Comment Re:Derivative Works (Score 1) 149
The position of artists complaining about training models on copyrighted work would make a lot more sense, and have more bite, if they first understood that the model isn't being trained preferentially or purposefully with their copywritten works--there's no targeting here. At least as far as models like Stable Diffusion goes, they were just one small piece of a massive scrape of widely *publicly available* images. Intent matters when we talk about laws and ethics, and there's no intent to duplicate specific artists works or unique styles, at least with that model. When you take a corpus like what Stable Diffusion uses and train a model that's only... a few gigs?... in final size, there's nothing resembling the original art in there.
The reason we have nothing in law that prevents people from looking at art and being inspired by it (including copying and extending techniques, themes, etc) is because 1) we couldn't ever police that effectively and 2) it would result in some really negative chilling effects on art and creativity. We *all* build on what came before and don't need to provide specific reference for every inspiration ot learning, only the major and significant ones... and unfortunately for artists, their art really isn't a major or significant individual contributor to the end result. All works are derivative works.
If someone uses the tool to recreate your specific style or otherwise pretend to be you, that's one thing... and the book should be thrown at those people, as they're genuinely trying to profit off your work and ride on your coattails. But just because a piece of art coming out of Stable Diffusion *appears* like your style doesn't mean it's copying it at all... maybe, you just both found a similar aesthetic sweet spot.
