> 1. They seem of the opinion that the copyright question is open.
I'm surprised it has taken this long to be honest. I've seen generated images with the blurred Getty watermark on them.
I'm surprised it has taken this long to be honest. I've seen generated images with the blurred Getty watermark on them.
The model is only around 4GB, and it was trained on ~5B images. At 24 bit depth, that'd be 786k raw data per image, which would be 3.5 petabytes of uncompressed information in the full training set. Either the authors have invented the world's greatest compression algorithm, or the original image data isn't actually in the model.
So, I think the argument is: if you look at someone else's (copyrighted) work, and produce your own work incorporating style, composition, etc elements which you learned from their work, are you engaged in copyright infringement? IANAL but I think the answer is "no" - you would have to try to reproduce the actual work to be engaging in copyright infringement, and not only do these models not do that, it would be extremely hard to get them to do so without feeding them the actual copyrighted work as an input to the inference procedure.
The blurred watermark is what makes its obvious they used Getty's (copyrighted?) images to train the model.
But the authors have the exclusive right to making more copies of their work, so if you'd want to make a thousand of these frankensteinbooks, you would need to get a thousand copies of the original books.
The images in Getty are provided with the intent to be used only as a catalogue for purchasing corresponding images without watermarks.
The difference in intent is very clear, and a judge would make a distinction between these.
Copyright can be fairly flexible in interpreting what constitutes a derivative work. The Getty water is evidence that an image belongs to Getty. If someone produces an image with the watermark and gets sued, they could say "your honor, I know it looks like I copied that image but let's consider the details of how my hypercomplex whatsit work..." and then judge, say a nontechnical person, looks at the defendant and say "no, just no, the court isn't going to look at those details, how could court do that?". Or maybe the court would consider it only if you paid 1000 neutral lawyer-programmers to come up with a judgement, at a cost of millions or billions per case.
Another scenario, what if i create a painting of a river by hand in acrylic and also draw a getty-watermark-looking thing on top using acrylic? As for why, i would put it there as an integral part of the piece, to allude to the fact of how corporations got their hands over even the purest things that have nothing to do with them, with the fake watermark in acrylic symbolizing it. You can make up any other reason, this is just the one i thought of as i was writing this. It wont look exactly like the real getty watermark, it will be acrylic and drawn by hand, so pretty uneven with colors being off and way less detailed. Doesn't feel like fraud to me.
What if i create a painting of a river by hand in acrylic and also draw a getty-watermark-looking thing on top using acrylic? As for why, i would put it there as an integral part of the piece, to allude to the fact of how corporations got their hands over even the purest things that have nothing to do with them, with the fake watermark in acrylic symbolizing it.
A human artist might well do that and make that defense in court. For all anyone knows, some GPT-3-derived-thing might go through such a thought process also (though it seems unlikely). However, the GPT-3-derived-thing can't testify in court concerning it's intent and that produces problems. And it's difficult for anyone to make this claim for it.
Edit: Also, if instead of a single work (of parody), you produced a series of your own stock photos, used the Getty Watermark and invited people to use them for stock photo purposes, then your use of the copyrighted Getty Watermark would no longer fall under the parody exception for fair use.
Even if it was plausible to believe the mechanics of how human brains process art was particularly similar to a diffusion model or GAN, I don't see "but human brains are deterministic functions of their inputs too" as being a successful legal argument any time soon. You'd have to throw out rather more of the legal system than just copyright if those arguments start to prevail...
You just described how modern browser cache images.
The diffusion model is revolutionary at scale, but doesn't mean it is doing anything drastically different than what is allowed right now, e.g. any AI based image beautifying/denoise filter, just the scale changes everything.
In art is very common that some author review their own work and made several paintings of the same subject. Artists made several attempts to conquer a panting, they draw studios or use different mediums. Photographers reproduce one portrait 20 years later to see how the subject changed. If you insert an IA image in the middle and copyright it, then any posterior painting on the same subject, even by the original author, would became derivative of the IA image that replaced it. This could even go so far as excluding the artists to review their most successful works.
In both cases, you take a series of bytes (the image data / the library source code) that is ultimately crucial to the functioning of your software, combine it with your own original code you wrote (training / compilation), and end up with a new output (the trained model / the binary executable) that is distinct from any of the original sources.
If you use a GPL'd library in your software, then it's uncontroversial to say that you have to follow the terms of the GPL. You can't say "well actually, the compiler is just reading your source code and learning what sort of binary it should produce, just like a human learns by studying source code, so I actually don't have to follow your licensing terms". No one would buy that. You clearly used that library, so you have to obey whatever terms come along with it.
Why is it fine to ignore the licensing terms for image data you incorporate into your software, but not third-party source code that you incorporate?
What if I read the code, understand its concepts, and re-implements another library that provide similar functionality without directly linking to the original repository, that is not an infringement, and actually how open source community has always been operating, like MariaDB to MySQL, or any projects that markets themselves as 'open source alternative' of some commercial software.
I would argue, the diffusion models are really good, it is possible that they capture the essentials of drawing that they learn no different than a human. Put in another way, it masters the imaging process at fundamental level
I agree, that's fine.
The analogous situation with image generators would be if the companies that trained the models had a human artist look at every image in their dataset, paint a unique but similar image, and then feed all of those images into the model, so that no copyrighted images were used in training without permission. But that's obviously not what they did. They just fed in the images unaltered, without getting permission.
It might be… which is why famously wine developers don't look at leaked windows code.
The contested issue is whether training a model requires permission from the copyright holder, because for most ways of using a copyrighted work - all uses except those where copyright law explicitly asserts that copyright holders have exclusive rights - no permission is needed.
A quick demo: I fed in the prompts "a getty watermark", "an image with a getty watermark", and "getty", and it spat out these: https://imgur.com/a/mKeFECG - not a watermark to be seen (though lots of water).
I was then able to generate an obviously-not-a-stock photo containing something approximating a Getty watermark, with the prompt "++++(stock photo) of a sea monster, art": https://imgur.com/a/mNC6XtQ - the heavily forced attention on the "stock photo" forces the model to say "okay, fine, what's something that means stock photo? I'll add this splorch of white that's kinda like what I've seen in a lot of things tagged as stock photos" and it incorporates that into the image as a to satisfy the prompt.
We can easily recognize that as attempting to mimic the Getty watermark, but it's not clearly recognizable as the mark itself, nor is the image likely to resemble much of anything in Getty's library.
The court delivers the judgement, not the model.
If courts can find against musicians whilst accepting they 'unconsciously' plagiarised key elements of a song in their own completely different song played by different musicians based on maybe hearing it in the background somewhere, they can certainly find against the creators of a model which has a sufficiently strong and obvious dependency on Getty IP they imported to output reasonably close approximations of Getty watermarks.
1: If that watermark is recognizable enough to be a getty watermark or something whose shape vaguely looks like a watermark. 2. Where that watermark is coming from.
From how the model is trained, it is possible that the model considers the watermark itself a style of the picture and mimics it. But it would be mission impossible to trace to a particular work that the inspiration is coming from.
AI and finding the best compression algorithm for an input are essentially the same problem.
I'm not sure what the statute of limitations is for infringement, but I bet it is two years or more.