I wonder if it's possible to assign a unique id to each image with a fingerprint, then feed it to an AI model when training. Then, on inference, be able to estimate the weights of each reference image used to produce an output. I bet they'd be a lot but still useful output.

The open web is dead, long live the open web. ⌐◨-◨

Building /warpcast and /farcaster

— gabrielayuso.com

If you're high on web3 growth, I'm your dealer @ intract.io | ex-TradFi bro @ Blackstone

I don't really know. That's something I'd like to learn.

There are people and companies who specialize in model introspection because it's mostly a black box.

Do different images get different weights when an AI model produces outputs? How are these weights decided?

Have you been following Holly Herndon and May Dryhurst’s spawn/holly+ project? They’re doing something like this for audio inputs to AI