Text-To-Image Accessibility
summary
Adobe Firefly’s current alt text like “variation 1 of 4” fails blind and low vision users because it doesn’t say which image matches which part of a prompt. The project builds an embed similarity score, trains an image-difference captioning model to summarize meaningful distinctions across four outputs, and prototypes alt text that flags detect AI mistakes like missing pedals or wrong candle counts. Goal is a prototype that lets BLV users pick the best image independently.
AI-written summary · about
chapters (5)
- 0:00Introduction and background
- 13:36WCAG guidelines for alt text
- 26:31AI mistakes and abstract objects
- 36:50Next steps and future goals
- 39:42Q&A session
related talks
55mBridging the Digital Divide: Empowering Accessibility Through AI Innovations
22mFAULT TEXT: The Adversities in Automated Alternative Text with Henri Helvetica
42mA Glimpse Into Accessibility
01h 00mAI and alt text
49mUsing AI to Streamline Inclusive Design Without Losing Humanity
22mAccessibility for Everyone
56mBeyond Colors and Captions: How to Provide More Inclusive Accessibility
17mPositive alt-itude! An Inclusion Tool for Your Accessibility