An IP attorney’s reading of the Stable Diffusion class action lawsuit (opens in new tab)

The LAION-5B dataset is metadata and URI pairs; all the images are publicly accessible on the Internet.

Stable Diffusion's U-Net is trained to remove noise from images in latent space, which the variational autoencoder (VAE) converts to and from pixel space. CLIP embeddings are used to improve the denoising step of the U-Net by using the correlations between human language descriptions of the pixel image to reduce latent noise. Neither the U-Net nor the VAR are trained to interpolate or reproduce images from the training set; if that happened the model would be overfitted and loss would be terrible on the validation set. The VAE is trained to produce a latent space that can accurately encode and decode any pixel image, and the U-Net is trained to remove gaussian noise from the latent space.

Stable Diffusion v2 16-bit is ~3GB of data. It was trained on hundreds of millions of images (minimum of 170M in the 512x512 step alone). That leaves a maximum of ~20 bytes per image that could conceivably be a copy, which is certainly not enough to directly reproduce either the style or contents of any individual image.

There is no artwork included in Stable Diffusion. There is a semantic representation of how images are composed of varied subjects represented in the latent space and what pixel probabilities over those subjects relate to human language phrases during decoding, and finally a method to remove noise from the semantic representation, e.g. starting with a blank or random canvas and interpreting what may be there, iteratively guided by CLIP embeddings. If you give Stable Diffusion an empty CLIP embedding you get a random human-interpretable image obeying the distribution of the learned latent space.

Where is the form to remove my reddit comments from chat gpt training data? Or my blog posts from gpt training data? I have a paragraph on the Internet that someone read and got an idea - I want my royalties.

These artists complaints are ridiculous, and are being made by people who don’t understand how things work.

If some other person draws a picture in their “style”, no one has to ask permission. That’s not a thing.

They either don’t understand how it works or they are just upset that a computer can make art as good as (or better than) they can in a fraction of the time.

All knowledge workers and creatives are going to face this in the future. It’s going to suck, but it would be great if we all could try to understand reality first.

Sounds like an opt-out dark pattern. US law is unbelievably aggressive when it comes to issues of copyright, and makes copyright itself opt-out, i.e. everything you produce is copyrighted, and you have to license it in order to remove that automatic copyright. But when it comes to building these models to reproduce imitations of other people's work, suddenly copyright gets loosey-goosey.

Notice that it's a legal posture that implicitly condemns Copilot, which ignores explicitly formulated opt-outs in the form of licensing.

The models are already released. They can't retroactively censor released models.

> This reads like damage control to me in an effort to temporarily silence the loudest critics.

I think it is to avoid any common law wrongs related to the publicity rights of the defendants. It seems like something that a legal team would flag as an unnecessary risk for the product. Removing their names and images from the training data doesn't impact the usefulness of the model while at the same time creating a much smaller surface area for collecting subpoenas.

They don't have to remove images from the training set, they're saying they're opting to do so, and using that as an argument as to why if there supposedly could be copyright infringement, they're not liable, because they allow it to be removed.

They could just as well not do anything and continue on - it's likely this case will be in defendents favor. Same as how Google can crawl the net, cache data, transform it, etc.

It seems highly foolish from a marketing perspective for the artist too.

I don't see how more people copying the artist style would not increase the value of originals.

A smart artist here should promote that their style is staying in the dataset. It is as good free publicity as they will ever get.

> Machines aren't people.

There's no reason why that is the significant detail. Why does it matter? If you can look at millions of images over your lifetime and faithfully reproduce famous works of art by hand, aren't you just as wrong?

They are never going to compensate the artists. It's cheaper to hire 1000 designers to make 100000 images of artistic styles they are going for

I compare it to reading a big stack of comic books and then trying to create your own, painstakingly reproducing the layout, color palettes and character styles. Or listening to (eg) Pink Floyd and then making a song that really matches their music. YouTube is full of those, it's really similar to SD images. But still not the same as sampling as that requires actually duplicating parts of the original.

