G-STAR paper: 32–64x token compression gives 20x faster autoregressive image generation
_akhaliq · x · 2026-11-14
A new paper proposes G-STAR (Group-Scaled Token Capacity), tackling the trade-off in discrete autoregressive image generation between long token sequences (quality but slow) and aggressive spatial compression (fast but lossy).
- Proves that under a fixed nominal information budget, group size should scale quadratically with the per-axis spatial compression factor, building on product quantization
- Introduces a spatial–group co-scaling framework enabling 32x and even effective 64x compression while preserving reconstruction, understanding, and generation quality
- Adds a semi-autoregressive decoder combining parallel feature computation with lightweight causal modeling to handle intra-group dependencies
- Competitive with leading models across five text-to-image benchmarks (GenEval, DPG-Bench, PRISM, LongText-Bench, Qwen-Image-Bench)
- At 1024x1024, generates each image in 2.5s on a single H200 — a 20x speedup over prior discrete autoregressive models
More from Multimodal
- Reddit User Releases AI-Generated Short Film 'Toni Part II' — Szlob · 2026-11-14
- AI virtual influencer series 'The Engagement' S1E6 airs on Rad TV — AIandDesign · 2026-11-14
- AI-Generated Series 'Liminal' Returns with Episode 3: A Door To Nowhere — AccomplishedDrive169 · 2026-11-14
- RTX 5070 Ti ComfyUI benchmarks: overclocking and launch flags yield surprising results — AnalElder · 2026-11-14
- AI-Generated 'Evil Toy Dimension' Is a Hellscape of Furious Dolls and Lava Suns — Affectionate_Film161 · 2026-11-14
- ZONOS3-Mini: MoE TTS model runs realtime on a Pixel 9 and MacBook Air — cocktailpeanut · 2026-11-14