Runway fine-tunes image models to reduce bias
Runway says fine-tuning Stable Diffusion on demographically varied synthetic images sharply improves group fairness, giving builders a practical, data-centric way to curb text-to-image models' skew toward lighter-skinned men.
Runway's research team has described a method for reducing stereotypical bias in text-to-image models. It works by fine-tuning on synthetic data built to vary in perceived skin tone and gender.
The researchers note that leading generative image models over-represent people with lighter perceived skin tones and men. To counter this, they generated training images from text prompts built out of combinations of ethnicities, genders, professions, age groups and other attributes. They then fine-tuned Stable Diffusion 1.5 and Stable Diffusion XL on the resulting dataset.
Runway reports that the resulting "diversity finetuned" (DFT) models improve a group fairness metric by 150% for perceived skin tone and 97.7% for perceived gender. Fairness is measured as disparate impact relative to the light-skinned and male subgroups the models default to. Compared with baseline models given the same prompt and seed, the DFT models produce more people with darker perceived skin tones and more women.
The post also looks at two factors that affect results. One is prompt qualifiers. The other is how the training data is composed, and the team found that a balanced distribution of perceived skin tones performs better.
Runway says it will release all the text prompts and the code used to generate the training images. The post does not give a release date. It also does not say whether the approach has been applied to Runway's own production models or how fine-tuning affects image quality.