QuadTok Visual Tokenizer Cuts ~10% of Image Tokens With Quadtree Allocation, Hits 2.08 gFID
Summary
QuadTok reserves tokens for intricate image regions via a hierarchical quadtree, saving ~10% tokens on ImageNet and ~9% zero-shot on COCO against fixed 256-token grids at matched reconstruction quality. A 947M GPT-style AR generator conditioned on quadtree topology reaches 2.08 gFID on ImageNet 256×256 and supports zero-shot spatial control. Code open-sourced on GitHub.
Originally reported by huggingface.co
Read the original article →Original headline: QuadTok Visual Tokenizer Cuts ~10% of Image Tokens With Quadtree Allocation, Hits 2.08 gFID