QuadTok: Quadtree Visual Tokenizer for Autoregressive Image Generation
What it is
QuadTok is a framework that uses a hierarchical quadtree structure to tokenize images, allowing for variable resolution based on visual complexity. It enables autoregressive image generation by leveraging the natural causality of the tree structure, achieving competitive FID scores on ImageNet while reducing token count compared to fixed grids.
The problem it solves
Fixed-resolution tokenization wastes capacity on homogeneous regions and fails to capture fine details in complex areas, leading to inefficient or lower-quality autoregressive image generation.
What you could use it for
- Autoregressive image generation with variable resolution
- Spatially controlled image synthesis
- Efficient visual tokenization for diffusion or AR models
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.