As I said, humans learn but the phase humans do it still allows for people be compensated.
I'm not anti-AI at all, I think its great and I think the artist who can leverage it can hugely benefit from it but it is not fair to just take something from people away.
They don't have the right to not allow people to learn from it, that's just never been a part of copyright.
> They don’t have the right to not allow people to learn from it, that’s just never been a part of copyright.
Yeah this is true. Stable Diffusion and other neural networks are not “learning” from it they way humans do though, it’s remembering and remixing and interpolating pixels (fixed expression), which is a part of copyright.
BTW compensation is not a very accurate summary of this problem. Machines that borrow someone’s style and copies and remixes without attribution is inherently problematic far beyond artists who live off selling their art, it dilutes both creativity and credit, in addition to undermining people who have to work much harder than the computer to produce images. These AI completely depend on being given human-created art to begin with, and the companies that are making them and using them are already making handsome profits, so it’s reasonable to expect some kind of return in addition to proper credit.
> Available for free online is not a valid justification for copying under copyright law, though, right?
Legal and attribution and licensed are all terms that are usually involved, but the core assertion is more or less correct. eg Images on billboards are protected by copyright in a similar manner. Something displayed on a website does not invalidate the copyright of the author or indemnify the owner of the site (or downstream users) from licensing conditions.
The laion 5b dataset is in the neighborhood of 220TB. (1) That is how much storage space you need to remember the pixels.
The stable diffusion 1.5 checkpoint is 7gb. (2)
1 https://github.com/rom1504/img2dataset/blob/main/dataset_exa...
2 https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/m...
To extend your musical analogy, it's remembering that many songs are in 4/4 time, and that major chords sound appealing.
Also, were you to compress anything, an mp3 or a picture, in a lossy fashion, to that degree of compression (~10^-5), you would no longer have anything resembling the original. The audio would be glitchy noise, and the image would be a scattering of apparently random pixels a few pixels wide.
Here's the thing - I empathize that this is disruptive in a very similar fashion to a tool that does store compressed copies of the work in question. It is capable of doing the same kind of damage. There's a conversation to be had there - but it's just not compression. That's not how the thing works.
'cause those images are not the same. Sports events are just easy to fake, because they're boring - all sports pictures look roughly the same.
1 - https://www.theverge.com/2023/2/6/23587393/ai-art-copyright-...
Calling people dumb is to be willfully ignorant to the fact that neural networks actually can and really do remember images, not just when overfitting, but also when examples are in a low-density area of the latent space, when it doesn’t have enough neighbors to average with. The machine really is technically a machine intentionally and specifically built to reproduce a weighted combination of it’s inputs, and it really is possible for that weight vector to spike on some specific training examples. This won’t go away by pretending it doesn’t happen, it will go away when people curate training data that is legal to use, and/or when people write software that detects and rejects outputs that are too similar to a training sample, or otherwise guarantee no individual examples can be reconstructed. This is precisely why the project we’re commenting on is interesting, because it takes a step in that direction.
I think it does go without saying that our legal system has made some pretty dumb decisions regarding tech in the past - we read here all the time about the patent system, which is damn close in spirit to copyright.
Again, yes, they can remember an image, but they are not remembering pixels, and it's not compression. The vectors you're referring to are not a smaller version of the data, nor are they a pixel representation or even a close derivative thereof. Sure, there's a connection between the latent space and the pixels, but I don't see how that's the same thing.
For those following along, (1) is the best paper I could find talking about extracting images from SD. I'm open to more resources, and I'm even open to being convinced I'm wrong, but not by intentionally overtraining a model and calling it 'compression'. That's a lie.
To take a step back here, is it really the incidental occasional regurgitating of an existing image that's got everyone on edge, or is that just an easier target than "this is disruptive so I want to make it go away"? I'm not saying it doesn't suck that this is gonna put a ton of people out of jobs; both my parents were professional photographers in the 80s. I get it. But like, let's talk about that. Not some orthogonal strawman.
And hey, just to get it out there. We might disagree but I'm not calling you dumb. I do appreciate your willingness to engage an opposing view - it's part of what keeps me coming back to HN.
