G-STAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation
THE QUICK READ
The paper proposes G-STAR, a spatial–group co-scaling framework for discrete autoregressive image generation. Based on product quantization, the authors argue that under a fixed nominal information budget, group size should scale quadratically with per-axis spatial compression, enabling 32× and even effective 64× compression. They also introduce a semi-autoregressive decoder combining parallel feature computation with lightweight causal modeling. The authors report competitive results on five text-to-image benchmarks and a 20× speedup, generating each 1024×1024 image in 2.5 seconds on one NVIDIA H200.
KEY POINTS
- The abstract states a design principle that group size should scale quadratically with per-axis spatial compression under a fixed nominal information budget.
- A semi-autoregressive decoder is proposed to combine parallel feature computation with lightweight causal modeling for intra-group dependencies.
- The authors claim competitive performance on GenEval, DPG-Bench, PRISM, LongText-Bench and Qwen-Image-Bench, plus a 20× speedup and 2.5 seconds per 1024×1024 image on a single H200.
READING LIMITS
Only the title and abstract were read; method details, experimental setup, baseline comparisons and ablations were not examined, so the compression, speed and benchmark claims remain unverified.
Source abstract [EN]
Discrete autoregressive models offer a scalable framework for image generation. However, they face a fundamental trade-off: long token sequences preserve representation quality but slow generation, while aggressive spatial compression reduces representation capacity. Based on product quantization, we show that, under a fixed nominal information budget, group size should scale quadratically with the per-axis spatial compression factor. Building on this principle, we propose GSTAR, a spatial--group co-scaling framework that enables \(32\times\) and even effective \(64\times\) compression while preserving strong reconstruction, understanding, and generation performance. Since larger groups introduce intra-group dependencies that must be modeled efficiently, we analyze their impact on inference latency and generation quality and introduce a semi-autoregressive decoder that combines parallel feature computation with lightweight causal modeling. With these designs, our highly compressed models perform competitively with leading models across five representative text-to-image benchmarks (GenEval, DPG-Bench, PRISM, LongText-Bench, and Qwen-Image-Bench). At \(1024\times1024\) resolution, our model delivers a \(20\times\) speedup over prior discrete autoregressive models, generating each image in 2.5 seconds on a single NVIDIA H200. Together, these results show that highly compressed discrete representations can enable high-quality autoregressive generation at low latency.
AI summaries use public titles and abstracts only. Research claims belong to the authors; consult the original paper.