Differentiable Hierarchical Visual Tokenization
Marius Aasan 1 , Martine Hjelkrem-Tan 1 , Nico Catalano 2 , Changkyu Choi 3 , Adín Ramírez Rivera 1 University of Oslo Deptartment of Informatics 1 {}^{1}\underset{\text{Deptartment of Informatics}}{\text{University of Oslo}} Polytechnic University of Milan Artificial Intelligence and Robotics Lab 2 · 2025
Read Differentiable Hierarchical Visual Tokenization by Marius Aasan 1 , Martine Hjelkrem-Tan 1 , Nico Catalano 2 , Changkyu Choi 3 , Adín Ramírez Rivera 1 University of Oslo Deptartment of Informatics 1 {}^{1}\underset{\text{Deptartment of Informatics}}{\text{University of Oslo}} Polytechnic University of Milan Artificial Intelligence and Robotics Lab 2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-02652.txt.
Description
Vision transformers chop each image into fixed square patches, ignoring what the picture shows. Marius Aasan and four colleagues offer a smarter way to cut up images in a 2025 paper. Their tokenizer adapts to image content down to single pixels, yet it can still slot into existing models, even ones already trained.
The method picks its splits by a layered model selection rule guided by information criteria. It performs well both at labeling whole images and at dense tasks that mark each pixel. It can even turn raster pictures into vector shapes right away. For teams with trained models on hand, the design promises finer image handling without starting training from scratch.
Catalog record
| CREATOR | Marius Aasan 1 , Martine Hjelkrem-Tan 1 , Nico Catalano 2 , Changkyu Choi 3 , Adín Ramírez Rivera 1 University of Oslo Deptartment of Informatics 1 {}^{1}\underset{\text{Deptartment of Informatics}}{\text{University of Oslo}} Polytechnic University of Milan Artificial Intelligence and Robotics Lab 2 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 67 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-02652