G-STAR from PKU & Meituan LongCat: 20x faster autoregressive image generation at 2.5s per image
meituan-longcat · hf · 2026-11-13
Researchers from Peking University, Meituan LongCat and partners published G-STAR, tackling the core trade-off in discrete autoregressive image generation: long token sequences preserve quality but are slow, while aggressive spatial compression hurts capacity.
- Key finding: with product quantization, under a fixed nominal information budget, group size should scale quadratically with the per-axis spatial compression factor.
- Framework: a spatial–group co-scaling scheme enabling 32× and effective 64× compression while preserving reconstruction, understanding, and generation quality, plus a semi-autoregressive decoder combining parallel feature computation with lightweight causal modeling.
- Results: at 1024×1024, a 20× speedup over prior discrete autoregressive models — 2.5 seconds per image on a single NVIDIA H200 — competitive with leading models on GenEval, DPG-Bench, PRISM, LongText-Bench and Qwen-Image-Bench.
Takeaway: heavily compressed discrete representations can enable low-latency, high-quality autoregressive generation.
More from Multimodal
- Someone rebuilt The Simpsons house in real life with AI and it looks oddly plausible — TommycoolGun · 2026-11-13
- Engineer builds 3D Ayodhya model with Chisel, shares the file publicly — gowthami_s · 2026-11-13
- 67B MiniMax H3-Image runs on a 16GB MacBook Air via MLX with ~85% less peak memory — FlimsyBrief4232 · 2026-11-13
- Viral image prompt: watch the alien invasion from your penthouse — charis_ai · 2026-11-13
- Ant Group's Duo v1: a reciprocal world model for real-time full-duplex interaction — inclusionAI · 2026-11-13
- Dev builds ComfyUI nodes to add batch generation to MiniMax H3 workflows — an80sPWNstar · 2026-11-13