An image of an archeologist adventurer who wears a hat and uses a bullwhip

I don't understand why problems like this aren't solved by vector similarity search. Indiana Jones lives in a particular part of vector space.

Two close to one of the licensed properties you care to censor the generation of? Push that vector around. Honestly detecting whether a given sentence is a thinly veiled reference to indiana jones seems to be exactly the kind of thing AI vector search is going to be good at.

Not worth it to compute the embedding for Indy and a "bull-whip archaeologist" most guardrails operate at the input level it seems?

Thinking of it in terms of vector similarity does seem appropriate, and then definition of similarity suddenly comes into debate: If you don't get Harrison Ford, but a different well-known actor along with everything else Indiana-Jones, what is that? Do you flatten the vector similarity matrix to a single infringement-scale?

> Not worth it to compute the embedding for Indy

If IP holders submit embeddings for their IP, how can image generators "warp" the latent space around a set of embeddings so that future inferences slide around and avoid them--not perfectly, or literally, but as a function of distance, say, following a power curve?

Maybe by "Finding non-linear RBF paths in GAN latent space"[0] to create smooth detours around protected regions.

0. https://openaccess.thecvf.com/content/ICCV2021/papers/Tzelep...