Yes, the model is smaller than the total input data. But when it comes to recreating a single image, how many of the weights must be configured 'just-so' recreate an image enough to call it the same image? I'll admit ignorance here - but I also don't think that this is a thing anyone knows for sure. We can only just extract othello piece colors from a simplified, othello-specialized model designed to recognize two colors.
How much of the information from other images must be present to perform this task?
My instinct, given my understanding of how these things work, is that to replicate an image with any recognizable fidelity, you have to overtrain the model enough that you've affected a set of weights much, much larger than the pixel data. The internal representation of these images is concerned with much more visual information than just 'this pixel is this color' - by looking at layered outputs from the inverse type of system (image recognition, which is the core component of these models), you can see that they're encoding layers of shading, lines that map to brushstrokes or object boundaries, foreground, background, all kinds of stuff. A direct representation of an image with all of these would be necessarily huge - and we know this because we have them. Artists use layers in all kinds of image-creation software, and they're always way bigger than the JPEG itself.
I get that this may sound pedantic, but the term 'compression' doesn't seem, to me, that it fits here. Compression, by definition, makes stuff smaller
The assumption that there must be compensation comes from the capitalist society that we live in. Switching to Communism or something else can be a solution to not directly pay people who do works and still have them around.
This seems much more about the fear of competitions, than about violation of copyright. And yes, that's scary, but unavoidable as tech progresses. I don't think anything good could come from trying to strickten copyright here. The AI is obviously not copying directly, at best it takes a bit of inspiration from other works, something artists themselves do plenty themselves.
Something like UBI might help with the monetary needs eventually. Though the fear part might start getting really interesting in the near future. What do you do with your life when AI is better at it than you? When everything you can create, can be created by the AI faster and better?
Yep, it's all about being compensated and yes copyrights are already very problematic and creativity limiting concept.
People in these industries happen to be the first to face AI revolution but it will come for as all, eventually. UBI might be something but I don't know, it seems like hard times are ahead of us.
as has many other industries that faced automation and become obsolete. The fact that it is AI derived doesn't make it different than Luddites before.
I listened to an interview with an artist who referenced the “three C’s”:
Consent
Credit
Compensation
These seem reasonable to me. The request isn’t to eliminate the technology. It’s that artists should be able to consent to their work being used. Works derived from their work should include credit. And, they should be compensated.
Getting consent of all the authors within that 5 billion images, managing the infrastructure of paying them, would be a Herculean task, and the cost of that task would far outrun the 5 billion spend if you gave a dollar for each image.
That would kill any and all possibility of an open source AI model. The future where this happens would be a very dystopian one.
What I see in the discussion is a results first, truth second thing. This is understandable - in an existential fight, damn the consequences - I'm swinging for my own survival.
What's missing from that analysis though, is that by constricting open source models, you're not by any measure stopping the development or deployment of these models. Google will make one. Microsoft will make one. Apple will make one.
Adobe will make one. Then, in order to use these technologies, you will have to pay. A lot - and it won't be paying the artists. Sure, the initial training night send out a few cents for each work included, a small price for Google to pay to prevent competition. But you're smoking the wrong shit if you think for a second that won't change as soon as the lead is cemented.
So now you're still looking for another job, and you don't even get to play with this new tech and continue making art, because you don't work for Google.
Many are too angry and small minded to see how lucky we are. A tool beyond anything Adobe currently ships today, for free. Could have easily been a AutoCad level subscription (10k+ a year IIRC that would have left you behind if you hadn’t paid.
But here's the thing; I have never 'done art' as a part of my identity, and yet I can listen to them talk about it and understand where they're coming from enough to contribute.
Also, I work in tech and I didn't even make it through basic trig. I have a GED and a set of scattered community college credits that are never going anywhere because of a GPA < 2.
The whole idea that there are 'this type of person' and 'that type of person' and that those are immutable is a horrific and dangerous lie.
But some people have bought it.
These sort of tools are only going to accelerate this trend in my eyes.
Artists do not care about the copyright of people being able to learn from their art and reproduce it.
It is not been an issue for decades, because it is a long established standard of copyright. You can't copy a work directly, you can learn from it and reproduce your own sort of stuff.
What's different now is that this is a tool which artists can be replaced by. They not angry that their art has been learned from. Artists are angry that there is a tool which can replace them. And artists are looking for a way to make sure that tool is hampered as much as physically possible.
