SAGA: Stable Acceleration Guidance for Autoregressive Video Generation

Autoregressive video diffusion models are highly efficient for long-horizon streaming, but they suffer from a persistent failure mode: the recursive reuse of generated latents as causal context accumulates small temporal errors. This leads to visible artifacts such as flickering, motion jitter, and structural drift. The authors investigate whether these instabilities can be mitigated at inference time without retraining the underlying generative model.

The authors analyze this instability through a spectral kinematic lens, identifying that discrete latent acceleration—a second-order temporal operator—acts as a high-pass filter that amplifies high-frequency temporal perturbations. To counteract this, they introduce SAGA (Stable Acceleration Guidance), which consists of two main components:

1. Structured Autoregressive Noise (SAN) Initialization: A noise strategy that decorrelates adjacent frames to neutralize short-range temporal bias while preserving long-range motion structure.
2. Acceleration-Domain Spectral Guidance: An inference-time guidance objective that projects latent acceleration onto a band-limited Slepian basis. This allows the model to suppress high-frequency "jitter" energy while preserving the low-frequency semantic motion dynamics.

Because SAGA operates entirely at inference time by modifying the latent trajectory during the denoising process, it is model-agnostic and can be applied to existing autoregressive diffusion backbones without any architectural changes or fine-tuning.

Experimental results across multiple 1.3B-parameter backbones (such as Self-Forcing and CausVid) demonstrate that SAGA consistently improves temporal consistency metrics. Specifically, on the Self-Forcing model, SAGA improved the aggregate Temporal Quality score from 97.30 to 97.91 and Image Quality from 69.60 to 70.51. Spectral analysis confirmed that the method successfully reduces high-frequency acceleration power, and human preference studies indicate that viewers perceive the resulting videos as more stable and visually faithful compared to unguided baselines.