G-STAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation
November 7, 2026
G-STAR is a spatial-group co-scaling framework for discrete autoregressive models that scales group size quadratically with the spatial compression factor.
HOW THIS AFFECTS YOU
●
builderYou can generate images or videos much faster by using highly compressed token representations.
●
researcherYou can balance the trade-off between sequence length and representation quality using group-scaled token capacity.