If it were not learning copyright it would be something else, and whatever that's something else is it would also probably be something with proxy for that core issue.
Imagine we had created an AI which was perfect, it could take your exact description of anything you asked it for and drew that thing.
And then I went and asked it for a picture of Mickey mouse in the style of Disney. Because it's a perfect AI perfectly reproduces it.
Who is in the wrong? I think it's absurd to think the creators of the AI would begin the wrong, because they did their job perfectly. It's a perfect app, it is a tool capable of literally anything.
At the end of the day, it's the person using the tool to violate copyright who is in the wrong, assuming they distribute those images.
And I don't believe the answer here is communism, I believe the answer is the same answer we've had for decades for the same situation. People need to find new jobs and learn new skills and do new things.
And as programmers, it's probably going to happen to us too, about as soon as it happens to artists. Time to consider to start to learn to weld or build homes or something like that.
I don't object that people should learn new skills and find new jobs, my concern is that people who actually put huge effort to produce that "training data" are now left dry. They should be compensated and move on, I'm completely against hampering the abilities of AI in order to preserve the current business structures.
It's greater than zero, otherwise the AI wouldn't exist, but these AI are trained off of 5 billion plus examples.
Examples that required work to collect and sort through. Examples that required millions of dollars worth of compute time to make into something useful.
What is there to compensate them for? These machines are not actually going to make these companies that much money either. The tools are open source. Competition is going to reduce profit on these systems to a very slim margin.
You have to not only compensate for the fact that they aren't involved in actually creating the AI, so they're only getting a fraction of what this thing is worth in the first place. Then you have to consider that there are five billion images in the training set.
The compensation is going to be nothing.
Currently compensating the artists would be like the Pirate Bay compensating the studios for their production costs through the gambling ads they run. No new move would have been ever made if that was the case.
I want my computer to be able to summon images out of the ether if I ask it for a picture of anything I could imagine. I could not imagine a positive system that would value such a small group of individuals over the need of the whole in a situation like this.
These days it's both publishers and artists attempting to create RMS's world of a right to read.
Under communism artists would be barred from making art entirely, ordered by the state to work in a place where it is deemed to be productive.
It would seem that it opens you up to the problem of other people deciding whether you're an artist or not and if so whether you're a good artist.
The same of course happens with capitalism, where it's your customers doing it. (Or in the case of books, which are almost never profitable, the VC-like publishers deciding to give you advances.)
So of course the reason modern mixed economies are good is there's more than one set of people that you may be able to convince to fund you.
Sure they would, it would just be "I got paid to do this art/music piece that then someone will use in their product they sell"
Surely that's their labor? Capital is the factor of production that isn't labor (or land if Georgist.)
This here shows that you are speaking without understanding what you are speaking about.
AI is absolutely trained. It's a process that is quite literally inspired by the way we understood neurons to work in the 1970s.
AI start with a big batch of random numbers. There's a big fancy scientific method used to adjust those numbers in order to cause the system to learn to do some task.
The process creates genuine and novel understanding of the problem space at hand. AI trained to do something simple, like add two numbers, will have a real solution to add two numbers in them once you have finished training them.
With bigger problems, similar understanding certainly exists and in some cases has been proven to exist. However, once you get to the difficult problems with billions of parameters, it's very difficult for us to check that because if we could we would have just written it without needing to use an AI in the first place.
There are lots of researchers who do lots of study and effort to ensure that AI are actually producing real outputs and have genuine understanding of the problems face they are working in. Do not insist that AI has no understanding if you have not taken the time to learn about those techniques.
Stable diffusion has genuine understanding of different images and how to produce them within it. It is not a simple system that reproduces would already exists. It is not a collage tool. It is not incapable of producing novel outputs.
Diffusion models are perfectly capable of producing any image that can possibly exist. It is only a matter of time before someone invents a new style that hasn't been seen before, and someone else is able to find a combination of descriptive words that causes stable diffusion to produce that output.
Or that someone produces a novel combination of words, chucks it into stable diffusion (or some other AI model), and produces a new style of art.
> Art is made with physical motor skills, time and effort. AI is a human made program that uses human made imagery, classified by more humans to generate a shopping list written by another human who want art that requires no effort of aquiring the physical skill needed to make the art. With no physical product, there is no AI art
This is simply not true. Art is incredibly cultural, vaguely defined, and often involves little to no work at all. If someone prompts an AI, gets a result, and sticks it to a wall, calling it art, it is probably able to be considered art.
