Artists File Class-Action Lawsuit Against Stability AI, DeviantArt, and Midjourney (technollama.co.uk) 140
An anonymous reader shares a report: What many of us had expected has finally happened, artist have sued for copyright infringement a couple of AI companies, as well as an art repository site {complaint here (PDF)}. Is this the end of AI tools? I don't think so, I'll try to explain why, this will not be a detailed look at the lawsuit, there will be more time for that, this is my own take on some of the technical issues that I think the complaint gets wrong, so this is not intended as an in-depth look at the law anyway, as I suspect this may not get to a trial, more on that later. I'm also aware that this is at a very early stage, things may change, and most importantly, nobody can be sure of what the result will be, this is my own early speculation on the first filing as it stands, I'll update and write further blog posts as needed.
Three artists are starting a class-action lawsuit against Stability.ai, Midjourney, and DeviantArt alleging direct copyright infringement, vicarious copyright infringement, DMCA violations, publicity rights violation, and unfair competition. DeviantArt appears to be included as punishment for "betrayal of its artist community," so I will mostly ignore their part in this analysis for now. Specifically with regards to the copyright claims, the lawsuit alleges that Stability.ai and Midjourney have scraped the Internet to copy billions of works without permission, including works belonging to the claimants. They allege that these works are then stored by the defendants, and these copies are then used to produce derivative works. This is at the very core of the lawsuit. The complaint is very clear that the resulting images produced by Stable Diffusion and Midjourney are not directly reproducing the works by the claimants, no evidence is presented of even a close reproduction of one of their works.
Three artists are starting a class-action lawsuit against Stability.ai, Midjourney, and DeviantArt alleging direct copyright infringement, vicarious copyright infringement, DMCA violations, publicity rights violation, and unfair competition. DeviantArt appears to be included as punishment for "betrayal of its artist community," so I will mostly ignore their part in this analysis for now. Specifically with regards to the copyright claims, the lawsuit alleges that Stability.ai and Midjourney have scraped the Internet to copy billions of works without permission, including works belonging to the claimants. They allege that these works are then stored by the defendants, and these copies are then used to produce derivative works. This is at the very core of the lawsuit. The complaint is very clear that the resulting images produced by Stable Diffusion and Midjourney are not directly reproducing the works by the claimants, no evidence is presented of even a close reproduction of one of their works.
Does this extend to all the viewers memories? (Score:2)
Do they want the right to neuralyze anyone after they see their works?
Re: (Score:2)
Do they want the right to neuralyze anyone after they see their works?
I would certainly think so. This is like claiming that the alphabet is copyrightable. As well, this is a full stop on any future art, because it's all been done, anything new is derivative
Re: (Score:2)
The problem is that "Viewers" know when they're stealing work if they steal it.
AI is like the guy in a trench coat in the alley. Are these counterfeits? Did his sister really take these photos or are they famous photographer's works stolen and printed and sold for profit? You don't know. Maybe many of them are his sister's photos and he is licensed to sell them. Maybe many of them are direct rip offs.
If it were a bunch of guys in trench coats that would be one thing. But these are large for-profit compan
Re: (Score:2)
Now that you mention it, I suspect it is provable that the image is vanishingly unlikely to closely match any of the training inputs, and I mean mathematically provable. In contrast a human artist can never prove a work is unique from everything they've seen before, because there is no record of all they have seen, nor any mathematical
Re: (Score:2)
That's a question I had for any copyright. Perhaps any lawyers here can explain legal definition of "copy" in this context?
One can argue that viewers memory is not a copy as it's more of processed impression (and who exactly know how the brains store it), but alas in the process of delivery the exact copies of the artwork are stored in multiple places - routers, bits on the wires, computer memory, browser cache etc.
Re: (Score:2)
Here's how 'copy' is defined in US copyright law:
"Copies" are material objects ... in which a work is fixed by any method now known or later developed, and from which the work can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device. The term "copies" includes the material object ... in which the work is first fixed.
As for 'fixation':
A work is "fixed" in a tangible medium of expression when its embodiment in a copy ... , by or under the authority of the author, is sufficiently permanent or stable to permit it to be perceived, reproduced, or otherwise communicated for a period of more than transitory duration. A work consisting of sounds, images, or both, that are being transmitted, is "fixed" for purposes of this title if a fixation of the work is being made simultaneously with its transmission.
