New Framework Lets Users Define Fairness in AI Image Generation

A new research paper proposes a lightweight framework that allows users to specify their own definition of fairness when generating images with text-to-image models like Stable Diffusion and DALL-E. The system intervenes at the prompt level during inference, avoiding the need for retraining or curated datasets. Users can choose from multiple fairness specifications, from simple uniform distributions to complex definitions informed by a large language model. This addresses the problem of these models often replicating societal biases, such as depicting lighter-skinned individuals for high-status professions like 'doctor' or 'CEO' while showing more diversity for lower-status roles like 'janitor'. The framework makes bias mitigation accessible to ordinary users without requiring technical expertise.