Demographic Bias in T2I Models

This field studies how text-to-image systems represent people across attributes such as gender, race, age, disability, and culture, because generated images can reinforce stereotypes or erase groups. Audits compare outputs for occupations, neutral and explicit demographic prompts, varied contexts and languages, sometimes against workforce statistics or community-informed standards. Mitigation approaches include prompt rewriting, embedding and attention steering at inference time, and fairness-aware guidance or training. A central challenge is defining fair representation for each setting: reducing measured bias can compromise prompt fidelity, while simple benchmarks may miss intersectional or context-specific harms.