There has been debate over the years about copies generated in the process of works going through computers. For a long time, about the only consideration of the subject was MAI Systems Corp. v. Peak Computer, Inc., 991 F.2d 511 (9th Cir. 1993), which was very bad -- if it was in RAM, well, that was a perceptible copy. Cartoon Network, LP v. CSC Holdings, Inc., 536 F.3d 121 (2nd Cir., 2008) fixed that a little by saying that the MAI court i
Artists should have to prove their own originality (Score:2, Insightful)
The artists in this lawsuit should have to prove their own originality as well. As in, they NEVER ONCE IN ANY WAY used their knowledge of prior work in their own art. All of the artists that had ever been to an art museum, or taken an art class, or looked at an art book should have their lawsuits thrown immediately.
Every single artist is building on and re-mixing their own knowledge of prior works. This is literally no different than what the AI content generators are doing.
Re: (Score:2)
Re: (Score:2)
A friend of mine who does work for a major comic book corporation tells me that while the source material used in some of the AI work is hard to spot, there are cases where parts of the original artist's signature are actually visible, if you know what to look for.
Re:Artists should have to prove their own original (Score:5, Informative)
This is false. Yes, it will sometimes draw "signatures", but they don't resemble anyone's actual signature. It just knows, "Hey, paintings often have some sort of squiggly thing in the corner, so I'm going to make up a squiggly thing in the corner".
These tools are not compositors. They don't cut and paste images together. Txt2img has no access to any images. I don't know how to put this in strong enough terms, but when someone downloads StableDiffusion onto their computer, they're not downloading billions of images comprising an amount of data best measured in petabytes. The checkpoints are only 4GB.
These tools work in a way that's like image recognition run in reverse. "Hey, in this random noise, you see this fuzzy shape that kinda looks like a horse? If I tweak it here and there, it'll be more recognizable as a horse."
Re: (Score:2)
The model doesn't store the bitmaps, but it still has a very good understanding of what some images look like. Ask it for the Mona Lisa for example.
Legally I suppose this would be like reproducing a work from memory rather than just physically copying something.
Re: (Score:2)
It's not just signatures, but watermarks as well. So you can be reasonably sure the art used to train the model was used without the artist's permission.
The crux of the issue is not that AI could replace artists, but that the work of artists was used without their consent to train their replacement, committing copyright infringement in the process. Artists relied on copyright law to protect them when they posted their work online.
Re: (Score:2)
but that the work of artists was used without their consent to train their replacement
They could just not share their art. But if they do share their art they can't be upset that someone (or some thing) learns how they drew it.
Re: (Score:2)
Hi, artist here.
I've played a bit with AI, and the source material is very easy to spot, especially if you use terms like neoclassical, romanticism, etc... If you enter "birth of venus, in the style of botticelli", some models will generate the actual painting, albeit with some of the elements mangled.
If you are familiar with the work of Hubert Robert, Leonardo, Poussin, etc, you can very clearly see the sources in what it generates. Hubert Robert painted the interior of the Louvre, and without menti
Re: (Score:2)
I'll take "Things You Can't Actually Do With AI Tools" for $1000!**
(** Within the caveats described here [stablediff...volous.com])
The irony is, while AI art tools DON'T do that, a lot of physical artists I know DO do that. A common practice is to collect "inspiration images", and then paint while looking at said images, and in a way the outcome is sort of a distorted collage of their inspiration images. The very thing that AI art tools DON'T do, that they accuse it of doi
Re: (Score:2)
Re: (Score:2)
No, you're not following: people don't do it, because it can't do it (within the above caveats).
Re: (Score:2)
The artists in this lawsuit should have to prove their own originality as well. As in, they NEVER ONCE IN ANY WAY used their knowledge of prior work in their own art. All of the artists that had ever been to an art museum, or taken an art class, or looked at an art book should have their lawsuits thrown immediately.
Every single artist is building on and re-mixing their own knowledge of prior works. This is literally no different than what the AI content generators are doing.
Exactly. If an exact same image is being claimed by another artist, that's one thing. But if an image is in a particular style, and looks similar, it's just new art.
It's like if I claimed Ansel Adams' picture of Yosemite was mine, yeah, I infringed. But I can take my own picture of Yosemite. The mountain is not his.
Same with other forms. If someone paints a nude woman, they don't own all works where she posed for others, or the physical positioning was similar by another woman, or the color palatte was
Re: (Score:2)
But I can take my own picture of Yosemite. The mountain is not his.
It's not quite so clear cut, at least in the UK:
https://www.copyrightuser.org/... [copyrightuser.org]
Re: (Score:2)
But I can take my own picture of Yosemite. The mountain is not his.
It's not quite so clear cut, at least in the UK: https://www.copyrightuser.org/... [copyrightuser.org]
Zoinks! Those two pictures in that case are different enough that the original owner of the copyright is the copyright "owner" of any red bus driving by Parliament own it, and if he declares it art, there shall be no more!