The reason we have copyright for a limited time is to promote the arts, not to give people a monopoly on a specific style so they can milk it. (although that is what seems to be happening)

> No matter what, putting images into your machine then selling the output generated with them and not compensating the original creators is going to be seen as problematic. Machines aren't people.

What about a company where you submit images and it tells you which faces are in them?

> If the image is freely viewable (say you can browse to it), and you just look at it, are you violating any rights?

The fundamental issue with this line of argument is that it equates the process of human vision and the consequences of that with that of a computer program ingesting that image and the consequences of that.

This anthropomorphization seems like a form of deep fallacy when considering the nature and impact of AI software. In the case of "seeing" an image, the two processes could not be any more unlike each other, both in content and context.

>> If the image is freely viewable (say you can browse to it), and you just look at it, are you violating any rights?

If I read Harry Potter, then turn around a write a book about a wizard with a z-shaped scar? Who works at a school for wizards? With a pet owl? Who is an orphan? At some point I have started to violate intellectual property rules. (Ignoring all the Harry Potter material that was itself lifted from prior public domain art.)

AI systems aren't just reading, they are generating material based on the stuff they have read. They and the people controlling them have to abide the copyright rules just like any other "author".

> I don't understand how using an image as input to a model is a copyright infringement.

Look at the extreme case, then. What if that one image is your only input, and your output is identical to it? What if your output is your input reflected over the x-axis? What if your output just crops the input? What if your output is your input cut into irregular pieces and randomly rearranged? Which outputs violate copyright?

Slightly less extreme: suppose your input is two images, and your output is those two images next to each other in a single image? Or your output is the second image, reduced in size and placed in the center of the first? What if both of the inputs are human figures, and your output is to cut out the face and hands of one image and put it onto the other?

> images that are derivative of that original image, the same way a counterfeiter would make a copy of it.

Only one of these outputs are anything a counterfeiter would do. Are any of the others copyright-violating?

> If the image is freely viewable (say you can browse to it), and you just look at it, are you violating any rights?

This isn't the kind of question that the lawyers of the defendants are going to ask the court.

They'll more likely ask if it isn't clearly fair use similar to Sony v Universal and Authors Guild v Google and then present evidence of significant non-infringing commercial use.

> It seems that violation would only come if you would use the model to produce images that are derivative of that original image, the same way a counterfeiter would make a copy of it. Have the skill to copy is not the same as actually copying.

Yeah, that's basically how the courts see it these days although for a different reason. They don't ask questions about skills or work or anything like that. They ask questions like, "is this supposed infringing work a replacement in the market for the plaintiff's work?".

The deeper questions about what the hell anyone meant by the words in the Constitution about Copyright wait for the highest courts to get involved, which is where we got this nice division between tools and what the tools are used for which allows for innovative fair use of copying other people's protected works with tools like VCRs, online book search and large language models.

Think of it as this way:

in order to create 5 very different illustrations you need to talk with 5 people. in the end 5 people will get money when they finish with their work.

an AI consumes these artists past output and instead of paying to these artists it will gather income to the owner. So by using the output of 5 people who have spent decades on perfecting their craft, the AI generates income by stealing their work, and the money flows to the owner only, who doesn't give back anything to these people.

so in essence AI in this form kills income stream for humans, since it gives back nothing.

> If the image is freely viewable (say you can browse to it)

This doesn't mean anything. If an unsecured SSH server is connected to the internet that lets anyone who connects to it in and gives them a root shell, it is still illegal to 'hack' that machine. The law cares about intent, not technicalities.

edit: Since HN decided to break with "You're posting too fast. Please slow down. Thanks." again, banning me from replying: This is obviously just an example intended to show that the law cares more about intent than technical measures.

@dang Calm down dude.

> It seems that violation would only come if you would use the model to produce images that are derivative of that original image, the same way a counterfeiter would make a copy of it. Have the skill to copy is not the same as actually copying.

With this sort of model's "creation" process, is something close to everything it generates derivative of everything it ingested, since had you ingested a different set of images you'd presumably have a different model with different weights?

