QuadTok: quadtree visual tokenizer saves 10% tokens, hits 2.08 gFID on ImageNet 256

Yucheng Mao · hf · 2026-10-08

QuadTok is a hierarchical quadtree visual tokenizer for autoregressive image generation, bridging 2D spatial binding and 1D sequence flexibility.

Highlights:

Dynamically allocates capacity: fine regions get more tokens, homogeneous regions stay coarse.
Saves 10% tokens vs a fixed 256-token grid on ImageNet and 9% zero-shot on COCO, with comparable reconstruction fidelity.
The tree structure naturally enables autoregressive generation: a 947M GPT-style model conditioned on the quadtree topology reaches 2.08 gFID on ImageNet 256×256.
Spatial correlation preserved by the quadtree enables zero-shot spatially controlled generation.

Code is available on GitHub.

More from Multimodal