Hacker News Reader

Comment by yazaddaruvala | Hacker News Reader

> That’s going to be hard to argue. Where are the copies?

In fairness, Diffusion is arguably a very complex entropy coding similar to Arithmetic/Huffman coding.

Given that copyright is protectable even on compressed/encrypted files, it seems fair that the “container of compressed bytes” (in this case the Diffusion model) does “contain” the original images no differently than a compressed folder of images contains the original images.

A lawyer/researcher would likely win this case if they re-create 90%ish of a single input image from the diffusion model with text input.

View on HN

There's a key difference. A compression algorithm is made to be reversible. The point of compressing an MP3 is to be able to decompress as much of the original audio signal as possible.

Stable Diffusion is not made to decompress the original and actually has no direct mechanism for decompressing any originals. The originals are not present. The only thing present is an embedding of key components of the original in a multi-dimensional latent space that also includes text.

This doesn't mean that the outputs of Stable Diffusion cannot be in violation of a copyright, it just means that the operator is going to have to direct the model towards a part of that text/image latent space that violates copyright in some manner... and that the operator of the model, when given an output that is in violation of copyright, is liable for publishing the image. Remember, it is not a violation of copyright to photocopy an image in your house... it's a violation when you publish that image!

Lossy compression isn't reversible but presumably the content when compressed tjis way is still covered by copyright.

Pedantically, yes, lossy compression is not 100 percent reversible. Practically, the usefulness of compression is that it does return the original content with as little loss as possible… so lossy compression is mostly reversible.

All of my other points remain unchanged by this pedantry.

You can't rip something and compress it badly enough to not violate copyright when you sell it. The point of compression is to throw away information about the original in ascending order of importance.

>You can't rip something and compress it badly enough to not violate copyright when you sell it.

While I doubt that specific case has been tested in court, arguably you could. If you created glitch art (https://en.wikipedia.org/wiki/Glitch_art) via compression artifacts, and your work was sufficiently distinct from the original work, I think you would have a reasonable case for transformative use (https://en.wikipedia.org/wiki/Transformative_use).

I'm pretty sure you were downvoted by someone who walks into MoMA and whispers to their partner that a three year old could have drawn that...

At what point does it become lossy enough that it's not protected, though? You can imagine a lossy compression algorithm that merely stores a 1 for images that are "more red" and a 0 for images that are "more blue." Such a compression algorithm would be storing some information about the thing it's compressing, but the closest reconstruction you could get from the compressed data is a red square or a blue square. Surely that's not copyright infringement? What about an algorithm that counts the fingers portrayed in an image and just reconstructs an image with the same amount of fingers? Where's the line?

Storing copies of training data is pretty much the definition of overfitting, right?

The data must be encoded with various levels of feature abstraction for this stuff to work at all. Much like humans learning art, if devoid of the input that makes human art interesting (life experience).

I think a more promising avenue for litigating AI plagiarism is to identify that the model understands some narrow slice of the solution space that contains copyrighted works, but is much weaker when you try to deviate from it. Then you could argue that the model has probably used that distinct work rather than learned a style or a category.

Even that approach seems highly vulnerable to fair use. If the model does not recreate a copyrighted work with enough fidelity to be recognized as such, then how can it be said to be in violation of copyright?

> 90%ish of a single input image

Oh, one image is enough to apply copyright as if it were a patent, to ban a process that makes original works most of the time?

The article authors say it works as a "collage tool" trying to minimise the composition and layout of the image as unimportant elements. At the same time forgetting that SD is changing textures as well, so it's a collage minus textures and composition?

Is there anything left to complain about? unless, by draw of luck, both layout and textures are very similar to a training image. But ensuring no close duplications are allowed should suffice.

But they are not original works, they are wholly derived works of the training data set. Take that data set away and the algorithm is unable to produce a single original pixel.

The fact that the derivation involves millions of works as opposed to a single one is immaterial for the copyright issue.

The training data set is indeed mandatory but that doesn't make the resulting model a derivative in itself. In fact the training is specifically made to remove derivatives.

Go to stablediffusionweb.com and enter "a person like biden" into the box. You will see a picture exactly like President Biden. That picture will have been derived from the trained images of Joe Biden. That cannot be in dispute.

Just because it generates you an image like Biden still does not make it a derivative either.

You can draw Biden yourself if you're talented and it's not considered a derivative of anything.

There is no need for rhetorical games. The actual issue is that Stable Diffusion does create derivatives of copyrighted works. In some cases the produced images contain pixel level details from the originals. [1]

[1] https://arxiv.org/pdf/2212.03860.pdf

> The actual issue is that Stable Diffusion does create derivatives of copyrighted works.

