Mitigating stereotypical biases in text to image generative systems

Runway proposes a method to mitigate social biases in text-to-image generative models by finetuning on synthetic data that varies in perceived skin tones and genders.

Their diversity finetuned (DFT) model improves group fairness by 150% for perceived skin tone and 97.7% for perceived gender.

The company will release all text prompts and code to generate training images.