It's pretty impressive that Stable Diffusion can compress over 200TB of already previously compressed images from LAION 5B down to a few gigabytes! And that it can search that database so quickly to copy and paste things from the right images! /s
There's a lot of good discussion to be had around the ethics of AI art, training on copyrighted materials, etc. But it is equivocally not just copy pasting from a database of images.
This isn’t true, but it’s 4am, and I regret that I can’t type out a full rebuttal on my phone.
One obvious counterexample is stylegan interpolations. If you interpolate between two images, the midpoint is usually unique — it’s often not obvious what the source material was. (E.g. Gwern’s anime interpolations; sure, they’re anime faces, but from where? “All of danbooru” might as well be “all styles of anime ever created.”)
Go to your favorite image search engine and try to replicate anything AI has generated. It's impossible. The results you get from image search are nowhere near as specific as what the AI produces. It's not even close. The only time image search can compete is when you are highly specific, e.g. "painting of the Mono Lisa", both AI and image search will produce very similar results. But for a generic prompt, image search will come up with nothing that gets even close to the query, while AI can produce a highly specific image.
DreamBooth really should have destroyed all doubt about this point point, as the AI can generated highly specific images of subjects that aren't even in the original training set.
Comments
"AI" is only copy-pasting elements from a database of existing images. It's effectively just a fancy image search engine.
You can use it to make procedurally-generated art, but it's still obvious what the source material was.
It's pretty impressive that Stable Diffusion can compress over 200TB of already previously compressed images from LAION 5B down to a few gigabytes! And that it can search that database so quickly to copy and paste things from the right images! /s
There's a lot of good discussion to be had around the ethics of AI art, training on copyrighted materials, etc. But it is equivocally not just copy pasting from a database of images.
Yes, neural network compression is, indeed, very impressive.
Just because you put a lossy compression step in the middle doesn't mean it's not copy-paste anymore.
This isn’t true, but it’s 4am, and I regret that I can’t type out a full rebuttal on my phone.
One obvious counterexample is stylegan interpolations. If you interpolate between two images, the midpoint is usually unique — it’s often not obvious what the source material was. (E.g. Gwern’s anime interpolations; sure, they’re anime faces, but from where? “All of danbooru” might as well be “all styles of anime ever created.”)
Maybe someone else can argue the point further.
Go to your favorite image search engine and try to replicate anything AI has generated. It's impossible. The results you get from image search are nowhere near as specific as what the AI produces. It's not even close. The only time image search can compete is when you are highly specific, e.g. "painting of the Mono Lisa", both AI and image search will produce very similar results. But for a generic prompt, image search will come up with nothing that gets even close to the query, while AI can produce a highly specific image.
DreamBooth really should have destroyed all doubt about this point point, as the AI can generated highly specific images of subjects that aren't even in the original training set.
I haven't seen any concrete examples of this yet.