Almost as dumb as George Harrison found guilty of a song while his inspriation was a completely different song.
By the way - I have copyright of any and all instances of the letter A. ;^)
We'll not be having it any more!
Re: (Score:2)
That is easy.
1. Take an image generated by AI.
2. Give the same image to an AI and ask AI to make based on it.
3. Show sketches and original image to the Judge and claim that you have obviously drawn the original image, you even have original sketches to prove it.
4. Profit.
Old artist yells at cloud. (Score:2)
Re: (Score:2)
Then it's a good thing they're fighting the copyright infringement and not the actual use of the tool.
Re: (Score:2)
It's basically an entire career path that's going to just go away along with all the current jobs that go with it.
The current trend of automation is destroying jobs and a much faster rate than we can create them. Declining birth rates would help but th
Re: (Score:3)
This is way different. This is an entire vocation going away almost overnight.
There used to be a whole army of people who walked around town every evening to light the streetlamps. Perhaps if they'd sued the Edison company, they'd still be employed.
Re: (Score:2)
This is way different. This is an entire vocation going away almost overnight.
Publishing is rife with this. Once upon a time there were writers, copy editors, typesetters, compositors, proofreaders, camera operators, strippers (i.e. the people who arrange and join together film negatives to make lithographic plates, get your mind out of the gutter), platemakers, and printers. Now half of those jobs are gone -- writers double as typesetters, copy editors double as proofreaders, and prepress technicians do the rest and in many cases can print directly to the press, or at least direct
Article gets it wrong (Score:5, Insightful)
The article goes way overboard in describing the technical details of the model and process, confusing the issue.
Copyright law exists to allow creators to control copies of their work... period. The exceptions are what are really at issue here: Do the defendant's actions fall within a recognized exception to copyright law?
The defendant's process appears to be:
1. Scrape copyrighted images off the internet
2. Use them as data / inputs into a machine learning model
3. Produce some output [the plaintiff characterizes these as derivative works, but that is very much unclear]
The article focuses on points 2. and 3. However IMO the issue is point 1.
-Defendant creates copies in their 'Downloads' folder as a necessary part of their process. This is a violation of copyright law. It may seem silly but it is true. It is just as illegal as downloading an Mp3 of a copyrighted song, for instance.
-Now, is there an exception to the blanket rule of copyright? The defendant will certainly argue Fair Use: for an academic purpose the images are temporarily copied/used and then deleted.
Whether or not this use is Fair Use is the meat and potatoes of this whole dispute. I think a good argument could be made either way, but here is the argument for why it is not. The defendant's use is not the narrowest possible use... they could, in theory, only use public domain images... there is no reason to use any specific copyrighted image, except for their convenience. Because the ultimate goal is profit seeking (producing so-called AI Art) this is a factor that cuts against fair use.
Re: (Score:2)
No it isn't. These images are published. You're *supposed* to download them. The artists' beef is about what you do with them after you've downloaded them. Copyright law clearly makes it illegal to republish the images, publicly display them, stuff like that, but everything else gets into dangerous territory.
Re: (Score:2)
I may have missed something... aren't we talking about non-public domain images, such as the ones on Deviant Art? Just because something is 'published' doesn't mean it can be freely copied. Just because individuals *can* download them with little consequence doesn't mean it's technically legal to do so. There's almost always fine print stipulating that the images are copyrighted and what can and can't be done with them.
Re: (Score:3)
Its not copying, you are free to view a publicly available image copyrighted or not, and take inspiration from it. If you don't want people or AI viewing your image don't put it on the internet ever.
The fact that there are people that think they can control what someone, or something can do with information after its publicly available shows that copyright has gone way too far.
Re: (Score:2)
If it's on the Internet it's published. If you go to the DeviantArt website, you are downloading the images, otherwise you wouldn't be able to see them.
"Public domain" means you can do whatever you want with it. That's not the same thing as published.
Re: (Score:2)
Re: (Score:2)
You have downloaded them when you view it, in order to view you need to download. If you mean keep a copy on your device for an extend period of time, then sure, but then again the AI isn't keeping a copy either.
Re: (Score:2)
Ah Slashdot. How far you've come from your roots.
In order to view them, you have to download them. You might mean that you don't have the right to save them to some kind of more permanent storage medium. I doubt something like that is enforceable, and it's not necessary to use the images to train an AI model anyway.
Re: (Score:2)
Re: (Score:2)
Yeah, so you replied to my post (I'm the GP). That post was copyrighted, and I grant a license only to view and create derivative works if you agree with me. You did not agree with me, so you're in violation. I'm going to need you to delete your post and pay a penalty. Thanks.