Nothing points to that, in fact even in this website they had to lie on how stablediffusion actually works, maybe a sign that their argument isn't really solid enough.

> [1] https://arxiv.org/pdf/2212.03860.pdf

You realize those are considered defects of the model right? Sure, this model isn't perfect and will be improved.

> You realize those are considered defects of the model right? Sure, this model isn't perfect.

You can call copying of input as a defect, but why are you simultaneously arguing that it doesn't occur?

I don't call these defects copying either but overfitting characteristics. Usually they are there because there's a massive amount of near-identical images.

It's both undesirable and not relevant to this kind of lawsuit.

Correction: if you draw a copy of Biden and it happens to overlap enough with someone’s copyright of a drawing or image of Biden, you did create a derivative (whether you knew it or not).

is that really how copyright law works? Drawing something similar independently is considered a derivative even if there's no links to it?

It's bad news for art websites themselves if that's the case...

No that’s not… at least in many countries. Unlike patents, “parallel creation” is allowed, this was fought out in case law over photography decades ago, because photographers would take images of the same subject, then someone else would, and they might incidentally capture a similar image for lots of reasons and thus before ubiquitous photography in our pockets, when you had to have expensive equipment or carefully control the lighting in a portraiture studio to get great results… well it happened and people sued like those with money to spare for lawyers are want to do, and thus precedent has been established for much of this. You don’t see it a lot outside photography but it’s not a new thing for art copyright law and I think the necessity of the user to provide their own input and get different outcomes outside of extremely sophisticated prompt editing… will be a significant fact in their favour.

The difference is that computers create perfect copies of images by default, people don't.

If a person creates a perfect copy of something it shows they have put thousands of hours of practice into training their skills and maybe dozens or even hundreds of hours into the replica.

When a computer generates a replica of something it's what it was designed to do. AI art is trying to replicate the human process, but it will always have the stink of "the computer could do this perfectly but we are telling it not to right now"

Take Chess as an example. We have Chess engines that can beat even the best human Chess players very consistently.

But we also have Chess engines designed to play against beginners, or at all levels of Chess play really.

We still have Human-only tournaments. Why? Why not allow a Chess Engine set to perform like a Grandmaster to compete in tournaments?

Because there would always be the suspicion that if it wins, it's because it cheated to play at above it's level when it needed to. Because that's always an option for a computer, to behave like a computer does.

You’re acting like the “computer” has a will of it’s own. Generating a perfect copy of an image would be a completely separate task from training a model for image generation.

There are no models I know of with the ability to generate an exact copy of an image from its training set unless it was solely trained on that image to the point it could. In that case I could argue the model’s purpose was to copy that image rather than learn concepts from a broad variety of images to the point it would be almost impossible to generate an exact copy.

I think a lot of the arguments revolving around AI image generators could benefit from the constituent parties reading up on how transformers work. It would at least make the criticisms more pointed and relevant, unlike the criticisms drawn in the linked article.

> There are no models I know of with the ability to generate an exact copy of an image from its training set

Is it "the model cannot possibly recreate an image from its training set perfectly" or is it "the model is extremely unlikely to recreate an image from its training set perfectly, but it could in theory"?

Because I am willing to bet it's the latter.

> You’re acting like the “computer” has a will of it’s own. Generating a perfect copy of an image would be a completely separate task from training a model for image generation.

Not my intent, of course I don't think computers have a will of their own. What I meant, obviously, is that it's always possible for a bad actor of a human to make the computer behave in a way that is detrimental to other humans and then justify it by saying "the computer did it, all I did is train the model".

In theory, you can:

- Open Microsoft Paint

- Make a blank 400 x 400 image

- Select a pixel and input an R,G,B value

- Repeat the last two steps

To reproduce a copyrighted work. I'm sure people have done this with e.g. pixel art images of copyrighted IP of Mario or Link. At 400x400, it would take 160,000 pixels to do this. At 1 second per pixel, a human being could do this in about a week.

Because people have the capability of doing this, and in fact we have proof that people have done so using tools such as MS paint, AND because it is unlikely but possible that someone could reproduce protected IP using such a method, should we ban Microsoft Paint, or the paint tool, or the ability to input raw RGB inputs?

>The difference is that computers create perfect copies of images by default

are we looking at the output of the same program? because all of the output images i look at have eyes looking in different direction and things of horror in place of hands or ears, and they feature glasses meting into people faces, and that's the good ones, the bad one have multiple arms contorting out of odd places while bent at unnatural angles.

Storing and retrieving photos, files, music, exactly identical to how they were before, is what computers do.

