University of Amsterdam develops inclusive AI image generator for non-English speakers

Most generative AI tools today rely on translation pipelines that convert non-English text into English before generating an image. This often leads to loss of cultural nuance and unfair disadvantages for billions of non-Anglophone users. NeoBabel instead learns directly from input text in multiple languages, preserving linguistic subtleties and producing more accurate, culturally aligned visuals.

The NeoBabel project brings together a multidisciplinary research team from the UvA’s Faculty of Science and their partners. Key contributors include:

Prof. Cees G. M. Snoek – Full Professor of Computer Science and principal researcher on the project, advocating for open science and transparent AI development.
Mohammad Mahdi Derakhshani – PhD student and first author of the NeoBabel research paper, who emphasises the importance of an open-source research pipeline.
Dheeraj Varghese and Marzieh Fadaee – Collaborative co-authors on the NeoBabel publication, contributing to model architecture and multilingual data integration.
Together, they worked with a combination of multilingual language models and large datasets, expanding publicly accessible image-text pairs to improve direct cross-lingual training.

Why it matters

As generative AI increasingly shapes creative industries, education, media, and digital communication, building systems that fairly represent the world’s linguistic diversity is both a technological and societal imperative. NeoBabel demonstrates how academic research can push AI toward broader accessibility and cultural relevance — fostering innovation that truly reflects global needs.

Source:
University of Amsterdam (2026). “An inclusive AI image generator for non-English speakers.”
Available via the University of Amsterdam news portal.