It's possible someone could talk a court into ruling that copyright holders have that kind of fine-grained control over their works. I think it's a really bad idea.
Re: (Score:2)
It has to download it in order to show it to you at all.
Turn the cache off completely, and the downloading will not cease.
Re: (Score:2)
That alone does not change the act - it gets copied. It was explained elsewhere here that it's the intention of use that matters here not the act. However, since browser must copy to display, the AI must copy to train. There is not intention to copy it "permanently" for later or otherwise. Now, the claim eludes that a "compressed" copy if stored in AI's NN parameters. This will be interesting to see litigated.
Re: (Score:2)
Oh, it's definitely being copied into the training set. The issue is, is that an infringing use, or is it a fair use?
Once the AI is trained, though, the image is no longer present. If it were, it would be an amazing leap forward in compression software! Fuck Claude Shannon.
Re: (Score:2)
Agreed - the ultimate goal is to profit in some manner. Either directly by producing AI art for money, or building an AI that can be used to sell a service, or indirectly by improving the reputation of the organization, group, devleoper, etc.,
To me it falls apart at each of the 3 levels.
1 - they used copyrighted works without permission ... ...
2 - for an application the builds on the work contained in the copyrighted works
3 - that ultimate creates new works based on what was learned in step 2
By this very pr
Re: (Score:2)
They used it with permission, if the art is on the internet and available to everyone to view that is implicit permission to view and derive works from it. Just like an artist does. If these artists can claim that every work they ever looked at that was copyrighted gave them explicit permission to derive their works from then they might have a point.
The real reason they are worried that they might loose income from AI generated art. The rest is mental gymnastics to justify their position. While I feel for t
Re: (Score:2)
'Derivative work' is a term of art in copyright law; it doesn't simply mean 'derived from in any fashion.'
A "derivative work" is a work based upon one or more preexisting works, such as a translation, musical arrangement, dramatization, fictionalization, motion picture version, sound recording, art reproduction, abridgment, condensation, or any other form in which a work may be recast, transformed, or adapted. A work consisting of editorial revisions, annotations, elaborations, or other modifications which, as a whole, represent an original work of authorship, is a "derivative work".
I think the link is too tenuous here to support the contention that the AI-generated works are derivative works, especially when you separate the work from the artist's style, which is not protected by copyright law.
Re: (Score:2)
That is silly as it would implicate anyone who's using browser to look at original pictures as browsers make fairly permanent copy in their caches. Forget browser cache, to display the image it must reside in the computer memory. It is a copy, however temporary. So it is in case of training AI. I see no categorical difference.
Re: (Score:2)
Wrong! It very much goes to intent and enforceability.
The browser copy goes into a cache, automatically without any intent of the user. It is then presumably deleted automatically at some point. This is actually pretty different from downloading and saving onto the desktop, and everyone buys the argument that it is 'temporary' and necessary to look at the picture at all.
Increasingly, images on the internet have a lot of fine print stating what can and can't be done with them. They may give you a license
Re: (Score:2)
Defendant creates copies in their 'Downloads' folder as a necessary part of their process. This is a violation of copyright law. It may seem silly but it is true. It is just as illegal as downloading an Mp3 of a copyrighted song, for instance.
An 'illegal mp3' that the artist embeds on their webpage such that it is cached on my computer and starts playing automatically when I visit their website? And I would be liable for damages for my visit?
My brain also encodes a copy of the artwork when I view it. That copy likely then informs my creation of subsequent artworks (whether I will it or not). Is all artistic creation by any person who hasn't lived their entire life isolated in a cave fundamentally infringing?
People feel like artists are getting
Re: (Score:2)
Straw men and grasping at straws are both flawed arguments.
Intent matters. Browser cache copies that you do not control don't matter. Your brain does not (quite obviously and in general) matter.
I simply took a poorly written defense that misunderstood copyright law and pointed out the flaws.
I have no dog in this fight because I think "AI Art" is irrelevant.
Re: (Score:2)
When you view an image on the internet the users intent is to remember it otherwise why view it? Unless its forced on you like advertising. Why is you remembering an inaccurate copy different from the AI? If AI can generate new and interesting images then why shouldn't it? Oh yes someone generated 1 billionth of the input into the work they must be compensated.
Re: (Score:2)
Do note that copyright is a strict liability statute. Intent does not matter, except in calculating damages. The most innocent infringer in the world is still an infringer, and can still be required to pay compensation.
Re: (Score:2)
You're (presumably) smart enough to realize that my comment about intent was in response to a ridiculous hypothetical.
Re: (Score:3)
Except there's a slew of case law already about processing bulk data, including copyrighted data, for analysis. And spoiler alert: it's fine. You can do that. Google does it every day. Learning vastly more from works that it analyses, in terms of bytes per work, than AI art tools learn from works that they analyze.
