Heinrich-Heine-Universität Düsseldorf

As machine learning-based text-to-image systems become increasingly prevalent in real-world applications, concerns regarding their potential biases have emerged. To complement existing evaluations, we move beyond single-attribute analysis to study intersectional demographic effects in generated outputs. This setting requires methods that account for the distributional nature and relationships between protected attributes. To address this, we introduce a framework for quantifying intersectional fairness in text-to-image systems that characterizes bias along two complementary dimensions: (i) intersectional diversity, measuring the coverage and balance of demographic groups, and (ii) attribute dependence, quantifying statistical dependencies between protected attributes. We evaluate three popular diffusion models across gender, age, and race in a wide range of professions. Our analysis reveals disproportionate discrimination against individuals with multiple marginalized identities, uncovered through limited intersectional diversity and systematic attribute dependencies. We further demonstrate that a simple prompt-based intervention improves intersectional fairness. We additionally release an interactive visualization tool for exploring fairness in our datasets, as well as supporting the analysis of additional models and demographic groups.