Save a photo on your computer, open it in a browser or photo viewer, you will get that photo. That is the default behavior of computers. That is not in dispute, is it?

All of this machine learning stuff is trying to get them to not do that. To actually create something new that no one actually stored on them.

Hope that clears up the misunderstanding.

You've made some errors in reasoning.

First, there is a legal definition of a "derivative work" and there is an artistic notion of a "derivative work". If the two of us both draw a picture of the Statue of Liberty, artistically we have both derived the drawing based on the original statue. However, neither of these drawings in relation to the original sculpture nor the other drawing is legally considered a derivative work.

Let's think about a cartoonish caricature of Joe Biden. What "makes up" Joe Biden?

https://www.youtube.com/watch?v=QRu0lUxxVF4

To what extent are these "constituent parts" present in every image of Joe Biden? All of them? Is the latent space not something that is instead hidden in all images of Joe Biden? Can an image of Joe Biden be made by anyone that is not derived from these "high order" characteristics of what is recognizable as Joe Biden across a number of different renderings from disparate individuals?

I can draw Biden, yes, but SD can only draw Biden by deriving it's output from the images on which it was trained. This is a simple tautology, because SD cannot draw Biden without having been trained on that data.

SD both creates derivative works and also sometimes creates pixel level copies from portions of the trained data.

Yes, and we are now using the artistic definition of “derived” and not the legal definition.

You cannot copyright “any image that resembles Joe Biden”.

This isn't about what can be copyrighted but that there are copyrighted images being used without following the legal requirements.

Can you draw Biden without ever having seen him or a picture of him? So,why is it that you are not deriving but SD is?

So is your mental image of Joe Biden, unless you know him personally.

If I were to take the first word from a thousand books and use it to write my own would I be guilty of copyright violations?

Words have a special carve out in copyright law / precedent. So much so that a whole other category of Intellectual Property exists called Trademarks to protect special words.

But back to your point “if you were to take the first sentence from a thousand books and use it in your own book”, then yes based on my understanding (I am not a lawyer) of copyright you would be in violation of IP laws.

I doubt it would be a violation.

Specifically fair use #3 "the amount and substantiality of the portion used in relation to the copyrighted work as a whole."

A sentence being a copyright violation would make every book review in the world illegal.

If I take a million copywritten images from magazines, cut them with scissors, and make a single collage, I would expect the resulting image to be fair use. Fair use is an affirmative defense, like self defense, where you justify your infringement.

People are treating this like its a binary technical decision. Either it is or isn't a violation. Reality is that things are spectrums and judges judge. SD will likely be treated like a remix that sampled copywritten work, but just a tiny bit of each work, and sufficiently transformed it to create a new work.

If I take a million copywritten images from magazines, cut them with scissors, and make a single collage, I would expect the resulting image to be fair use.

That’s not how it works. Your collage would be fine if it was the only one since you used magazines you bought. Where you’d get into trouble is if you started printing copies of your collage and distributing them. In that case you’d be producing derived works and be on the hook for paying for licenses from the original authors.

That’s not how fair use works. It’s not a binary switch where commercial derivatives automatically require licensing. Such a college would be ruled transformative and non competitive.

Me having bought the magazines also has nothing to do with anything. Would apply equally if they were gifted or free or stolen.

If I make software that randomly draws pixels on the screen then we can say for a fact that no copyrighted images were used.

If that software happens to output an image that is in violation of copyright then it is not the fault of the model. Also, if you ran this software in your home and did nothing with the image, then there's no violation of copyright either. It only becomes an issue when you choose to publish the image.

The key part of copyright is when someone publishes an image as their own. That they copy an image doesn't matter at all. It's what they DO with the image that matters!

The courts will most likely make a similar distinction between the model, the outputs of the model, and when an individual publishes the outputs of the model. This would be that the copyright violation occurs when an individual publishes an image.

Now, if tools like Stable Diffusion are constantly putting users at risk of unknowingly violating copyrights then this tool becomes less appealing. In this case it would make commercial sense to help users know when they are in violation of copyright. It would also make sense to update our copyright catalogues to facilitate these kinds of fingerprints.

how is that any different from new human artist that study other artists work to learn a style or technique. In fact it used to be that the preferred way for painters to learn was to repeatedly copy paintings of masters.

What you and many other in the thread seem to be oblivious about is that algorithms are not people. Yes, it may come as a shock to autistic engineers, but the fact that a machine can do something to what a person does does not warant it equal protection under the law.

It seems like you're using "autistic" as an insult here. If that's not your intention you might want to edit this comment to use different verbage.

What do you mean, autism is well established as a personality trait that diminishes empathy and the ability to understand other people's desires and emotions, while having a strong affinity to things, for example machines and algorithms.

