GeoDiv: Measuring Concept Diversity of Images Across Geographical Regions

Abhipsa Basu, Mohana singh, Venkatesh Babu Radhakrishnan

Keywords: Geographical Diversity, Image Generation Diversity

Abstract: Image datasets—real and synthetic—often lack geographical diversity in how concepts are portrayed across regions. Existing metrics rely on curated datasets or visual dissimilarities, limiting interpretability. We propose GeoDiv, a metric that leverages large language models to identify region-specific attribute variations for a concept, uses a VQA model to measure their prevalence in images, and computes entropy over the resulting distributions. Applied to real and synthetic datasets (including Stable Diffusion and Flux.1-Schnell) across three concepts (house, car, bag) and six countries, GeoDiv reveals higher diversity in real-world images, with the UK and Japan being least diverse and Colombia the most. Our results underscore the need for geographical nuance in generative models and we believe that GeoDiv as a step toward measuring and mitigating regional biases.

Submission Number: 10