Sociological Measurement of Stereotypes in Text-to-Image Models
A framework for auditing stereotypes in text-to-image generative models that defines a stereotype not as any statistical bias but, following its sociological meaning, as the over-representation of attributes generally associated with a concept. It operationalizes this as the OASIS toolbox: a Stereotype Score for distributional over-representation, WALS for spectral variance along an attribute, StOP to reveal the attributes a model internally ties to a concept via optimized prompts, and SPI to trace how those attributes emerge across the diffusion model's denoising steps.
Published as a conference paper at ICLR 2025 OASIS UNCOVERS: HIGH-QUALITY T2I MODELS, SAME OLD
OASIS (Open-set Assessment of Stereotypes in Image generative models) is a toolbox for quantifying and tracing the origins of stereotypes in text-to-image (T2I) models. It replaces the usual statisti…