Legislation is driven by people who are, on aggregate, not autistic. So it's entirely appropriate to presume that a person not understanding how that process works is indeed autistic, especially if they suggest machines are subjects of law by analogy with human beings.

It's not that autists are bad people, they are just outliers in the political spectrum, as you can see from the complete disconnect of up-voted AI-related comments on Hacker News, where autistic engineers are clearly over-represented, versus just about any venue where other professionals, such as painters or musicians, congregate. Just try to suggest to them that a corporation has the right to use their work for free and profit from it while leaving them unemployed, because the algorithm the corporation uses to exploit them is in some abstract sense similar to how their brain works. That position is so for out on the spectrum that presuming a personality peculiarity of the emitter is the absolutely most charitable interpretation.

This argument's pedantic and problematic for artists; take away a human's "dataset" and processes and they are also unable to produce a single original "pixel".

That is not true. The dataset is needed, the same way that examples are used by a person learning to draw. But the dataset alone is not capable of producing images not derived from any part of it (and there are many examples of SD results that seem so far to be wholly original), so you can’t reduce stable diffusion to being only derived from the dataset. It may “remember” and generate parts of images in the dataset - but that is a bug, not a feature. With enough prompt tweaking, it may even generate a fairly good copy of pre-existing work - which was what the prompt requested, so responsibility should lie on the prompt writer, not on SD.

But the fact that it often generates new content, that didn’t exist before, or at least doesn’t breach the limits of fair use, goes against the argument made in the lawsuit.

The model can generate original images, yes, and those images might be fair use. But it can also generate near verbatim copies of the source works or substantial parts thereof, so the model itself is not fair use, it's a wholly derivative work.

For example, if a publish a music remix tool with a massive database of existing music, creators might use to create collages that are original and fall under fair use. But the tool itself is not and requires permission from the rights owners.

So, is any sort of creation that relies upon copyrighted or patented works copyright infringement? Is any academic research or art that references brands or other creations illegal? This is such a clear case of fair use that it could be a textbook example.

> Oh, one image is enough to apply copyright as if it were a patent, to ban a process that makes original works most of the time?

The law can do whatever its writers want. The law is mutable, so the answer to your question is “maybe”.

Maybe SD will get outlawed for copyright reasons on a single image. The law and the courts have done sillier things.

All the handwringing about generative AI brings to mind the aphorism about genies returning to bottles. There can be lawsuits and laws--and there may even be cases where an output by chance or by tickling the input sufficiently looks very close to something in the training set. But anyone who thinks this technology will be banned in some manner is... mistaken.

So as a code author I am pretty upset about Copilot specifically, and it seems like SD is similar (hadn't heard before about DeviantArt doing the same as what GitHub did). But I agree with this take: the tech is here, it's going to be used, and it's not going to be shut down by a lawsuit. Nor should it, frankly.

What I object to is not the AI itself, or even that my code has been used to train it. It's the copyright for me but not for thee way that it's been deployed. Does GitHub/Microsoft's assertion that training sidesteps licensing apply to GitHub/Microsoft's own code? Do they want to allow (a hypothetical) FSFPilot to be trained on their proprietary source? Have they actually trained Copilot on their own source? If not, why not?

I published my source subject to a license, and the force of that license is provided by my copyright. I'm happy to find other ways of doing things, but it has to be equitable. I'm not simply ceding my authorship to the latest commercial content grab.

I doubt Microsoft sees fragments of Windows source code as a particular crown jewel these days. That said, some of it is decades old code that was intended for the public to see (unlike, presumably, anything in a public GitHub repository). And some of it is presumably third-party code licensed to Microsoft that was likewise never intended for public viewing. So, while it would be a good gesture on the part of Microsoft to scan their own code--if they haven't done so--I could see why it might be problematic. (Just as training on private GitHub repos would be.)

tl;dr I think there's a distinction between training on copyrighted but public content and private content.

Private third-party GitHub repos is another good example. If licenses don't apply to training data, as GitHub has asserted, why not use those too? Do they think they'll get in trouble over it? Why doesn't the same trouble apply to my publicly-readable GPL-licensed code?

I assume there's something in their terms of service about not poking around in private repos and using the code even for internal purposes except for necessary maintenance like backups, court orders, etc.

I am not a lawyer but I also assume Microsoft's position, at least in part, is that they can download and use code in GitHub public repos just like anyone else can and developing a public service based on training with that (and a lot of other) code isn't redistributing that code.

Copyright is not the only law. Something might be permitted by copyright law (as fair use, an implied license, etc)-yet simultaneously violate other laws-breach of contract, misappropriation of trade secrets, etc.

