> That makes no sense. The copyright holder has access to their own inventions, of course.
No, not talking about the copyright holder, I'm talking about the hypothetical individual(s) creating infringing copies. If those people need to already have a copy of the image to extract a copy of the image from the generative image model, then I'm saying the argument that the model itself is infringing seems weak. Or it's at least not an open-and-shut case.
>> A sufficiently advanced model could, in theory, generate any image. You could then, again in theory, find an embedding for any image. Does said model then infringe on all copyrighted images?
> Without the slightest doubt.
I'll continue to argue otherwise. This proposed model is not a compressed archive that reproduces a set of infringing works when decompressed. Instead, you already have to have a copy of an image to find an embedding. Otherwise, the chances of the model spitting out copies of infringing works is exceedingly improbable. (A program that outputs random noise also has a vastly improbable chance of spitting out a copyrighted work, but that's hardly keeping copyright holders up a night.) Furthermore, in being able to produce any image, the model is not going to contain every image, and provided a copyrighted image produced after the creation of the model, you could still find an embedding. From a copyright perspective, suing the creator of this model would be like suing someone over distributing an image of random noise, claiming that because you can find an "embedding" which produces your copyrighted work (really just the difference between the two images), the noise is infringing.
Now, if you want to sue someone for distributing an embedding into this model for infringement, that's another matter entirely. That makes perfect sense.
In reality, I acknowledge that models like Stable Diffusion are going to be a bit more muddy. There definitely is some overfitting going on, so some images are literally present. However, it's a case-by-case thing. Given the requirements (framed as an "attack" no less) for extracting those images, a particular release of SD might or might not be found to infringe. Other models, with better training and better datasets, could avoid the overfitting problem.
> The method of storing the information is pretty much irrelevant to copyright. Your link argument has been tried by pirates and it's not working too well, although it depends on the country and legislation.
Unless you agree that Pi itself is a copyright violation, I think you misunderstand. I'm not making the same "link" argument made by pirates. The index and length needed to find copyrighted embeddings in Pi is just a different encoding for the same data, similar to a compressed version of the same data, though I'm sure the Pi embedding would in fact tend to be absurdly larger than the original. Again, I'm saying that Pi isn't infringing here, but the Pi-rates with their "links" would be where the infringement happens.