QuadTok: Quadtree Visual Tokenizer for Autoregressive Image Generation
Yucheng Mao, Zeyuan Chen, Xiaojun Shan, Xiang Zhang, Divyansh Srivastava, Bingnan Li, Zhuowen Tu
University of California, San Diego
Published October 7, 2026 · on the feed October 8, 2026 · ▲ 12 · arXiv 2610.10497
The abstract
We introduce QuadTok, a novel framework for visual tokenization and autoregressive image generation. Compared to traditional approaches using 2D grids or 1D token sequences, we propose a hierarchical quadtree structure, bridging the gap between 2D spatial binding and 1D sequence-level flexibility. The QuadTok tokenizer dynamically allocates representational capacity to visually intricate areas while leaving homogeneous regions at a coarse resolution. Compared with a fixed 256-token grid, our ImageNet-trained tokenizer saves approximately 10% of tokens on ImageNet and 9% when transferred zero-shot to the COCO dataset, while maintaining comparable reconstruction fidelity. Furthermore, the natural causality introduced by the tree structure seamlessly enables autoregressive image generation. Conditioned on a quadtree topology supplied before generation, our 947M GPT-style generative model achieves a 2.08 gFID on the ImageNet 256 times 256 benchmark. Additionally, leveraging the strong spatial correlation preserved by the quadtree structure, the QuadTok generator enables zero-shot spatially controlled image generation capabilities. Code: https://github.com/myc634/QuadTok.
The Brief
2:01A short podcast summary of the paper. Written and voiced by AI.
Transcript
QuadTok. Quadtree Visual Tokenizer for Autoregressive Image Generation. By Yucheng Mao and six others, from the University of California, San Diego. Published on the seventh of October.
Visual tokenizers for autoregressive image generation have tended to come in two forms. One lays tokens out on a 2D grid, which keeps them bound to places in the image. The other turns the image into a 1D token sequence, which gives flexibility at the level of the sequence. The authors frame these as two advantages that have been hard to have at once.
QuadTok proposes a hierarchical quadtree structure as the bridge between 2D spatial binding and 1D sequence-level flexibility. A quadtree lets representational capacity follow the image. The tokenizer dynamically gives more of it to visually intricate areas, and leaves homogeneous regions at a coarse resolution.
The tree does a second job. Its structure brings a natural causality with it, and that ordering is what makes autoregressive generation possible on top of the tokens. The generative model is conditioned on a quadtree topology that is supplied before generation begins, and produces the image within it.
Against a fixed grid of two hundred and fifty-six tokens, the tokenizer, trained on ImageNet, saves roughly ten percent of tokens on ImageNet itself. Transferred zero-shot to COCO, it saves nine percent. In both cases reconstruction fidelity stays comparable.
On generation, a GPT-style model with nine hundred and forty-seven million parameters reaches a gFID of two point zero eight on the ImageNet benchmark at a resolution of two hundred and fifty-six by two hundred and fifty-six. And because the quadtree preserves strong spatial correlation, the generator can also perform spatially controlled image generation zero-shot.
The quadtree serves three purposes at once here: it spends tokens where the image is complex, it supplies the order an autoregressive model needs, and it keeps enough spatial correlation for control to come zero-shot. A grid offered spatial binding, a sequence offered flexibility, and the tree is the authors' way of keeping both.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free