As if Google isn't a profit-seeking entity? It's repeatedly won cases on th
Re: (Score:2)
The devil always lies in the definitions.
The existing case law fits those actions into fair use or some other exception.
It is definitely not settled whether 'Art' or 'Pictures' = raw data _in this specific case_ or whether they are unauthorized copies created and used without a license.
Therefore it is also not settled whether an exception to copyright applies. Feel free to write an amicus brief to the court explaining how everyone is wasting their time.
Re: (Score:2)
The attorney who filed this case literally believes that AI art tools work by looking up and mashing together latent images of the billions of images in the training dataset.
Latent images - which are already highly compressed data representations - are 64 bytes x 64 bytes x 4bits x 32 = 66kb each. Multiply by several billion. Apparently he thinks that if you want to install a few StableDiffusion checkpoints on your home PC you're going to need a petabyte of storage space.
If anybody is wasting the court's t
Re: (Score:2)
Thank you for boldfacing a properly used word; referring to reading a person's own unambiguous words as "mind-reading", and then immediately launching into namecalling, which is always the sign of an intelligent person who has the arguments on their side.
I think our time here has come to a middle.
Re:Article gets it wrong (Score:5, Insightful)
The defendant's process appears to be: 1. Scrape copyrighted images off the internet 2. Use them as data / inputs into a machine learning model 3. Produce some output [the plaintiff characterizes these as derivative works, but that is very much unclear]
Every single artist practices by reproducing (as exactly as possible) works by the greats on the field. It's how we learn technical skills, composition, etc. I have never met a working artist who hasn't sat in a gallery and sketched a Matisse or whomever to practice their skills and to understand visual things like light and shadow, composition, subject, etc better.
As an artist of more than 40 years, I wonder if I changed a few words if this makes it sound any less evil? Let's see ...
1. Look at hundreds of images in galleries and online
2. Use them as data / inputs into my brain
3. Produce some artwork based on the general rules I've learned and what I've seen that intrigues or influences me.
I think the issue is simply the scale at which a machine can do step 1, because that step is no different than what any artist does, just at a smaller scale. We've all learned our craft on the shoulders of giants and create work that in some ways are influenced, consciously or unconsciously, by what we've seen over our lifetimes of looking at art. It's a rare thing that anyone produces anything remotely "new" these days, sorry.
I wonder if that is the fundamental issue for artists unsettled by that technology? That it's technology reproducing what an artist does (viewing an being influenced) on such a scale? I don't know, carry on.
Re: (Score:2)
Re: (Score:3)
Unless the AI stores the exact image (which it doesn't) its flawed memory too.
Re: (Score:2)
I wonder if the difference is that there is a human-artist her vs AI being a machine (a sophisticated xerox hence)?
Re: (Score:2)
Whether or not this use is Fair Use is the meat and potatoes of this whole dispute.
This question has generally gone in favor of fair use, on the grounds that the use is transformative.
For example, Google Book Search stores an exact (ish, given OCR errors) copy of many books without permission. This is unavoidable in order to be a search engine for the contents of books. The courts, in Authors Guild v. Google, 804 F.3d 202 (2d Cir. 2015), held that the use for which the copying occurred was a fair one because it was transformative -- creating something new (Google for books) -- even thou
Re: (Score:2)
Thanks for posting! Perfectly illustrates my point. Every website has a TOS. The AI researchers are violating it simply by downloading these images. They're doing it on a scale large enough for someone to go after them
Re: (Score:2)
Every website has a TOS. The AI researchers are violating it simply by downloading these images.
TOS only matter if you have to agree to them ('click license') to visit and see the images (such as a click license when you sign up for the site, or when you first visit the site, etc.). If not then they aren't actually a contract, and thus only normal copyright law applies.
Re: (Score:2)
The problem here is that you as the artist must grant those rights to DeviantArt in order for them to transmit your image from their servers. You have to grant the right for derivative works for DeviantArt to resize your image for mobile devices. Everyone who signed up to DeviantArt believes this was the intent, and DeviantArt went out of their way to explain to prospective users that they retained control over their work.
Nobody believed that they were giving DeviantArt a license to train AI models on
Pro-AI agruement (Score:3)
I don't think so, I'll try to explain why, this will not be a detailed look at the lawsuit, there will be more time for that, this is my own take on some of the technical issues that I think the complaint gets wrong, so this is not intended as an in-depth look at the law anyway, as I suspect this may not get to a trial, more on that later. I'm also aware that this is at a very early stage, things may change, and most importantly, nobody can be sure of what the result will be, this is my own early speculation on the first filing as it stands, I'll update and write further blog posts as needed.
