SAGA stabilizes autoregressive video generation without retraining

Autoregressive video diffusion models excel at streaming and long-horizon generation but suffer from temporal errors like flickering, motion jitter, and structural drift due to reused latent context. A research team (Vo et al.) identifies this failure mode from a spectral kinematic perspective, pinpointing discrete latent acceleration as a signal of unstable high-frequency perturbations. They propose SAGA, a training-free guidance approach that integrates acceleration domain spectral guidance via finite-window Slepian projections with a structured autoregressive noise initialization strategy. This combination suppresses short-range temporal correlations while preserving long-range motion structure.

SAGA applies directly to existing chunk-wise autoregressive diffusion models without retraining or modifying the backbone. Extensive experiments show consistent temporal quality gains across multiple models. On Self-Forcing, Temporal Quality improves from 97.30 to 97.91, and Image Quality from 69.60 to 70.51. Spectral analysis and human preference studies confirm that SAGA reduces temporal instability while maintaining visual fidelity. This offers a practical, cost-effective upgrade for any autoregressive video system.