That's kinda sorta analogous to human creation, but a human can much more actively choose what to think about, what to ignore, what to filter out.

The human process involves an explicit creative judgement step that I don't think the image-generation-by-model process can - and that creative transformation is key, legally, to a derivative work being able to itself be copyrightable and to not be infringing.

Ultimately, this is just something that has to be solved with legislation, not a court case. It's too novel a setup for a court case to deal with under existing frameworks.

I think one issue is just that of scale. I personally tend to agree that there's something icky with just slurping up literally everyone's content, then producing a tool that will then proceed to put them out of business en masse. But proving that illegal under current law is certainly going to be a challenge.

I have not read the original complaint but it surprises me that the lawsuit doesn't have a much stronger focus on this aspect. Copyright law is very concerned about not destroying the market for a given work through infringement, but this is a case about destroying the market for entire artists at a stroke.

But that's a hard argument in court. There's no legal basis for claiming damages because the entire market itself is being destroyed.

Though I'm not sure it's any weaker than the other claims trying to be made. The basic problem is, this isn't illegal in any sense. I don't just mean "illegal in that it must be banned" but any level of gradation in between, in the licenses, in requiring compensation, in any sort of regulation whatsoever. Technology has simply outrun law again.

I don't.

Information comes with many different rights: copy-right is the right to make copies; "moral rights" were mentioned in a few of my UK job contracts and that's "the right to be identified as the author of a work"; database rights are for collections of statements of fact that are not eligible for copyright but which were deemed to be worth protecting anyway for much the same reasons.

Even if copyright is totally eliminated from law by the mere existence of these AI[0], we may well retain the aforementioned "moral rights". And even if it is totally legal, there's also a strong possibility of it being considered gauche to use an AI trained on the works of those that don't like this.

[0] https://kitsunesoftware.wordpress.com/2022/10/09/an-end-to-c...

An artist's style is not copyrightable so I doubt it makes much of a difference. My guess is that showing good faith will make the lawsuit go over easier, because there's nothing illegal about paying someone to copy someone else's style (and not just a replica).

I think this restriction was more about trying to shut up some very vocal people on social media and less about the law.

Copying an artists style is legal in every jurisdiction in which Stability operates.

There's a big difference between 'rework my drawing to look like it was painted by Goya' and 'render this drawing in the style of Lisa Frank' or any living visual artist famed for a specific identifiable style as opposed to a particular image.

Comics are one example of an area where individual artists might develop a large body of work in a very distinctive style. You probably know what a Tintin comic (by Belgian artist Hergé) looks like. And lots of Manga artists have very specific and instantly identifiable styles. Individual artistry is a little less obvious with popular western comics because the best-known titles tend to be superhero franchises where the characters/story world are owned by a corporation and individual artists come and go.

I'm not sure I understand the point you're making. Its disheartening that artists can opt-out of having a computer algorithm make derivative versions of their creations?

I'm probably on the opposite side of the fence. I do find it disheartening that it's opt-out instead of opt-in. The training set should be limited to public domain and CC-0 until such a time it can comply with attribution; then other CC works could be incorporated.

> Who knows how long will it take to progress the tech to the point where anyone will be able to train and run models unrestricted without dealing with lawyer nonsense.

These are orthogonal issues at this point.

The one concern I do have is that the “lawyer nonsense” (read: AI companies playing fast and loose with current laws) will stack the regulatory deck against AI technology unnecessarily - essentially because of an unforced error that brings negative attention to the technology.

Put another way, these companies are asking to have a spotlight put on them by being so flippant about copyright and ethics issues. This spotlight could have been avoided with better behavior, and the tech would still appear magical and remain one of the most impactful jumps in tech in decades.

Where are all these self trained artists who learned their craft in a bubble, devoid of outside influence from other artists? Is it because there's a better paintbrush now, or is it because that paintbrush is not the 'real' way?

This reminds me of the backlash against the wacom community on deviantart in the early days.

Rent seeking by owners of AI machines is OK, but not by copyright owners?

Strongly disagree, IP law (despite it's misuse by a certain mouse mascot'd company) is extremely important and protecting artists work and their livelihood.