I can say one thing for sure. The summary was not written by AI. AI would not have rambled on about nothing. This has to be one of the worst cases of run-on sentences I've seen in a long time. I have a lot of tolerance for people who's first language is not English and are making a genuine effort at writing but this is not that. The irony is that the whole first paragraph could have been eliminated with no consequence.
Truly new (Score:3)
An AI trained by copying art... (Score:3)
If an AI is trained by copying art from a variety of artists, wouldn't that by definition mean that anything that the AI creates by definition is derived from the originals?
And if the AI model can't show what it took from each of the trained images to create what it's final image is, it would be necessary to assume that all artists' works could have influenced the AI, with some element copied in some in some manner, further reinforcing the idea that the AI's work is derived from the work of the original artists.
Finally, if the AI were not trained by anything at all, would the AI be able to create anything close to what it did with the artists' works? Again, the artists' works were necessary for the AI to create its art, again becoming a derivative work.
Consider Pharrell Williams being sued, and losing, for a song that the estate/family of Marvin Gaye claimed was a rip off of one of Marvin Gaye's songs. Pharrell acknowledged he was influenced by it.
This is an oversimplification of everything, but at its heart, given laws around derivative works in copyright, this could be a problem for AIs trained on copying lots of data. This is different than a human seeing it while out and about, here someone or some entity deliberately took copies of images and fed them to an algorithm. That's very different from a human seeing something and being influenced by it.
At first I thought fair use might be a good defense for the AIs, but while the work created is novel and it can be argued it is transformative, how much of the artists' works were used to train the AI. Was it complete images? If so, that may be another issue and it may make it more difficult to argue successfully about fair use.
It will be interesting to watch this court play out and to see what kind of precedent ultimately gets set.
Re: (Score:2)
Re: (Score:2)
If an AI is trained by copying art from a variety of artists, wouldn't that by definition mean that anything that the AI creates by definition is derived from the originals?
My understanding is that the U.S. Copyright Office does not consider an AI able to produce copyrightable works of art.
This would imply an AI would also be unable to produce derivative works, since a derivative work needs to contain some originality worth of its own copyright protection.
Stable Diffusion reproduces work (Score:2)
Re: (Score:3)
If a logo shows up over and over and over again in the training dataset, yes, it's going to learn it, in the same way it's going to learn anything else that shows up over and over and over again, whether it's a cat, George Washington, the Mona Lisa, or whatnot.
This is not relevant to the topic of a work showing up once in a dataset, among billions of other images.
If a work is greatly replicated through the training dataset and gets overtrained, and it's not something that should be overtrained (like the Mo
Good artist imitate (Score:4, Insightful)
"Good artists imitate, great artist steal" has been a motto in the art industry for a long while.
The AI just happens to do it better than others.
There are generative methods, where you can upload your own photos and ask this to be reinterpreted in "Monet" or "Picasso" style. The AI already has learned from libraries of all these artists and can imitate any style at will.
Is this any different than a very talented artist visiting all museums, and being commissioned to make imitation art on demand?
The AI "learns" from the masters. Yes, it has initially "seen" the art of all these artists. And if that was legal (i.e.: no "warez" art libraries were used), then it should be treated like any other being. Human or artificial does not matter.
Re: (Score:2)
Re: (Score:3)
Depends on what you mean.
Imitating styles is perfectly legal and does not violate copyright.
Imitating actual works (which the attorney has no actual examples of for his clients)? That depends on a lot of factors, including how transformative they are.
Transformative Use (Score:2)
One argument against this assertion could be that the use of images from the Internet for training AI image generation systems constitutes fair use under copyright law. Fair use is a doctrine that allows for the use of copyrighted material without permission for certain limited and transformative purposes, such as for the purpose of criticism, commentary, news reporting, teaching, scholarship, or research. The use of images for training an AI system could be considered a transformative use, as the images ar
My challenge to the plaintiffs. (Score:2)
Present a computer with Stable Diffusion installed on an otherwise vanilla OS. Give it no network connection.
Generate an image with SD to prove it's working.
Now, plaintiff, can you please show me even a single art work stored on the computer other than the original work it just generated?
In a just world, the plaintiffs will be made to pay the defendants' legal costs.
Can you see the original artwork in the final? (Score:2)
Re: (Score:2)
Re: (Score:2)
Did Andy Warholl violate copyright when he painted the Campell Soup design?
Re: (Score:2)
Well, only if it copied it.
If it independently made it, then it's fine.