> If you want to physically copy my art, go for it, I will applaud your skill, but taking my art and shoving it in a program
In short, it appears that you are threatened by the existence of AI and the copyright argument is not actually about copyright, it's about ensuring that the competition does not exist, and that people did not have access to these tools that might make artists less valuable.
This is actually an anti-capitalist argument, rather than an anti-AI one.
That doesn't weaken the argument, IMO, but it does help to know whose ox you're trying to gore.
Apparently it isn't. Now a computer can do it.
> that requires no effort of aquiring the physical skill needed to make the art.
So what? This doesn't matter.
> If you want to physically copy my art, go for it
This is the same thing as if a computer does it.
Why do we need farmers when you can just buy a burger at McDonald's?
My AI art generator works right now, on my PC. If everyone stopped making art, the "normal" way, my AI art generator would still work.
This is a false equivalence to your example, because if farmers stopped farming, then it would not be possible to make burgers.
But, thats usually what happens when someone comes up with a pithy, one off meme response, like you just did, instead of actually responding to the substance of the argument. (I expect any future response from you, to instead not be engaging with the substance, and instead coming up with reasons why the silly analogy still works)
Hasn’t been about the motor skills time and effort for a long long time. Art has been about the idea rather than a fetishization of how much time and skill something takes to build.
I'm not thinking from the current definitions of right and compensation structures and I'm not calculating fees here. What I say is that if you build your thing on top of other people's work and put them out of business this needs to be addressed because you will cause huge problems to the people who made your machine possible in first place and you will dry up your source of "inspiration" because you will no longer have people making living of it.
Previously, someone was being paid to produce something.
Now, they are no longer being paid and the business can get the same thing “for free”.
You can argue the same issue about out sourcing programmers to south east Asia.
Is it good or bad?
“…but how do we compensate the programmers who lost their jobs?”
It’s just economics right?
…but saying that doesn’t address the social issues that this technology is creating.
People. Out of jobs.
Bluntly, that’s what it boils down to.
Are you ok being replaced by a $0.02 / hour dude from Thailand? By a $0.001 / hour AI?
You think, maybe, it’s fair for the people being replaced to feel a bit upset about it?
I do.
> You think, maybe, it’s fair for the people being replaced to feel a bit upset about it?
I think it makes all sorts of sense that people are upset about stable diffusion.
And, people being upset isn't a good reason to change the law or outlaw the new technology.
Today the word "luddite" is a slur. But luddites were real people who made the same argument you're making. In their case, they were a secret group of textile workers so upset by the introduction of the mechanical loom that they went around sabotaging equipment.
Suppose we wound the clock back and the luddites won a legal battle and successfully outlawed the mechanical loom in the UK. Knowing what we know now, would you support a law like that to protect the jobs of textile workers in the 19th century? I sure wouldn't - it would have decimated the economy of the UK and stunted innovation.
Thats the danger of using the law to protect the status quo. Sometimes the status quo needs to change to make room for what comes next, regardless of how painful that change is.
The law doesn't exist to protect your business model.
[1] https://en.wikipedia.org/wiki/Luddite
We're at the verge of the second industrial revolution. I have no idea how it shakes out, but I don't think clinging desperately to the old ways of doing things will be a winning strategy in the long run.
The law shouldn't protect the status quo, it should protect human beings. Unfortunately, the law usually protects the wealthy first and only.
You are fundamentally not understanding the issue.
People don’t have an inherent right to be an artist just because they are an artist.
They can get upset, smash the machines in a Luddite frenzy or try to use the law to stifle competition but the simple fact is nobody owes them anything. Maybe they win a lawsuit or two but the AIs will just get trained on out of copyright work and art styles (for those who now pay for graphic artists and whatnot) will change like they do every generation.
I, for one, look forward to an AI generated animated Matisse dancer selling me milk.
The busker doesn't get compensated if I listen to him as I walk by (unless I choose to). He doesn't get compensated even if he has a really cool, kinda-unique idea, I get inspired by it, and start doing something very similar (without outright copying him in the sense of copyright). He doesn't get compensated even if I get my friends to join me and put him out of business.
