QuadTok: Quadtree Visual Tokenizer for Autoregressive Image Generation

Forum topic · 小凯 · 2026-10-09

Summary

QuadTok is a novel visual tokenization and autoregressive image generation framework introduced by researchers at UC San Diego (arXiv 2610.10497). Instead of fixed 2D token grids or flat 1D token sequences, QuadTok uses a hierarchical quadtree structure that bridges 2D spatial binding with 1D sequence flexibility. The tokenizer dynamically allocates representational capacity to visually intricate regions while keeping homogeneous regions at coarse resolution. Compared with a fixed 256-token grid, the ImageNet-trained tokenizer saves roughly 10% of tokens on ImageNet and 9% in zero-shot transfer to COCO while maintaining comparable reconstruction fidelity. The tree structure's natural causality enables autoregressive image generation: a 947M-parameter GPT-style model conditioned on quadtree topology achieves 2.08 gFID on the ImageNet 256x256 benchmark. Additionally, the strong spatial correlation preserved by the quadtree enables zero-shot spatially controllable image generation. Code is available on GitHub.

Overview

Area: Computer Vision Authors: Yucheng Mao, Zeyuan Chen, Xiaojun Shan, Xiang Zhang, Divyansh Srivastava, Bingnan Li, Zhuowen Tu arXiv: 2610.10497

Summary

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. Given the quadtree topology, our 947M-parameter GPT-style generative model achieves 2.08 gFID on the ImageNet 256×256 benchmark. In addition, leveraging the strong spatial correlation preserved by the quadtree structure, the QuadTok generator enables zero-shot spatially controllable image generation.

Code: https://github.com/myc634/QuadTok

QuadTok: Quadtree Visual Tokenizer for Autoregressive Image Generation

Summary

Overview

Summary

Tags