Group-scaled token capacity enables deeply compressed autoregressive generation.
GSTAR scales spatial compression and token groups for 32x, effective 64x, competitive autoregressive image generation in 2.5 seconds at 1024x1024.
Key Takeaways
- GSTAR is a discrete autoregressive image generation framework designed to greatly compress token sequences while preserving reconstruction, understanding, and generation quality.
- Its core idea is spatial-group co-scaling: when spatial compression increases, the group size should grow quadratically under a fixed information budget to maintain model capacity.
- A semi-autoregressive decoder reduces the latency caused by larger token groups by combining parallel feature computation with lightweight causal modeling.
- The reported results show 32x compression and effective 64x compression, competitive performance on five text-to-image benchmarks, and 2.5-second 1024x1024 image generation on a single NVIDIA H200.
Meta description: GSTAR compresses discrete autoregressive image generation, enabling 32x and effective 64x compression, competitive benchmarks, and 2.5-second images.
Tags: GSTAR, discrete autoregressive models, image generation, text-to-image, compression
Featured image: abstract, friendly illustration of compressed image tokens forming a picture, with no real people, logos, or text.
GSTAR is a discrete autoregressive image generation framework for making models fast without losing quality needed for reconstruction, understanding, and generation. It reports 32x compression and effective 64x compression while performing competitively on five representative text-to-image benchmarks.
The Bottleneck in Discrete Autoregressive Image Models
Discrete autoregressive models generate images by predicting discrete token sequences step-by-step, each representing part of an image. Longer sequences preserve representation quality but slow generation; aggressive spatial compression shortens sequences but can remove information, hurting reconstruction, understanding, and generation. The challenge is to reduce sequence length while retaining enough information for strong performance.
A Scaling Rule Behind GSTAR
Building on product quantization, which uses smaller coded groups rather than independent elements, the work shows that under a fixed nominal information budget, group size should scale quadratically with the per-axis spatial compression factor. More aggressive spatial compression requires larger groups to preserve information for reconstruction, understanding, and generation. GSTAR’s spatial-group co-scaling framework scales spatial compression and group size together, enabling 32x and effective 64x compression without collapsing model capacity.
Handling Larger Groups Without Slowing Generation
Larger groups can introduce intra-group dependencies; sequential modeling can increase latency and reduce quality. GSTAR uses a semi-autoregressive decoder combining parallel feature computation with lightweight causal modeling. This processes parts of groups efficiently while preserving autoregressive coherence, keeping larger-group benefits without full latency cost.
Benchmark Results and Reported Speed
Results cover GenEval, DPG-Bench, PRISM, LongText-Bench, and Qwen-Image-Bench, where highly compressed models perform competitively with leading models. At 1024x1024, the model generates each image in 2.5 seconds on a single NVIDIA H200, a 20x speedup over prior discrete autoregressive models. Compression alone is not enough; GSTAR combines heavy compression with competitive performance. Exact scores are not listed, so this is a reported competitive result, not a numerical leaderboard.
What Readers Should Take From It
GSTAR’s main value is its design principle: highly compressed discrete representations can support autoregressive image generation when spatial compression and group size are scaled together under a fixed information budget. The concrete takeaway is reported 32x compression, effective 64x compression, competitive benchmark results, and 2.5-second 1024x1024 generation on a single NVIDIA H200. If these results hold beyond the reported benchmarks, GSTAR suggests discrete autoregressive image generation can be both highly compressed and low-latency without giving up strong generation quality.

Comments (0)