Bias in text-to-image models: standardized evaluation across Stable Diffusion, DALL-E 2, and CLIP

AI Impact Summary

The piece details how biases from training data (LAION-5B, MS-COCO), data filtering, CLIP-based inference, latent-space directions, and post-hoc safety filters influence outputs from Stable Diffusion v1.4/v2 and DALL-E 2. It emphasizes bias is sociotechnical, not solvable by a single algorithm, and requires tooling, red-teaming, and multi-faceted evaluation across these systems. For engineering teams, implement bias auditing with tools like the Stable Bias project's Average Diffusion Faces, Face Clustering, and Colorfulness Explorer; broaden data sources; and embed governance to mitigate reputational or regulatory risk from biased imagery.