Also, the software lacks agency, so really the question is what did the user do with it that caused it to output a similar or identical image? Sounds to me like it'll fall under the Sony rule; a technology that can be used to infringe, if it is capable of substantial noninfringing uses, is not itself contributing to infringement.
Biggest artistic upset this decade, no doubt ... (Score:2)
I don't want to try to predict what the LAW will decide here. But I find it really interesting how these AI programs, manipulating images freely accessible to them via web pages, have managed to anger and upset the majority of artists and make them fear for their future relevancy.
I feel like the "best solution" would be to develop a standard where artists can tag their images online with their name and creation date (and official name of the work, if they have one). Then, these AI programs should be require
Notion of "compressed" form? (Score:2)
The text of the complain has a notion of "stored compressed images". It eludes that the system processes the images and stores their representation in a "compressed" form.
At what point compression becomes another form? If I created a text recipe to paint a copyrighted picture would it also be a compression? I suppose jpeg compression does not create derivative work as it possibly reproduces the work with human-indistinguishable quality and creates nothing new but my text description is different.
Re: (Score:2)
Re:A fair middle ground? (Score:4, Interesting)
IMHO, this is why it's so frustrating to me to see anti-AI artists focus on StableDiffusion in particular. DALL-E, Midjourney, etc are authoritarian. They make the rules. They decide who can play. Everything is closed source, hidden, and pay-to-play. For an OS analogy, they're Windows, while StableDiffusion is Linux. Free. Open. Unlimited. The only license terms can basically be summed up as your standard open source license terms plus "I agree not to do evil stuff". Yet it's SD that gets most of the hate. I find this so weird.
Even if your works are in the training dataset, it can't recreate your works, either in whole or in part, even if the user uses the exact same label as in training. Exception: if it's "overtrained" to your works, for example, if your works don't exist just once in the dataset but thousands of times (like, say, the Mona Lisa or whatnot); then it can actually learn to reproduce specific works (in the same way it learns how to reproduce, say, a cat - it doesn't know the difference between an object, a painting, a style, or whatever). That extra learned information of course comes at the cost of learning from the other images in the dataset, so - with limited exceptions (like the Mona Lisa) - nobody wants any particular image or images to be overtrained.
These tools learn about 1 byte's worth of information from each training image, give or take. So your whole portfolio might be a tweet or two's worth of information. The reason they can learn styles so well with such little information is, frankly, because artists' aren't very original (sorry!). Nobody is working in a vacuum. Your work already resembles other artists. You invoke motifs that it already knows. It doesn't have to learn how to draw a dragon just because you draw a dragon; it already knows what a dragon looks like. All it needs to know is what's the same or different about you and your style vs. others. Including your work in the dataset just helps associate these preexisting concepts in the network with tokens related to labels that mention you, as well as to - ever so slightly, to a minuscule degree - helping to train the concepts better.
Careful here not to confuse copyright with trademark. Superhero characters are trademarked. Copyright applies to specific works; it does not protect things like characters.
Warner Brothers could certainly sue Midjourney for trademark violation, but I doubt they'd have success, whereas I think they'd have good success suing an actual user who tried to create an image that violated their trademarks. If you type "Captain America fighting The Hulk" into Midjourney, there's no question what you were trying to do. The thing is, it's pretty easy to deliberately violate trademark with any tool. Nobody is suing Adobe because even an amateur can use it to quickly draw Mickey Mouse.
The other here is that, to be honest, companies like Disney, Warner Brothers, etc are salivating over these tools. Indeed, even if Midjourney, DALL-E, SD, etc all disappeared overnight, they'd just train their own model on their own content and use it to make more at well. They'd still need artists - the outputs usually aren't professional quality right out of the box, and usually need touchup - but it would (and reportedly will) significantly lower their production costs.
Re: (Score:2)
Re: (Score:2)
Yeah, nah. [howtodraw.pics]
Re: (Score:2)
Re:Original work (Score:5, Informative)
This isn't even remotely accurate. And the attorney's "understanding" of how the software works is just as laughably bad.
Rebuttal site to the lawsuit: http://www.stablediffusionfrivolous.com/ [stablediff...volous.com]
AI tools are NOT collagers. They're denoisers, and what they do is basically like seeing shapes in clouds. Applied image recognition, thrown into reverse: "this fuzzy shape is kind of like an X. How can I change it to make it look more recognizable as an X?" And like image recognition, built on statistics, accumulated across billions of images, millions of instances of common types of "X".
In a generated image, nowhere are you looking at anything that came from one image, or a handful of images. Everywhere, you're looking at generation based on all images. Every individual image's contribution is vanishingly small. Like a byte or so per image on average. But collectively, they define both what is recognizable as "something", and thus conversely, how you can make random noise look more like said "something".
