Who Defines Fairness? Target-Based Prompting for Demographic Representation in Generative Models

Researchers propose a lightweight inference-time framework to mitigate demographic bias in text-to-image models like Stable Diffusion and DALL-E without requiring model retraining. The method allows users to define fairness preferences—from uniform distributions to LLM-informed definitions—and constructs demographic-specific prompt variants to achieve target outcomes. Evaluation across 36 prompts and 30 occupations demonstrates the framework can shift skin-tone distributions toward declared fairness targets.