I think the problem with this being AI based is that predictability and replicability are basically non-existent, and that realistically speaking most people's machines would not be able to reliably generate those images.
also, robohash works because it's simple, any sha256 can easily map to an image, I'm not sure how you could map any pubkey into an embedding vector without also risking the generation of some very unsavory content
oh and you also run into the risk of the same image gen resulting in completely different things in different clients. For example if my client uses checkpoint A and yours uses checkpoint B we could end up with VASTLY different images **for the same npub**
idk.. I tried the same prompt several times. I think with large prompts, our AIs are quite predictable.
I think I may not have explained myself properly... you need to somehow map 460c25e682fda7832b52d1f22d3d22b3176d972f60dcdc3212ed8c92ef85065c into a list of words for a prompt. the simplest way to do this (have a word list similar to bip39 and just pick n bytes and pick a word from the list) would result in very easily replicatable images, and the complex way could (mapping each n bytes into a value in the embedding vector) would result in **unpredictably** unsavory images (especially when dealing with different models)
Those things can be easily filtered out just like they do today on public AI-generating services.