Re: (Score:2)
Re: (Score:3)
Yeah, try actually reading the above link.
Copyright is based on works. You have to show that a work borrows a meaningful amount from your work and uses it in a non-transformative manner. Styles are not copyrightable. "My painting contributed 1 byte or so to a 4 gigabyte checkpoint and they mentioned my name and made a painting that doesn't look like anything I've ever painted but it's in a similar style to how I paint so now I want money." That doesn't cut it.
Re:Original work (Score:4, Insightful)
as long as the artists can produce evidence of direct infringement, which is the whole point, and which they apparently can't.
i agree that discussing the term "collage" is pointless, but discussing semantics or the particulars of the actual process. no evidence of copy, no case.
but, eyah, who knows what will happen. i'm actually quite amused with how all this is ai thing is shaking current dogmas about intellectual property, this would be a good opportunity to finally abandon that fundamentalism but somehow i'm afraid we're only going to make it worse in the end.
Re: (Score:3)
The genie can't go back in the bottle. Stable Diffusion in particular guarantee that, by being open source software that anyone can run. Make it illegal, people are going to keep using it, and companies will continue to contract out to whoever can do jobs well and cheap - which will be "artists who use AI art tools to speed up their workflows". Furthermore, big IP holders like Disney and the like can just go train their own models on their own assets even if all the existing models get banned. And there's
Re: (Score:2)
AI art generators came from research using neural nets for computer vision. Interesting results were obtained when the neural nets tried to make sense of noise. There are theories that dreams and hallucinations in humans also arise from neural noise when the regular sensory inputs are shut down or impaired. Refinements train the net not just on images but also descriptions of the images such that given noise and a description, it can generate an image based on what it learned.
It's no more copying than if y
Re: (Score:2)
Image generation has basically two AIs competing against each other. First one tries to create an image of what ever we want, e.g. a fish. The second AI tries to identify AI generated noise from actual images of fish. If the second AI is fooled by the first AI, first AI gets points. If second AI is not fooled, it gets points. This way both AIs can get better when competing against each other.
This has the following effect:
1. Images drawn by the first AI are really original. You don't even need to show ANY im
Re: (Score:2)
That is correct for GANs, an earlier technology. Diffusion models work differently.
Re: (Score:2)
Re:Original work (Score:4, Interesting)
The concept of creativity is not the same as a calculation.
Facts not in evidence. We have no solid understanding of how creativity is created by the chemicals in our brains. Chemical reactions within silicon, which you are labeling calculations, may be just as capable of creating creativity as sodium reactions in your brains neurons. Until we even understand how our creativity works, let alone AI creativity, your statement is simply unproven.
It remains to be known if a calculation can ever be so complex that it becomes creativity
While not 100% proven, it is almost certain that artificial chemical reactions can create creativity because we already see it in humans. If creativity developed within our organic brains, AI will almost certainly get there eventually. And for all we know we may be very close.
AI may not, and in my opinion almost certainly won't, develop creativity in the exact same way our organic brains have, even discounting the fact that the chemical makeup of electronics is different than our organic brains. If AI researchers are rewarding creative behavior, the possibility of something similar to convergent evolution is likely. AI creativity and human creativity likely won't be the exact same, but the mechanisms which create this behavior could be what biologists call "analogous structures".
Re: (Score:2)
Formula for creativity is following:
1. Take bunch of existing work.
2. Pick two of them in random.
3. Combine them.
4. Add created work to list of existing works and go back to step 1 if needed.
Only way to create something completely new and original is to create something that is perfectly random and hope you get something good out of it. But perfectly random things are almost always just noise, so this method is pretty impossible for humans. And usually it is not necessary. As we can see from evolution, smal
Re: (Score:2)
The parameters are introduced by human beings, though.
I was able to generate imagery, then iterate on that more and more, to the point where an image was generated, which does not resemble prior art.
Re: (Score:2)
Re: (Score:2)
It does take quite some learning, and in order to get a good result (which does what you want), you need quire a few prompt parameters and iterations.
As for "creativity versus skill", that's a lengthy debate. But just for brevity purpose: I have been frustrated most of my life in that regard, because I do have creativity, but I lack the skill in most areas. I can definitely write, and I have written quite a few things, but I suck at drawing, for example. The images I have in my head were permanently locked
Re: (Score:2)
however it maybe an issue if the starting parameters included an specific atrist's style? and showed intent if you followed the brances that got yuo closer to the original artists work... meh idk maybe thats a case
Re: (Score:2)
Artists have been imitating or iterating on other artists' styles for centuries.
For example, a painter who specialized in natural landscapes and had a specific style could have his style "copied" by another painter who specializes in portraits or mech-art.