G-STAR scales token groups for up to 64x image compression
Original title: G-STAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation
At 1024x1024 it generates an image in 2.5 seconds on one NVIDIA H200, about 20x faster than prior discrete autoregressive models.
G-STAR is a spatial-group co-scaling method for discrete autoregressive image generation based on product quantization. The authors argue that, under a fixed information budget, group size should scale quadratically with the per-axis spatial compression factor. They report effective 32x and 64x compression using a semi-autoregressive decoder that combines parallel features with lightweight causal modeling. The compressed models are evaluated on five text-to-image benchmarks.
Unconfirmed: Speed and quality claims are the authors' measurements; independent replication is not described.
- Benchmarks named are GenEval, DPG-Bench, PRISM, LongText-Bench, and Qwen-Image-Bench.
- The paper reports a 20x speedup over prior discrete autoregressive models.
- A semi-autoregressive decoder handles dependencies inside larger token groups.
- Authors say reconstruction, understanding, and generation quality stay competitive.
- Vendors
- NVIDIA
- Products
- NVIDIA H200
- AI models
- G-STAR
From the original article
1 min read at huggingface.co
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.
By Xu Huang, Ye Huang, Xiaotong Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.