There's several fundamental fallacies in your argument here:
Firstly, "AI" is a very strong word for the algorithmic lovechild of blind watchmaker programs and image compression. There is no actual thought going on there, simply an impressively efficient (if rather lossy) database with an integrated search engine/parser attached to it. This itself is an impressive innovation- the genetic algorithms here have found a way to render down images to a form can be easily stored and then reconstituted in a useful way.
That leads us to the second main issue- Describing what these neural networks are doing as cutting and pasting is a straw argument. The actual process is more complicated, more akin to a very fancy mixture of tracing, texture interpolation, rotoscoping, and what artists call photobashing. Photobashing involves the manipulation of existing images as a foundation for another work, the name derived from the crudest form involving "smashing images together" like a collage to create a composite image. The more advanced form involves painting over top of an existing image or collage of images, drawing forms, textures, and colors from this base. It can make for very impressive-looking results but is easily spotted by an experienced artist who knows what they're looking for- and it has an infamous dark side where many artists have been caught using photos under copyright or other artist's works instead of personal or properly licensed stock. These programs blatantly do exactly this in ways that are extremely obvious to an artist's eye, even with the stable diffusion algorithm's data storage format inherently helping to disguise this sort of image stitching.
That leads into the third problem- It may not be literally tracing an image, but when a program uses minor variations of the same silhouette for every image of a horse it produces there is clearly a template it has distilled from its input. If the template was created by tracing many works and averaging them together, that makes it worse, not better- its creator has simply had it carry out automated plagiarism many times.
We artists object not because this is something new and scary, but a sort of fraud we are very familiar with in a new form.
I disagree with pretty much everything. Emad himself declared Stable Diffusion to be art history compressed into 4GB, and 'cut n paste' is mentioned nowhere in the complaint. It constantly states SD as a *new form* of collage tool, which is essentially what it is, except the collage bits have been atomized into postmodern grey goo, which you can morph into new forms by prompting.
Regarding webscraping: You do know that Google settled with Getty Images and had to make all kinds of adaptions to keep their image search, right?
The only thing that matters here is if the companies involved used unlicensed artworks in the creation of a commercial tool, and if the Fair Use defense applies to a machine that can mimic artistic style.
They did, and we'll see.
update: Oh, and you totally didn't mention the paper that proved that SD can put out training data verbatim: https://arxiv.org/abs/2212.03860
There's several fundamental fallacies in your argument here:
Firstly, "AI" is a very strong word for the algorithmic lovechild of blind watchmaker programs and image compression. There is no actual thought going on there, simply an impressively efficient (if rather lossy) database with an integrated search engine/parser attached to it. This itself is an impressive innovation- the genetic algorithms here have found a way to render down images to a form can be easily stored and then reconstituted in a useful way.
That leads us to the second main issue- Describing what these neural networks are doing as cutting and pasting is a straw argument. The actual process is more complicated, more akin to a very fancy mixture of tracing, texture interpolation, rotoscoping, and what artists call photobashing. Photobashing involves the manipulation of existing images as a foundation for another work, the name derived from the crudest form involving "smashing images together" like a collage to create a composite image. The more advanced form involves painting over top of an existing image or collage of images, drawing forms, textures, and colors from this base. It can make for very impressive-looking results but is easily spotted by an experienced artist who knows what they're looking for- and it has an infamous dark side where many artists have been caught using photos under copyright or other artist's works instead of personal or properly licensed stock. These programs blatantly do exactly this in ways that are extremely obvious to an artist's eye, even with the stable diffusion algorithm's data storage format inherently helping to disguise this sort of image stitching.
That leads into the third problem- It may not be literally tracing an image, but when a program uses minor variations of the same silhouette for every image of a horse it produces there is clearly a template it has distilled from its input. If the template was created by tracing many works and averaging them together, that makes it worse, not better- its creator has simply had it carry out automated plagiarism many times.
We artists object not because this is something new and scary, but a sort of fraud we are very familiar with in a new form.
I disagree with pretty much everything. Emad himself declared Stable Diffusion to be art history compressed into 4GB, and 'cut n paste' is mentioned nowhere in the complaint. It constantly states SD as a *new form* of collage tool, which is essentially what it is, except the collage bits have been atomized into postmodern grey goo, which you can morph into new forms by prompting.
Regarding webscraping: You do know that Google settled with Getty Images and had to make all kinds of adaptions to keep their image search, right?
I don't think that any of your arguments have merrit in court. As i wrote in my Substack: https://goodinternet.substack.com/p/the-image-synthesis-lawsuits-cometh
The only thing that matters here is if the companies involved used unlicensed artworks in the creation of a commercial tool, and if the Fair Use defense applies to a machine that can mimic artistic style.
They did, and we'll see.
update: Oh, and you totally didn't mention the paper that proved that SD can put out training data verbatim: https://arxiv.org/abs/2212.03860