randon31415 , 5 hours ago

How close does the deep fake need to be to the original? What we saw with DALLE2 was that each person whose face was restricted made a hole in the latent space of all faces. After enough celebrities faces were restricted, there were so many latent space holes that the algorithm couldn’t make faces at all since every producable face was a certain “distance” away from a restriction.

Sure, you can make lora training on unconsenting people illegal and also make particular prompts illegal, but there is enough raw data and vague prompts to mold a generation into something that looks like it was done the illegal way without breaking either of those two restrictions.