> Have they actually trained Copilot on their own source? If not, why not?

People have posted illegal Windows source code leaks to GitHub. Microsoft doesn’t seem to care that much because these repos stay up for months or even years at a time without Microsoft DMCAing them-if you go looking you’ll find some right now. I think it is entirely possible, even likely, that some of those repos were included in Copilot’s training data set. So Copilot actually was trained on (some of) Microsoft’s proprietary source code, and Microsoft doesn’t seem to care.

The question is not whether there's some of their code that they don't mind being incorporated, but whether there's any at all that they wouldn't allow to be. And more importantly, not used for their own bot, but for someone else's.

If licenses don't apply to training, then they don't apply for anyone, anywhere. If they do apply, then Copilot is violating my license.

IANAL, but they likely believe their unpublished source code contains trade secrets. They may believe that training a public model is okay on published source code (irrespective of its copyright license), but that doing so on unpublished source code containing trade secrets might legally count as a voluntary relinquishment of their trade secrets (if we are talking about their own code) or illegal misappropriation of the trade secrets of others (if they trained it on third party private repos)

I seriously doubt Microsoft / GitHub would care if Copilot or a similar model were trained on their proprietary source code. An advanced code completion tool does pose any significant risk of someone building a competitive product to GitHub or any other Microsoft products.

This is an intelligence augmentation tool. It’s effectively like I’m really good at reading billions of lines of code and incorporating the learnings into my own code. If you don’t want people learning from your code, don’t publish it.

Microsoft is not training copilot on your proprietary code that you keep on your own systems, just like they are not training it on their proprietary code.

Oh, one image is enough to apply copyright as if it were a patent, to ban a process that makes original works most of the time?

The software itself is not at issue here. If they had trained the network on public domain images then there’d be no lawsuit. The legal question to settle is whether it’s allowable to train (and use) a model on copyrighted images without permission from the artists.

They may actually be successful at arguing that the outputs are either copies or derived works which would require paying the original artist for licenses.

Then I think any work of art or media inspired by past sources would fall into this category. It's a very grey line, and I haven't seen anyone or any case law put it into proper terms as of yet.

Does "inspired" equal to "learned by software neural network"?

Olivia Rodrigo is a good case study here. Good For You was so heavily inspired by Paramore that Hayley Williams was given songwriter credit despite having no involvement in its making.

So humans can already run afoul of copyright this way, the bar for NNs might end up lower.

In that vein, surely MD5 hashes should also be copyrighted, as they are derived from a work.

Not really, since one of the major characteristics is being able to recover the copyrighted work from the encoded version.

Since md5 hashes don't share this property, they're not "in that vein".

If an encrypted file for which there is no key is treatable as derivative by law, then so should be an md5 hash. Both require vast brute force to extract/establish the original data, but both could be said to contain a derived representation of the work in question.

lol thinking about this more:

I understand people’s livelihoods are potentially at stake, but what a shame it would be if we find AGI, even consciousness but have to shut it down because of a copyright dispute.

what if they shut us down because of a copyright dispute? :-)

Seriously!!!

I didn’t say it cuz I didn’t think it would resonate, but it’s a whole new world we are quickly entering.

I think the result will be image sharing websites where you have to agree to have your image read into the model.

I think it is likely github will do the same with copilot.

maybe a fair price to pay for free repo hosting. wouldn't want my private repos being used for training though

Image sharing sites routinely steal artwork from the web. My business has a unique logo with the business name in it. It has repeatedly shown up on such sites, despite repeated DMCA takedown requests.

Simply appearing on a shared hosting site should not be enough.

Didn't yesterday someone proclaim generative models can't destroy anything worth protecting? It was about chatGPT but the principle is the same.

The real tragedy is being marketed to so heavily that we construe enforcing copyright on llm/diffusion companies with shutting down an AGI. I blame companies like openai purposefully marketing themselves poorly since nobody is going to enforce false advertising laws on something they don't understand.

That could be a funny movie.

Special agents from the MPAA sent to assassins an Android who can spew out high quality art.

> In fairness, Diffusion is arguably a very complex entropy coding similar to Arithmetic/Huffman coding.

so, digits of pi anyone?

> In fairness, Diffusion is arguably a very complex entropy coding similar to Arithmetic/Huffman coding.

Not the way it's used in Stable Diffusion models. Compressed data can be decompressed knowing only the decompression algorithm. To recover data from a stable diffusion model, you need to know the algorithm and the prompt.

A critical part of the information _isn't_ in the data you decompress, it has to come from you. (And this isn't that relevant, but it would be lossy, perceptual compression like jpeg or mp3, not lossless compression like Huffman or Arithmetic coding.)

