Arioch wrote: ↑Tue Aug 23, 2022 12:04 am
I very much doubt that they're doing image searches in real time; the site mentions that there are "automatic and manual" controls in place to try to prevent pornographic or graphically violent images from being used as source, so it's more likely that they are collected in batches and preprocessed. But the collection method isn't really relevant to the argument; the only place they could be sourcing images from is the Web.
Again, we may be talking past one another on this. But I have no disagreements with the fact that the training data comes from the web.
However, I would disagree with the idea that a "search" makes up any part of creating an output. I think once the model is trained, it just works with whatever correlations it managed to learn by studying the data.
I'd consider it more akin to the AI taking inspiration from a vast amount of training data rather than cordoning off a specific section and plagiarizing.
So, when one asks for a chair made out of broccoli, I don't think the AI is searching a database for broccoli and again for chair and then trying to make an image out of it's source material. I think the resulting image will be the result of whatever effect the entirety of the training data had on the neural net weights. This makes it incredibly unlikely that it will plagiarize, I would think.
That should be setting alarm bells off in your head.
Again, I don't agree with the assertion that originality requires a human level understanding of motives. That old google-art bot that just made fractal dogs probably had less "understanding" than Dall-E. Yet I don't think there's a single human on earth that would be ready to start a copyright case over it's work.
One of the examples shown on their website is of a famous Vermeer painting that DALL-E has made alternate versions of. None of them are exact copies of the original, but they are clearly done in Vermeer's style and they're obviously based on that particular painting. Now, if Johannes Vermeer was still alive, and someone was trying to commercially sell one of these DALL-E renders... do you think he'd be okay with that?
If you asked a human for particular Vermeer painting in the style Vermeer, they'd probably also be plagiarizing to get you wanted.
I'm not claiming that the AI is incapable of plagiarizing. If you ask it to plagiarize someone explicitly, it probably will. However, I just doubt that any innocuous request, like, for example, a texture pack, is likely to plagiarize anything. In fact, the fact that none of the Vermeer alternates were exact copies of the original probably hints at the fact that AI isn't doing anything shady when creating it's outputs, like ripping off lesser known artists.
Of course, if people decide to use Dall-E to plagirize Vermeer or anyone else's style. That would be plagirism plainly. But, that already happens with human artists, because humans plagirise, and some might use an AI to do it, but I don't it's fair to use this as proof that the AI is incapable of creating non-plagirized outputs.
And, I'll admit right now that there is a lot of contextual stickyness when it comes to the AI's output. I've got some stuff written up on this if you want to see it, but it will generally import a lot of implicit values along with whatever you explicitly ask for.
For instance, if you ask for a painting by someone from the middle ages, it's difficult to get figures that are wearing non-medieval clothing.
Again, this is a part of the charm, and what allows it to create such detailed outputs with short prompts. However, this shouldn't cause any more accidental plagiarism than humans alone are guilty of, as a cascade of related matters ends up making two works look more than similar than pure probability would dictate.
Again, none of this is to say that diffusion models are on par with, or will be on par with human artists. Understanding is a key point in creating consistent charachters that are depicted consistently from multiple angels and in relation to various objects. (Again, I have a lot of detailed examples where the AI does strange things with object relation, if you want some stories."
For example, I remember there was an AI that managed to "generally" become superhuman at every Atari game they threw at it except the temple of montezuma. This being because games like starfighter were basically reaction time based, and Temple of montezuma required a player to navigate a temple and collect keys and solve puzzles and head towards various goals and doorways that needed to be completed in a certain order... you get the point.
I still stick to my point that a human level of understanding isn't required for the creation of a non-plagirising AI, and this would especially be the case when it comes to things like backgrounds and texture assets and the like.