The price floor on art commissions is already very low and AI effectively makes that cost zero, while providing zero compensation to the thousands of artists. Without their work, there's no Stability AI. From an ethical standpoint Stability is in the wrong, and from a legal one I think the class has a very strong case to recover damages.

I think this is why it was very important that their first release of the model weights (the v1.4 model) was and will continue to be very important. While training an entire model from scratch requires an insane amount of data and compute right now (and will continue to be out of reach of individuals for some years I think), fine-tuning the model can be done on a consumer graphics card (they call it Dreambooth)... and I believe we will see further refinements that allow more and more useful tuning and features being built by individuals on top that original model, making it more and more powerful for specific uses. So even if future efforts change the functionality, or nerf the output in some way, there will always be people developing tools on the original weights. That cat is out of the bag.

ChatGPT has shown that attempting to train the model to decline certain types of requests is of limited effectiveness and readily circumvented. The ability to make custom checkpoints and tools like Dreambooth will further limit these restrictions.

It would be a spectacularly shitty world where IP protection is only granted to entities with a legal budget that eclipses the GDP of Antigua, and not to smaller independent creators.

The ends of having an useful model like stable diffusion doesn't really justify just ignoring the IP rights of tens of thousands of creators who were already having a pretty rough time making ends meet. That's just a shitty thing to do.

I’ve heard this argument on numerous occasions but I have never heard someone justify it or why they believe it.

Are there specific similarities that make you believe these are equivalent scenarios? Not just “it feels thematically similar”.

I think there is a pretty big difference though. In the case of sampling you can play the original against the new media and show that they are 'the same'. I have no idea how you would go about doing something similar for Stable Diffusion. "Those three pixels look different when you remove image X from the training set" is probably not a convincing argument to anyone.

I'd rather not risk stunting progress in areas like language translation, malware scanning, DDOS prevention, spam filtering, product defect detection, scientific data analysis, autonomous vehicles, voice dictation, narration/text-to-speech engines, drug discovery, protein folding, optimization in production lines/agriculture/logistics, detecting seizures and falls, weather forecasting and early-warning systems, etc. just to let Getty Images have what they feel entitled to.

Best outcome in my opinion would be for the output to be judged on a case-by-case basis, like human works are, not for machine learning on data without "proper authorization from the owner" to inherently count as infringement.

I hope the exact opposite. AI, including AGI if we ever get there, cannot be allowed to be strangled in its crib by artificially limiting the information it can learn from in the name of IP maximalism. IP law already goes way too far, the line should be drawn here.

I hope this technology will become so ubiquitous that laws and "proper authorization" won't matter.

I agree that it should both be legislated, and substantially different to human learning. In the US (or UK) however, most legislators are such intellectual lightweights that they have no hope of grasping even the basics of what they are legislating.

> Diffusion models can be compared to a superhumanly talented artist that can be cloned in unlimited fashion by anyone having the software and hardware means.

How can you claim with a straight face that this is a better explanation of what an NN is?

An NN is simply an approximation of a multi-valued function, whose parameters are adjusted by minimizing the difference between the output of the NN and the output of the real function for a certain input. It is much much closer to "a giant archive of compressed images being used to interpolate between them" (though it's not that) than it is to a "superhumanly talented artist".

> the lawsuit will fail when the court hears an explanation from an expert

So how often does this happen? Somehow I'm too cynical to believe that a judge would rule against the intellectual property industry. The whole thing is based on absurd concepts to begin with, concepts that can be reduced to the ownership of unique numbers. Once a society accepts that, what difference do explanations make?

Is this requirement to register specifically a feature of the DMCA? It seems quite surprising if, as the article claims, “people who don’t have registered copyrights cannot enforce their copyrights in court.”

That would mean that the vast majority of artwork posted online is essentially free to exploit in the USA, since I’m sure most people do not routinely register their works with the copyright office before posting them.

Unless they're legally obligated to show proof of copyright registration for the takedown notice to be a valid, it would be risky to assume they didn't register it just because they didn't show proof.