QuadTok: Quadtree Visual Tokenizer for Autoregressive Image Generation
QuadTok's quadtree tokenizer lets a 947M model reach 2.08 gFID while using fewer tokens.
QuadTok is a hierarchical quadtree visual tokenizer for autoregressive image generation that allocates more tokens to intricate regions and fewer to homogeneous areas. Compared with a fixed 256-token grid, the ImageNet-trained tokenizer uses about 10% fewer tokens on ImageNet and 9% fewer in zero-shot transfer to COCO, with comparable reconstruction fidelity. A 947M GPT-style generator conditioned on a supplied quadtree topology achieves 2.08 gFID on ImageNet 256x256 and enables zero-shot spatially controlled generation. Code is released at the authors' GitHub repository.
- Quadtree tokenizer spends tokens on visually complex regions.
- Saves about 10% of tokens versus a fixed 256-token grid.
- A 947M GPT-style model reaches 2.08 gFID on ImageNet 256.
- Supports zero-shot spatially controlled generation from the quadtree topology.
- AI models
- QuadTok
From the original article
1 min read at huggingface.co
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.
By Yucheng Mao, Zeyuan Chen, Xiaojun Shan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.