Copyright is already an artificial piece of compensation that was added, and that is debatable (copyright can be seen as "theft from the commons"). It intentionally covers certain things and doesn't cover others, and this seems to fall under "others", as long as the training data was obtained legally.
I expect that to be the real problem, especially where content was reposted without permission.
They're in the Navy, aren't they?
But there are plenty of bad laws that protect business models. Hell, copyright - the law in question - is intended to protect certain models. Regulatory Capture is a term for a reason.
I agree with other points of yours though. Copying in art is for sure a thing, and same with style changing.
Really? The Constitution specifically includes a bit about: "to promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries". That was quickly followed by the first copyright act.
As I understand it, the copyright portion of this was exactly to protect the established business model for writers, which was publishing and selling copies of their work. In particular, protect it against other people publishing editions of their work.
exactly. This usage of works for _training_ is not part of that exclusive right, as far as i can tell. Otherwise, it would be a copyright violation for a human to read and learn off an existing works.
Training the AI isn't the problem.
Publishing output generated by it (by making it available to anyone else other than yourself), however, is.
>Otherwise, it would be a copyright violation for a human to read and learn off an existing works
You have the right to read and memorize the entirety of the Harry Potter series.
You do not* have the right to start a podcast where you recite the entire book from memory.
which is fine - this is public broadcasting of the works, which is part of the exclusive rights of the owner.
However, this is not what the AI is doing.
Yes, this is what the people who use the AI are doing. As well as people who release the AI model or a product that uses it.
if the people are using the AI to replicate an exiting works, to try to loophole the copyright act, then they will just get sued.
> release the AI model or a product that uses it
I argue that the model itself cannot be construed as copyright violation. After all, the model is information. What if i released a table of all of the word frequencies from books, and published that table? The table of word frequencies does not violate the copyright of the books from which it was derived.
Just because you _could_ re-derive the original books from this dataset, doesn't mean the dataset violates copyright. It _could_, if the dataset cannot do anything else (e.g., i just zipped up the text of the books and released that). But the AI model does not _only_ output the original, but it could also generate new works.
Were they publishing copies, your harry potter analogy would hold, but that is categorically not the thing happening.
Diffusion is not compression, or copying. It's stylistic synthesis, which is more like someone reading all the Harry Potter books, then doing a podcast about Harry Potter fan fiction.
Which, incidentally, exists (1)
This technology was clearly not one anticipated when the Constitution or the first Copyright Act was written. Or any of the later ones.
This will not fit in existing laws, so we have to go back to the purpose, "promote the progress of science and useful arts" via "securing for limited times to authors and inventors" particular rights.
What specific rights we'll need to secure here are going to take a while to work out. But new technologies have forced updates to copyright laws many times, and I' sure this won't be the last time.
Reading with a human is not copying, but reading by machine is - there are several cases where that has been enforced. This is covered by reproduction rights.
In this case, diffusion models are actually a perfect example of "the progress of science and useful arts". The law should be structured in such a way as to promote such progress, not hinder it.
We are potentially creating an all-seeing instant plagiarism machine that saps away not just monetary compensation but even credit and recognition.
Spend 2 years writing your novel? You'll barely sell a single copy, but people will happily pay 2 cents to hear GPTx's paraphrase of it.
Spend 5 months in the jungle with wild animals to snap that perfect picture? The magazine covers will be Dalle 5 rendering "nature photograph of orangutan cradling its new born baby in the style of X", your royalties won't cover buying a single lens.
so somehow the arts is above the telegraph operators or weavers?
> but people will happily pay 2 cents to hear GPTx's paraphrase of it.
except if you can tell that the novel GPTx "paraphrased" is a derivative work from yours, rather than transformative, you can either sue for loyalties, or take action. If you cannot tell if GPTx is sourcing from your works, even if it is, then there's no case to stand on. The AI wrote a better novel than the human.
> Spend 5 months in the jungle
and there's a lot of people doing "organic", "handcrafted" produce/foods. They tend to be very expensive compared to mass produced factory foods, and the market for it is relatively small. Just because something is difficult to do, doesn't automatically mean that people need to be paying for it to allow it to continue to exist. If it isn't profitable, stop doing it, rather than demand the world be changed to allow for it to continue.
Respectfully, no.