Stable diffusion (or any likelihood-based generative model) is a learned compression algorithm. It is not the "container of compressed bytes". You can use a trained generative model to compress images, by combining it with some kind of entropy coding / arithmetic coding.

In this sense, stable diffusion is more analogous to the JPEG algorithm than it is to a specific collection of JPEG files. As it stands, the originals trainng data is not stored, even in a compressed way.

Great. Now the defence shows an artist that can recreate an image. Cool, now people who look at images get copyright suits filed against them for encoding those images in their heads.

Just because I look at an image does not mean that I can recreate it. storing it in the training data means the AI can recreate it.

There's a world of difference that you are just writing off.

> storing it in the training data means the AI can recreate it.

No it doesn't, it means that abstract facts related to this image might be stored.

The pedantry gets tiring. If the AI can't recreate it exactly, it can recreate a likeness that is compelling enough that the average person would think it was the same. If it can't now, it will as it gets better. That's the point of using the training data.

That is not the point of using the training data. It's specifically trained to not do that.

See https://openai.com/blog/dall-e-2-pre-training-mitigations/ "Preventing Image Regurgitation".

That's probably a very relevant point. (I'm guessing.) If I ask for an image of a red dragon in the style of $ARTIST, and the algorithm goes off and says "Oh, I've got the perfect one already in my data"--or even "I've got a few like that, I'll just paste them together"--that's a problem.

That's extremely not how it works. If there's only one training example it's not going to remember anything like actual visual details of it.

I realize that's not how it works. My point was that they're apparently taking deliberate steps to try to make sure the model trains over a large number of images and doesn't overfit on a small sample given a sufficiently specific "in the style of," etc.

Actually that's partly how it works.

A trained model holds relationships between patterns/colours in artwork and their affinity to the other images in the model (ignoring the English tagging of images data within this model for a minute). To this degree, it holds relationships between millions of images and the degree of similarities (i.e. affinity weighting of the patterns within them) in a big blob (the model).

When you ask for a dragon by $ARTIST it will find within it's model an area of data with high affinity to a dragon and that of $ARTIST. What has been glossed over in discussion here is that there are millions of other bits of related images - that have lower affinity - from lots of unrelated artwork which gives the generated image uniqueness. Because of this, you can never recreate 1:1 the original image, it's always diluted by the relationships from the huge mass of other training data, e.g. a colour from a dinosaur exhibit in a museum may also be incorporated as it looks like a dragon, along with many other minor traits from millions of other images, chosen at random (and other seed values).

Another interesting point is that a picture of a smiling dark haired woman would have high affinity with Mona Lisa, but when you prompt for Mona Lisa you may get parts of that back and not the patterns from the Mona Lisa*, even though it looks the same. That arguably (not getting Mona Lisa) is no longer the copyrighted data.

* Nb. this is a contrived example, since in SD the real Mona Lisa weightings will out number the individual dark haired woman's many times, however this concept might be (more) appropriate for minor artists whose work is not popular enough to form a significantly large amount of weighting in the training data.

> If the AI can't recreate it exactly, it can recreate a likeness that is compelling enough that the average person would think it was the same

That's the opposite goal of this image model. Sure you might find other types of research models which are meant to do that but that's not stablediffusion and the likes.

Why does this argument apply to an Artificial Intelligence, but not a human one? A human is not breaking copyright just by being able recreate a copyrighted work they've studied.

It depends to what degree it's literal copying. See e.g. the Obama "Hope" poster. [1] Though that case is muddied by the fact that the artist lied about the source of his inspiration. Had it in fact been an older photo of JFK in a similar pose, there probably wouldn't have been a controversy.

[1] https://en.wikipedia.org/wiki/Barack_Obama_%22Hope%22_poster

This just sounds like really fancy, really lossy compression to me.

Compression that returns something different from the original most of the time, but still could return the original.

It's not really. It's more like making an entire compression scheme that is very good at compressing images encountered in real life, rather than say, noisy images.

No, it means there is a 512 bit number you can combine with the training data to reproduce a reasonable though not exact likeness (attempts to use SD and others as compression algorithms show they're pretty bad at it, because while they can get "similar" they'll outright confabulate details in a plausible looking way - i.e. redrawing the streets of San Francisco in images of the golden gate bridge).

Which of course then arrives at the problem: the original data plainly isn't stored in a byte exact form, and you can only recover it by providing an astounding specific input string (the 512 bit latent space vector). But that's not data which is contained within Stable Diffusion. It's equivalent to trying to sue a compression codec because a specific archive contains a copyrighted image.

