Q-ARVD: Quantizing Autoregressive Video Diffusion Models

Researchers introduced Q-ARVD, a quantization method built to reduce inference cost in autoregressive video diffusion models without hurting output quality.

The paper finds that standard diffusion quantization fails on ARVDs because frame sensitivity is uneven and weight outliers are highly heterogeneous.

Q-ARVD addresses these issues with frame-weighted optimization and adaptive dual-scale outlier handling.

This model-specific approach aims to make streaming and interactive video generation more practical by easing the compute bottleneck.