ToFu: Visual Tokens Reduction via Fusion for Multi-modal, Multi-patch, Multi-image Tasks

Vittorio Pippi 1 Matthieu Guillaumin 2 Silvia Cascianelli 1 Rita Cucchiara 1 Maximilian Jaritz 2 Loris Bazzani 2 1 University of Modena and Reggio Emilia 2 Amazon · 2025

Read ToFu: Visual Tokens Reduction via Fusion for Multi-modal, Multi-patch, Multi-image Tasks by Vittorio Pippi 1 Matthieu Guillaumin 2 Silvia Cascianelli 1 Rita Cucchiara 1 Maximilian Jaritz 2 Loris Bazzani 2 1 University of Modena and Reggio Emilia 2 Amazon (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-2503-04444.txt.

Description

Vittorio Pippi and co-authors shrink the visual appetite of multimodal AI models. Large multimodal models read images as long strings of visual tokens, and multi-image tasks multiply that cost until the computation groans. Existing trims are tied to one encoder design, demand retraining, or handle only single images. Their 2025 paper proposes ToFu, a token fusion method that needs no training and works with any encoder.

The idea is housekeeping: walk through the visual tokens in order, keep the distinctive ones, and merge each redundant token into a partner it resembles. On an established interleaved-image benchmark, the fused inputs held performance while cutting the token count. The team stress-tested further on a new comparison benchmark of their own, where many images must be weighed against each other at once. The authors pitch fusion as a free lunch the field left on the table.

Catalog record

CREATORVittorio Pippi 1 Matthieu Guillaumin 2 Silvia Cascianelli 1 Rita Cucchiara 1 Maximilian Jaritz 2 Loris Bazzani 2 1 University of Modena and Reggio Emilia 2 Amazon
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE84 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-04444

Browse the full library · About the Tower of Babel