Minimizing Bias in AI-Generated Images
summary
Generates well over a thousand AI images using MidJourney plus Flux, Runway and an aggregator called PO, and shows that explicit prompting is required to counter stereotypes: use a structure of editorial photo (style), subject, action, physical characteristics, clothes, background and fine details. Finds wins for photorealistic and modern corporate style illustrations, recurring failures for large groups, readable faces, left‑handed writing, accurate mobility aids and subtle hearing devices, and recommends reference images, inpainting and iterative tweaks.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 2:07AI image bias and inclusion
- 7:48AI reinforcing stock image stereotypes
- 13:55Prompting for inclusive images
- 22:16Structuring prompts for consistency
- 31:23Representing non-visible disabilities
- 33:47AI limitations in disability representation
- 42:14AI contextual understanding failures
- 51:08Live demo: AI image iteration
- 1:01:10Tools for better AI prompts
- 1:05:20Future of inclusive AI images
related talks
53mCreating Beautiful, On-Brand Photography with AI
33mAI Images: Redefining Quality with Innovation
49mCreation as Curation [Keynote]
47mAI Image Generation: Visual Artistry with Midjourney, Adobe Firefly, and DALL-E3
57mAImages - how neural networks will affect the production of games
14mAI for Brands
36mArt in the Age of Algorithms
28mUser Research GenAI Better Together