Daniel Demmel
-
Intrigued by multi-modal (text + vision) models like LLaVa, I tried an experiment to create a browser extension that walks the DOM, finds images without good alt, converts the image to Base64, sends it to LLaVa 7B 1.5 (running in LlamaCPP's server) and injects the rich description back into the image tag's alt. Needs much more work and testing, but amazing what a ~5GB (quantised at 5bit) model can do!
Edit: now on Github: https://github.com/daaain/image-alt-text-generator-extension