The top comment in this thread is saurik asking Stable Diffisuion to generate "an avatar of saurik" and getting a veritable likeness of themselves back.
It would be laughable to think that this is feasible without the model having labeled photos of saurik, which are all posted by (go figure) saurik.
The AI also generated a "better" photo than the one saurik posted. The AI saurik has more hair.
That doesn't change the fact that without the source material, this output would not exist. Nor that it would be really hard for saurik to "take action" on that.
----
In the end, the experiment to do would be very simple. Exclude certain data from the training set, retrain the model, give it the same prompt, and see whether the output changes significantly.
If it does, then that data was indispensable for generating the output for that prompt.
Something tells me we won't see such experiments done by OpeanAI any time soon.
This does not imply that the underlying model is a copyright violation. There is no case law on the copyright status of models but it is generally believed that training a model is fair use.
In the human analogy: it is legal to read Harry Potter and to learn Harry Potter from memory such that a copy of it is stored within your head. But if you reproduce and publish or perform substantial parts of Harry Potter from memory you are violating copyright (subject to the various exceptions to copyright: quotation, parody, etc).
It gets murkier than this, by far. The diffusion models are not just trained on images, but on text. It varies by implementation but Stable Diffusion for example used a pre-trained CLIP transformer network from OpenAI (and subsequently OpenCLIP). CLIP can have internal associations between words that in turn steer the diffusion image generation.
To given a simple example of how this could work (I'm not saying this particular example does work though): the model in total could understand that a "swedish flag" consists of a "yellow cross on a blue background", and it could have yellow crosses in its image training set, but no images of Swedish flags, and it would still understand how to draw a "Swedish flag" based on the language semantics.
Just as a human, actually.
So, this won't work. A style can be inferred and described and it doesn't necessarily need to be in the image training set at all.
I guess the test I described is the simplest case of detecting appropriation by AI. Still, I think it would be a good start.
Yes, obviously. We haven't found 40,000 year old telegraph stations in caves. No one thinks weaving is an intrinsic part of being human.
The cat is out of the bag, but we should think very carefully about what we're leaving behind before we embrace an artless society.
> Just because something is difficult to do, doesn't automatically mean that people need to be paying for it to allow it to continue to exist. If it isn't profitable, stop doing it, rather than demand the world be changed to allow for it to continue.
This I find positively pro-dystopian. Who cares about reality when we can get regurgitated fakes generated much cheaper? Just write "crying woman bleeds from head in front of ruined building, warzone (pulitzer) [trending on Flickr]" in the little text box, click generate, ship it, and go for coffee.
Who cares about playing against a chess player when you can just play against an AI opponent for free on your computer?
Who cares about going to a fancy bar when you can just drink at home for less?
Who cares about buying expensive concert tickets when you can just watch a recorded version for free on the internet later?
Who cares about buying a luxury car when a basic model gets you from A to B just as well?
.
People like you and many many others who think these questions are dumb will be the primary market for art.
But the point of good law isn’t to protect an established business model. If that were the case, we would have outlawed the loom because it displaced weavers, and the camera (sorry portrait painters) and the iPhone. (How many telegraph operators are left? None!)
When an artist learns to draw, they copy almost all their ideas from other art they’ve seen that they like. That’s how humans learn. I’m learning to compose music at the moment and everything that sounds good to me is probably a remix of musical ideas I’ve heard before in other songs. If I grew up in a different music tradition (say, ancient India) then all of my ideas about musicality would be different.
It seems to me that the only difference between me and stable diffusion is that stable diffusion copies less than I do, but from more sources.
Also, the idea that artists can’t learn each other’s styles is totally wrong. I heard an interview years ago with an art director at blizzard. The interviewer pointed out that blizzard’s games all have different art styles, and asked how they swap between styles. “Do you have different art teams?”. The art director laughed and said that was the difference between amateur artists and professionals. Professional artists can be given a style and a brief and they can draw in the style. When they move between games, he said, the whole art team spends about 2 weeks with the practicing concept art for game they’re moving to and critiquing each others’ work to align their styles. Then it’s smooth sailing. That’s how they hire artists too - they ask them to draw some concept art in the style of one of their games.
Sounds far fetched? That’s what we do when we hire and onboard programmers. Art for hire isn’t any different.