AI Briefing
KOSign in

G-STAR Achieves 20x Speedup in Discrete Autoregressive Image Generation with 32x Compression

·2026.11.14 03:44

Key point

The framework scales group size quadratically with spatial compression to enable 32x compression while maintaining competitive performance on five text-to-image benchmarks.

Details

  • G-STAR is a spatial-group co-scaling framework for discrete autoregressive image generation that achieves 32x and effective 64x compression.
  • The model delivers a 20x speedup over prior discrete autoregressive models, generating 1024x1024 images in 2.5 seconds on a single NVIDIA H200.
  • Performance remains competitive with leading models across GenEval, DPG-Bench, PRISM, LongText-Bench, and Qwen-Image-Bench.
  • A semi-autoregressive decoder combines parallel feature computation with lightweight causal modeling to handle intra-group dependencies introduced by larger groups.
  • The approach is based on product quantization, where group size scales quadratically with the per-axis spatial compression factor under a fixed nominal information budget.
image-generationcompressionautoregressive-modelsarxivinference-speedup
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Reddit38% of 2026 arXiv papers flagged as AI-written, up from 20% in early 2025RedditSparseRead Reduces Token Usage by Up to 92.9% for Web-Facing AI AgentsVercel BlogMiniMax H3 Image and Highspeed models added to AI GatewayHacker NewsNext-Concept Modeling Achieves 21x Faster Generation Than Diffusion Baselines on LM1BTLDR AIOpenAI launches Images 2.6 with 50% faster generation and GIF supportGeekNewsArtCraft releases open-source IDE for AI image and video creation with 62-model catalog