> It's equivalent to trying to sue a compression codec because a specific archive contains a copyrighted image.

This is the most salient point in this whole HN thread!

You can’t sue Stable Diffusion or the creators of it! That just seems silly.

But (I don’t know I’m not a lawyer) there might be an argument to sue an instance of Stable Diffusion and the creators of it.

I haven’t picked a side of this debate yet, but it has already become a fun debate to watch.

You can't (successfully) sue the creators of Stable Diffusion because they're an academic group in Germany, a country that has an explicit allowance in copyright law for training non-commercial models.

Exactly, the quarrel here is between the users of Stable Diffusion, some of which are deliberately, legally speaking with intent (prompt crafting to get a specific output demonstrates clear intent), trying to use Stable Diffusion to produce images that are highly derivative of and may or may not be declared legally infringing works of another artist, and the artists who’s works are being potentially infringed upon.

You can’t sue Canon for helping a user take better infringing copies of a painting, nor can you sue Apple or Nikon or Sony or Samsung… you can sue the user making an infringing image, not the tools they used to make the infringing image… the tools have no mens rea.

> It's equivalent to trying to sue a compression codec because a specific archive contains a copyrighted image.

That's plainly untrue, as Stable Diffusion is not just the algorithm, but the trained model—trained on millions of copyrighted images.

But in fairness, even a human could know how to violate copyright but cannot be sued until they do violate it.

SD might know how to violate copyright but is that enough to sue it? Or can you only sue violations it helps create?

I would assert (with no legal backing, since this is the first suit that actually attempts to address the issue either way) that the trained model is a copyright infringement in itself. It is a novel kind of copyright infringement, to be sure, but I believe that use of copyrighted material in a neural net's training set without the creator's permission should be considered copyright infringement without any further act required to make it so.

I think that is a very fair argument. It may win in court it may lose. I’m excited for the precedent either way.

That’s said, it does raise the question, “should this precedent be extended to humans?”

i.e. Can humans be taught something based on copyrighted materials in the training set/curriculum?

I think this is a reasonable question for the uninitiated—those for whom "training a neural network" seems like it would be a lot like "teaching a human"—but for those with deeper understanding (tbh, I would only describe my knowledge in both these areas as that of an interested amateur), it is a) a poor analogy, and b) already a settled question in law.

To address (b) first: Fair Use has long held that educational purposes are a valid reason for using copyrighted materials without express permission—for instance, showing a whole class a VHS or DVD, which would technically require a separate release otherwise.

For (a): I don't know anything about your background in ML, so pardon if this is all obvious, but at least current neural nets and other ML programs are not "AI" in anything like the kind of sense where "teaching" is an apt word to describe the process of creating the model. Certainly the reasoning behind the Fair Use exception for educating humans does not apply—there is no mind there to better; no person to improve the life, understanding, or skills of.

Stable Diffusion is essentially a Compression Codec though. It's one optimised to compress real world images and art, by using statistics gathered from real world images and art.

It's like the compression that occurs when I say "Mona Lisa" and you read it, and can know many aspects of that painting.

I will admit to knowing the overall underlying technology better than the details of what specific implementations consist of. My understanding is, though, that "Stable Diffusion" is both a specific refinement (or set of refinements) of the same ML techniques that created DALL-E, Midjourney, and other ML art generators, and the trained model that the group working on it created to go with it.

So while it would be possible to create a "Public Diffusion" that took the Stable Diffusion refinements of the ML techniques and created a model built solely out of public-domain art, as it stands, "Stable Diffusion" includes by definition the model that is built from the copyrighted works in question.

If you spent a decade trying to draw it, wouldn't your brain have the right "weights" to execute it pretty exactly going forward?

Except with computers, they don't need to eat or sleep, converse or attend stand-ups.

And once you're able to draw that one picture, you could probably draw similar ones. Your own style may emerge too.

Just thinking. Copywriters, students, and scribes used to copy stuff verbatim, sometimes just to "learn" it.

The product of that study could be published works, a synthesis of ideas from elsewhere, and so on. We would say it belonged to the executor, though.

So the AI learned, and what it has created belongs to it. Maybe.

Or, once we acknowledge AI can "see" images, precedent opens the way to citizenship (humanship?)

Don't think stable Diffusion can reproduce any single image its trained on, not matter what prompts you use.

It does have Mona lisa because of over fitting. But that's because there is too much Mona lisa on internet.

These artist taking part in suit won't be able to recreat any of their work.

I think there's a chance they might be able to recreate some simpler work if they make the prompts specific enough. When you set up a prompt you're essentially telling the system what you want it to generate - if you prompt it with enough specificity you might be able to just recreate the image you had.

