交历史
来自：Trong-Thuan Nguyen
[v1] 周四，2026年7月9日 00:55:10 UTC (2,178 KB)
English / 原文
[Submitted on 9 Jul 2026]
Title:SAGA: Stable Acceleration Guidance for Autoregressive Video Generation
View a PDF of the paper titled SAGA: Stable Acceleration Guidance for Autoregressive Video Generation, by Thanh-Nhan Vo and 4 other authors
Abstract:Autoregressive video diffusion enables efficient streaming and long-horizon video generation, but repeatedly reusing generated latents as causal context can amplify temporal errors, resulting in flickering, motion jitter, and structural drift. In this paper, we investigate this failure mode from a spectral kinematic perspective and identify discrete latent acceleration as an effective signal for revealing unstable high-frequency temporal perturbations. To this end, we propose SAGA, a training-free \textbf{\textit{s}}table \textbf{\textit{a}}cceleration \textbf{\textit{g}}uidance approach for \textbf{\textit{a}}utoregressive video generation. SAGA integrates an acceleration domain spectral guidance objective based on finite-window Slepian projections with a structured autoregressive noise initialization strategy that suppresses short-range temporal correlations while preserving long-range motion structure. Without retraining or modifying the backbone, SAGA can be directly applied to existing chunk-wise autoregressive diffusion models, which is the prevalent setting for high-quality generation. Extensive experiments show that SAGA consistently improves temporal quality across multiple autoregressive diffusion models. On Self-Forcing, SAGA improves Temporal Quality from 97.30 to 97.91 and Image Quality from 69.60 to 70.51. Moreover, spectral analysis and human preference studies demonstrate that SAGA reduces temporal instability while maintaining visual fidelity.
Submission history
From: Trong-Thuan Nguyen [
]
[v1]
Thu, 9 Jul 2026 00:55:10 UTC (2,178 KB)