VaaWIT: Visual-Aware Adaptation of Large Language Models for Multilingual Web Image Translation

Researchers introduced VaaWIT, an end-to-end framework for multilingual text translation inside web images.

It combines dual-stream attention with a visual-aware adapter to better fuse fine-grained visual cues and language reasoning.

The method showed strong results on eight tasks across three public benchmarks.

The work tackles a major gap in image-embedded text translation, with potential benefits for accessibility and cross-language search in web content.