/3/ - 3DCG

>>934312
>You cant sell AI fan art of an artist who holds copyright because you have to put their copyrighted work into the database and then remix it
Diffusion is not "remixing". Diffusion is a method that just statistically says how noise should be reduced with information from a text prompt. It takes white noise and passes over it several times, each time using math to figure out how the noise should best be reduced to match the prompt, based on a model. The model is basically what tells it what a square looks like, what a circle looks like, what a nose looks like, etc. But that model is not a collection of a bunch of different people's artworks, it's just a bunch of math formulas that are organized in a hierarchy. There is not a single shred of art in a Stable Diffusion model. It is NOT a database of art. There is no copy or remix of anything. None of the art that was fed into it was preserved, it was used to give the model statistical information about what things CAN look like. If you fed it a bunch of Picasso paintings it'd never be able to replicate Picasso because things like "nose" and "eye" lose all meaning when it comes to cubism and depend very heavily on context, which is something AI does not understand. It can generate generic anime waifus because they all look very similar and it's easy to statistically describe how a nose should look in anime style. Try telling it to replicate something as simple as Calvin and Hobbes, and it can't. Pic related is the closest thing I could generate, and it's still only close because it knows to associate "Calvin and Hobbes" with "stuffed animal" and "tiger". It is fully incapable of replicating the art style. Don't listen to the idiots on /g/ that think this will be the end of artists. It's a great tool for playing around and conceptualizing shit with, and I think it could genuinely harm the stock photo market, but artists really have nothing to worry about.