Mitigating Stereotypical Biases In Text To Image Generative Systems

Runway researchers tested whether further training on deliberately varied synthetic portraits could change whom an image model depicts by default. Their study fine-tuned Stable Diffusion models on generated images made from prompts combining occupations, ages and demographic descriptors, then evaluated the adapted models using neutral prompts without ethnicity or gender qualifiers.

The authors report that their group-fairness metric improved by 150% for perceived skin tone and 97.7% for perceived gender, with more women and more people of perceived darker skin tone in the outputs. Those measures rest on automatic classification of perceived gender and skin tone, rather than establishing fairness in every sense. The skin-tone classifier agreed with the authors' own labels on only 63.4% of a 750-image validation sample (gender agreement was 99.9%), so the skin-tone result in particular rests on a classifier that disagreed with human labels more than a third of the time.