Bochu Ding

The Latent Gaze

This project investigates how generative AI models “interpret” social identity, and the biases embedded into those visual interpretations. Each portrait pixel represents the median value across hundreds of generated images prompted by simple, yet loaded, social designations — like felon, nurse, and leader.

This corpus of portraits was generated by a Comfy-UI pipeline employing text-to-image model Stable Diffusion XL. Thereafter, a vision model was then deployed to adjust each generated image to align facial features (eyes, nose, mouth). Collapsing the output of these datasets into single synthetic images consolidates the variations, or lack thereof, in the data. Does the dominance of race, ethnicity, complexion and apparel appear through recurrent pixel traces in these portraits?

An interactive exhibit can be found here.