G-STAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation
The authors report generating 1024×1024 images in 2.5 seconds on one NVIDIA H200—20 times faster than prior discrete autoregressive models.
G-STAR compresses spatial representations by 32×, and effectively 64×, while increasing token group size to preserve capacity. The paper finds that, at a fixed nominal information budget, group size should grow quadratically with the per-axis compression factor. Its semi-autoregressive decoder combines parallel feature computation with lightweight causal modeling to handle dependencies within those groups. The authors report competitive results against leading models across five text-to-image benchmarks.
HF Daily Papers' note
G-STAR compresses spatial representations by 32×, and effectively 64×, while increasing token group size to preserve capacity. The paper finds that, at a fixed nominal information budget, group size should grow quadratically with the per-axis compression factor. Its semi-autoregressive decoder combines parallel feature computation with lightweight causal modeling to handle dependencies within those groups. The authors report competitive results against leading models across five text-to-image benchmarks.
HF Daily Papers' note
score 5