G-STAR: 32-64x token compression makes autoregressive image generation 20x faster
pmttyji · reddit · 2026-11-14
PKU's DAGroup released the G-STAR paper and models tackling the core trade-off in discrete autoregressive image generation: long token sequences preserve quality but are slow, while aggressive spatial compression hurts capacity.
Method
- Based on product quantization, they show group size should scale quadratically with the per-axis spatial compression factor under a fixed nominal information budget, enabling 32x and effective 64x compression while preserving reconstruction, understanding and generation quality.
- A semi-autoregressive decoder combining parallel feature computation with lightweight causal modeling handles the intra-group dependencies larger groups introduce.
Results
- Competitive with leading models across GenEval, DPG-Bench, PRISM, LongText-Bench and Qwen-Image-Bench (GSTAR-Pro-8B averages 76.2, behind only GPT Image, Emu3.5 and Qwen-Image).
- At 1024x1024, 20x speedup over prior discrete autoregressive models: 2.5s per image on a single NVIDIA H200. GSTAR-Flash-1.7B is 21.57x faster than Janus-Pro-7B; GSTAR-Pro-8B is 8.36x faster than Emu3.5-34B.
Paper, code and weights are all open.
Related event: G-STAR from PKU and Meituan Speeds Up Autoregressive Image Generation 20x(4 posts)→
More from Multimodal
- New method distills all of Monet's works into a single painting — vincesitzmann · 2026-11-14
- 'Peak prompt engineering': full prompt for having an LLM draw a JS-only masterpiece — adonis_singh · 2026-11-14
- Ticket roll-paper interaction demo built with Claude Opus wows AI design circles — floguo · 2026-11-14
- AI-generated synthwave meme charts an eBay leather jacket crashing from $480 to $12 — yungcontent · 2026-11-14
- Seedance 2.5 demo: a 1080P split-scene clip with prompts shared — LudovicCreator · 2026-11-14
- Matthew Berman open-sources Scarlet, an AI video editing agent that speeds up his workflow — MatthewBerman · 2026-11-14