SAGA: Stable Acceleration Guidance for Autoregressive Video Generation
SAGA is a training-free method to stabilize autoregressive video generation by addressing flickering and structural drift caused by recursive latent reuse. It targets high-frequency temporal noise in the latent acceleration domain during denoising. The approach uses structured initialization and spectral guidance to suppress unstable, high-frequency kinematic energy while maintaining essential low-frequency motion structure, leading to perceptibly better temporal quality.