The DAGroup team from Peking University, together with Meituan Longcat, SJTU, and co-authors, has released G-STAR (arXiv:2611.13004) — an open framework for autoregressive image generation using deeply compressed discrete tokens. A 1024x1024 image is generated in 2.5 seconds on a single NVIDIA H200, which is approximately 20 times faster than previous discrete autoregressive models. Weights for two models and inference code have already been uploaded to Hugging Face.

image
image

What Happened

On November 8, 2026, the paper G-STAR (arXiv:2611.13004) was published on arXiv, in which the DAGroup team from Peking University, Meituan Longcat, SJTU, and co-authors describe a group-scaled token capacity approach for autoregressive image generation. The claimed speed is 1024x1024 in 2.5 seconds on a single NVIDIA H200: 21.57 times faster than Janus-Pro-7B and 8.36 times faster than Emu3.5-34B. Two models were released: G-STAR-Pro-8B with 32x compression (GenEval 0.84, DPG-Bench 83.38) and G-STAR-Flash-1.7B with an effective 64x compression (GenEval 0.76, DPG-Bench 80.18). In instruction-based editing, Pro-8B scores 7.45 on GEdit-Bench — nearly on par with GPT-Image-1 (7.53) and Qwen-Image-Edit (7.56). Weights for both models, the tokenizer, the SR2x refiner, and inference code have been published in the DAGroup-PKU/G-star repository on Hugging Face.

Context

Discrete autoregressive generators render an image as a sequence of codes from a tokenizer. Previously, the depth of spatial compression was limited by the capacity of a single token: strong compression destroyed fine details, while weak compression left a sequence too long to generate quickly token by token. Therefore, previous discrete autoregressive models like Janus-Pro and Emu3.5 took tens of seconds to generate a 1024x1024 frame. G-STAR changes the problem formulation: the bit budget per position is fixed, while the number of codes in a group G grows quadratically with the degree of spatial compression alpha, so that 32x compression and an effective 64x via lossless 2x2 block packing do not necessarily have to destroy details — this is a testable hypothesis, not a slogan. Parallelism provides a semi-autoregressive decoder that computes features for all slots in a group simultaneously, while code consistency is maintained by a single causal layer with a KV-cache — a compromise between diffusion parallelism and autoregressive causality, an engineering-simple construction.

Why This Matters for the Industry

The claimed speed, according to source assessments, for the first time brings discrete autoregression into a range suitable for interactive production, making it comparable to diffusion pipelines, which directly reduces the inference cost of visual generation. Open weights for Pro-8B and Flash-1.7B allow reproducing the results locally and building products without dependence on paid APIs, while Flash-1.7B is sized appropriately as an on-premises base for cheap edge and on-device rendering. If independent reproductions confirm the numbers, a wave of follow-up work should be expected: transferring the approach to video and unified any-to-any models, fine-tunes and LoRA on open weights, first third-party comparisons with modern diffusion baselines on real traffic, as well as product wrappers for batch generation of product cards, banners, and A/B variants of a single composition. In the longer term, this direction leads to unified multimodal pipelines where image and video are simply sequences of discrete tokens on par with text, with a predictable reduction in the cost of such generation.

Why This Matters for Users

Both models can be downloaded from the DAGroup-PKU/G-star repository on Hugging Face and run following the instructions in the README on a single H100 or H200 GPU. Generation outputs a 1024x1024 image, which the SR2x refiner upscales to 2048x2048. Pro-8B can edit an existing image based on a text prompt: changing style and text while preserving composition, and the readability of text and quality of faces in the frame withstand even 64x compression. No public API or hosting has been announced, so the entry barrier for practical experiments is self-hosted or rented inference, meaning a realistic minimum for a prototype or internal POC on your own cases (faces, text, UI graphics) is one card.

What Is Still Unknown / Limitations

All quality metrics, including GenEval, DPG-Bench, and GEdit-Bench, are author-reported and have not yet been confirmed by independent community verification. The speed comparison was made against previous discrete autoregressive models, not modern diffusion systems, so "approximately 20 times faster" is correct relative to this class of baselines, not in an absolute sense. The claim that the gap of 7.45 versus 7.53 and 7.56 on GEdit-Bench lies within statistical error does not follow from the provided data: they contain no variances, confidence intervals, or sample sizes. Claims of first-in-class status are based on author materials and require careful verification. Scenarios for transferring the approach to video, edge rendering, and the emergence of product wrappers are interpretations that are fair provided the numbers are reproducible.

Sources

Author

Look at AI, editorial team