← Key papers

Paper review · November 14, 2026

G-STAR: High-Compression Autoregressive Image Generation by Scaling Group Size

G-STAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation

Generative mediaEfficiency & infra Advanced Upvotes 40 cs.CV LongCat

By increasing both the spatial compression ratio and the number of codes per token and adding a semi-autoregressive decoder, the model generated a 1024×1024 image in 2.5 seconds on a single H200.

Problem

Autoregressive image generation based on discrete tokens generates tokens one at a time in sequence. This makes it slower at higher resolutions. Stronger spatial compression reduces the number of generation steps, but each token has to carry more information, which increases quantization loss.

Method

Each spatial token is represented by G codes using Product Quantization. The authors proposed a joint spatial–group scaling principle: when the per-axis compression ratio grows by α, growing G by α² keeps the total nominal bit budget unchanged. They built a 32× tokenizer initialized from Qwen-ViT and formed an effective 64× tokenizer through 2×2 lossless packing. Dependencies between codes within a group are handled by a semi-autoregressive head. A parallel trunk does the heavy computation in one pass, and only a single-layer causal Transformer handles sequential processing. With this design they trained GSTAR-Pro-8B (32×, G=16) and GSTAR-Flash-1.7B (effective 64×, G=32).

Results

Limitations

The nominal bit budget is not the actual entropy or amount of semantic information. The joint scaling principle is therefore an approximation, not a guarantee of equal information. The benchmark figures for the GSTAR models are cut off in the provided text, so the exact difference from baseline models cannot be confirmed. The speed comparison assumes a single H200 GPU and an optimized inference implementation.

Why it matters

It presents a design principle that uses strong compression to reduce slow speed, a weakness of autoregressive image generation. It makes it more feasible to use unified multimodal generation models, which handle text and images in a single model, in real services.

Key terms

Product Quantization(PQ)
A method that splits a feature vector into several channel groups and quantizes each group separately. It increases the expressiveness of a single token without increasing sequence length.
Spatial compression ratio
The ratio of the number of pixels along one side of an image to the size of one side of the token grid. At 32×, one side of 1024 pixels becomes 32 tokens.
Semi-autoregressive head
A prediction head that computes code representations in parallel and handles dependencies between codes only with a lightweight sequential module, aiming for both speed and quality.

Original

Authors: Xu Huang, Ye Huang, Xiaotong Li, Zhengrong Yue, Jialv Zou, Minhao Jing, Haodong Li, Wenhao Yan, et al. · Posted 2026-11-08 · License CC BY 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.