Kind of like recreating your image one object at a time. It might not be exact, but close enough.

People have tried, unless the thing you want to recreat has been seen by it a lot (over trained) you won't get the same image. You don't have that much fine grained control via text only.

Best you can do is to mask and keep inpainting the area that looks different until it doesn't.

> if you prompt it with enough specificity you might be able to just recreate the image you had

At some point the input must be considered part of the work. At the limit you could just describe every pixel, but that certainly wouldn’t mean the model contained the work.

Does SD have to recreate the entire image for it to violate copyright?

As a thought experiment, imagine a variant of something like SD was used for music generation rather than images. It was trained on all music on spotify and it is marketed as a paid tool for producers and artists. If the model reproduces specific sounds from certain songs, e.g. the specific beat from a song, hook, or melody, it would seem pretty straightforward that the generated content was derivative, even though only a feature of it was precisely reproduced. I could be wrong but as far as i am aware you need to get permission to use samples. Even if the content is not published those sounds are being sold by the company as inspiration, and therefore that should violate copyright. The training data is paramount because if you trained the model on stuff you generated yourself or on stuff with appropriate CC license, the resulting work would not violate copyright, or you could at least argue independent creation.

In the feature space of images and art, SD is doing something very similar, so i can see the argument that it violates copyright even without reproducing the whole training data.

Overall, i think we will ultimately need to decide how we want these technologies used, what restrictions should be on the training data, etc, and then create new laws specifically for the new technology, rather than trying to shoehorn it into existing copyright law.

Do you know that the final trained model is only 2GB? There is no way it can reproduce anything verbatim. There is also Riffusion that can generate music after being trained on FFTs of music.

And how that's different from gzip or base64, which can re-create original image when given appropriate input?

well I guess it wouldn't be different, only there aren't any companies zipping up millions of images and then offering people the chance to get those images by putting in the text prompt that recreates them without paying any fees to the artists whose images were used.

Search engines do that.

good point, but didn't Google Image search lose some case and have to change their behavior?

If it's what I'm thinking about, I think they were forced to have decentralized image caching (i.e. the "user" is the one downloading images, Google just indexes).

LAION-5b is also just an indexer (in terms of images).

That’s my point, Diffusion[1] does seem to be “just like” gzip or base64.

And it would be illegal for me to sell or distribute zipped copies of images without the copyright holder’s consent. Similarly there might be an argument for why Diffusion[1] specifically can’t be built with copyrighted images.

[1] which is just one part of something like Stable Diffusion

A lossy compressor isn't just like a lossless compressor. Especially not one that has ~2 bytes for each input image.

How many bytes make it an original work vs a compressed copy?

Usually judges would care more about whether the bytes came from than how many of them there are.

Since SD is trained by gradient updating against several different images at the same time, it of course never copies any image bits straight into it. Since it's a latent-diffusion model, actual "image"ness is limited to the image encoder (VAE), so any fractional bits would be in there if you want to look.

The text encoder (LAION OpenCLIP) does have bits from elsewhere copied straight into it to build the tokens list.

https://huggingface.co/stabilityai/stable-diffusion-2-1/raw/...

The important distinction then is using another program or device to analyze the bits but without copying them, that takes its own new impression? Like using a camera?

Well, theoretically more like a vague memory of it or taking notes on it.

“any fractional bits would be in there if you want to look.”

What do you mean by this in the context of generating images via prompt? “Fractional bits” don’t make sense and it’s more misleading if anything. Regardless, a model violating criteria for being within fair use will always be judged by the outputs it generates rather than its composing bytes (which can be independent)

Fractional bits makes perfect sense. Do you know how arithmetic coders work?

One, of your compressor is specialised enough…so you can see how slippery this argument can be.

I agree with you. My intuition is also that SD itself is not a violation of copyright.

That said it can sometimes be in violation of copyright if it creates a specific image that is “too close to another original” (just like a human would be in violation even if they never previously saw that image).

But the above is just my intuition (and possibly yours) that doesn’t mean a lawyer couldn’t make the argument that it’s a ”good enough lossy compression - just like jpeg but smaller” and therefore “contains the images in just 2 bytes”.

That lawyer may fail to win the argument, but there is a chance that they do win the argument! Especially as researchers keep making Diffusion and SD models better and better at being compression algos (which is a topic people are actively working on).

So it's fine to distribute copyrighted works, as long as they're jpeg(lossy) encoded? I don't think the law would agree with you.

If I compress a copyrighted work down to two bytes and publish that, I think that judges would declare it legal. If it can't be uncompressed to resemble the copyrighted work in any sense, no judge is going to declare it illegal.

Reply on news.ycombinator.com