SoylentNews is people
SoylentNews
The AI software Stable Diffusion has a remarkable ability to turn text into images. When I asked the software to draw "Mickey Mouse in front of a McDonald's sign," for example, it generated the picture you see above.
Stable Diffusion can do this because it was trained on hundreds of millions of example images harvested from across the web. Some of these images were in the public domain or had been published under permissive licenses such as Creative Commons. Many others were not—and the world's artists and photographers aren't happy about it.
In January, three visual artists filed a class-action copyright lawsuit against Stability AI, the startup that created Stable Diffusion. In February, the image-licensing giant Getty filed a lawsuit of its own.
[...]
The plaintiffs in the class-action lawsuit describe Stable Diffusion as a "complex collage tool" that contains "compressed copies" of its training images. If this were true, the case would be a slam dunk for the plaintiffs.But experts say it's not true. Erik Wallace, a computer scientist at the University of California, Berkeley, told me in a phone interview that the lawsuit had "technical inaccuracies" and was "stretching the truth a lot." Wallace pointed out that Stable Diffusion is only a few gigabytes in size—far too small to contain compressed copies of all or even very many of its training images.
Related:
Ethical AI art generation? Adobe Firefly may be the answer. (20230324)
Paper: Stable Diffusion "Memorizes" Some Images, Sparking Privacy Concerns (20230206)
Getty Images Targets AI Firm For 'Copying' Photos (20230117)
Pixel Art Comes to Life: Fan Upgrades Classic MS-DOS Games With AI (20220904)
A Startup Wants to Democratize the Tech Behind DALL-E 2, Consequences be Damned (20220817)
(Score: 5, Interesting) by ledow on Tuesday April 04 2023, @12:30PM (5 children)
I can fit a lot of copyright-infringing images into "several gigabytes" without even trying.
As in, probably in the region of several million.
What will kill this is the same reasoning as is used to kill the worst kind of images - child exploitation images - when even having them in RAM briefly is considered "processing" and, by extension "possession" of the same images.
However you look at it, this is either a "creator" which is committing plagiarism of copyright material and passing it off as its own, or it's simply regurgitating images which are owned by other people.
https://waxy.org/2022/08/exploring-12-million-of-the-images-used-to-train-stable-diffusions-image-generator/ [waxy.org]
Also... legally it only needs to be storing just one representation of just one of those images against the licensor's wishes to be illegal, it doesn't need all the terabytes of source data to prove anything.
Fact is, they are stuffed coming or going, because their tool is generating copyrighted and trademarked data on demand, and if that was even a guy printing T-shirts he could be shutdown. A global reference service provided for free to the world? Good luck explaining that.
(Score: 0) by Anonymous Coward on Tuesday April 04 2023, @01:43PM
If "commercial AI" is hobbled in favor of open source "pirate AI", I'll take it as a good outcome. People already have their hands on SD-based models and they aren't letting go, ever.
(Score: 2) by ilsa on Tuesday April 04 2023, @04:31PM (3 children)
Agreed. If they try to push the case based on how the tech works, they lost before they even filed cause they dunno WTF they're talking about. The model doesn't contain any training data at all. It contains a construct that was generated from that training data.
If they do lose because of this, I hope it doesn't set a precedent that sets back future lawsuits that are more intelligently crafted.
The correct argument is that the training data uses data sources that they cannot reliably prove were public domain. Further, the system is not capable of unique results. It is only capable of producing an amalgam of what it was trained on, with no unique insights or variations. If the model was trained on one single picture of a woman, and you said "give me a picture of a woman", it will won't produce a picture of a generic woman, but that exact woman it was trained on, along with the clothing and the background of the image. It doesn't understand what a "woman" is, beyond the context that the image it trained on was tagged as "woman".
That means if they trained on copyrighted images, it is virtually guaranteed to reproduce those copyrighted images, which the Ars article demonstrates. Therefore, and again as Ars points out, the argument actually has nothing to do with AI per se, but of fair use.
(Score: 2, Touché) by Anonymous Coward on Tuesday April 04 2023, @05:59PM (2 children)
"My ZIP file doesn't contain any images at all! It contains a construct that was generated from the images!"
Since it has been shown that one can extract a replica of training images from this "construct", albeit with compression artefacts, how can it be claimed that's this "construct" doesn't contain the training data?
(Score: 3, Interesting) by ledow on Wednesday April 05 2023, @07:15AM (1 child)
It doesn't contain identical copies of the training data in the exact format provided, that much is likely true. The system doesn't "understand" JPEG or similar, and it likely isn't storing them like that at all.
But it doesn't need to in order to be copyright infringement.
Literally any "derivative work". And that's such a broad definition that it encompasses basically the entire "AI" training database and any resulting intermediary format and output whatsoever, if it's sufficiently close to the original data to be recognisable to a human (which is kind of the whole point, no?).
(Score: 0) by Anonymous Coward on Wednesday April 05 2023, @10:14AM
Yeah I mean we can already see some stuff here: https://youtu.be/Ok44otx90D4 [youtu.be]
See also: https://en.wikipedia.org/wiki/Rogers_v._Koons [wikipedia.org]
Where it doesn't even need to be a very close copy.