G-STAR: 20x Faster Image Generation on Compressed Tokens

🤖 G-STAR: 20x Faster Image Generation on Compressed Tokens

DAGroup (Peking University), along with Meituan Longcat, SJTU, and others, released G-STAR (arXiv:2611.13004) — an autoregressive image generation framework on discrete tokens. The codebook size grows quadratically with compression at a fixed bit budget, achieving 32x and effective 64x through lossless 2x2 block packing. A 1024x1024 image is generated in 2.5 seconds on a single NVIDIA H200.

🌍 Discrete AR generators operate on tokens with 32-64x compression without quality loss and at the speed of diffusion models. This reduces inference costs and paves the way for real-time generation and unified LLM pipelines.

👤 Both models are open on Hugging Face (DAGroup-PKU/G-star) and run on a single H100/H200. Pro-8B edits images based on text, maintaining text and face quality even at 64x.

Source 1: https://arxiv.org/abs/2611.13004 Source 2: http://dagroup-pku.github.io/G-STAR/