Stable Attribution
stableattribution.com
stableattribution.com
https://www.stableattribution.com/?image=e89f1e94-067b-4ab8-...
https://pbs.twimg.com/profile_images/1434747464/square_400x4...
Given that I only said "saurik" and SD came up with something that not only looks more than quite a bit like me but is more than quite a bit similar to the pose of my profile photo, I'd say clearly that photo would be one of the most important photographs in the database to show up when asking "which human-made source images were used by AI to generate this image"...
...and yet, whatever algorithm is being used here--which I'm guessing is merely "(full) images similar to this (full) image" as opposed to "images that were used to make this image"--isn't finding it; which, to me, means this website is adding more noise than signal to the discussion (in that I think people might learn the wrong lessons or draw the wrong conclusions).
The first one, generated by AI has you looking at the camera. The other one has you looking at an instrument. They don’t look much like each other to me, in terms of pose or anything.
The first images suggested by Stable Attribution looks a lot more like the AI image to me, in terms of pose and everything.
As another comment[1] points out:
> This appears to be just looking for the nearest neighbors of the image in embedding space and calling those the source data.
Stylistically similar images are not the same as source images.
Sounds very much like a human would do it.
If I 'know' how to recognize saurik and I know how anime is supposed to look like, I can check my digital photo library for a picture of saurik and than use that picture as a template to draw an anime version of saurik. If someone later asked me what pictures I used the photo is the only one I'd present. Not the thousands of anime pictures I have seen teaching me what anime looks like, nor the picture my eyes took meeting saurik.
> these people aren't 'luddites'
This is just total bullshit. I know plenty of artists who are embracing this technology to make all sorts of things that tools like SD were not designed to do, like psychedelic music videos, etc.
What the author means is that a few loud blue check marks on Twitter who claim to be artists have been tweeting, get ready for it, inflammatory claims.
It’s called Sony v Universal, and the legal doctrine for fair use that resulted is a test for “commercially significant non-infringing use”, of which a tool used for inpainting to remove power lines, latent space psychedelic visuals, and photo booth-painterly-style all are.
Imagine if Stable Diffusion was made illegal. Someone accuses me of using this illegal tool for one of these non-infringing uses, that is an image that doesn’t look like anyone else’s image as far as the court is concerned for copyright. I put the image on my website. If the image itself is not at all infringing, then what is the evidence that Stable Diffusion was used? Should the police be issued a warrant to search my private property for proof that I used Stable Diffusion without a shred of evidence or based on a tool that will always have both false positives and negatives?
My conclusion is there is an argument to be made for infringement in some cases, but it's based on degrees instead of absolutes. If infringement is defined as "copyrighted works were used in this dataset", then at a certain point (low enough learning rate) it becomes impossible to tell if infringing data was used. You'd be working with weight amounts that are so miniscule they could be rounding errors, yet by that definition would still be infringing.
And since any arbitrary data can be used with some set of keywords, the standard for what constitutes "infringing" changes with each model. As in, it would probably be hard to have a benchmark test that can definitively state "this model violates copyright." Any number of keywords can be trained on to obfuscate the prompt needed to reproduce the data, assuming there was even a high enough LR for the data to be reproduced similarly enough.
I'm unsure if there can ever be one standard for when a set of a bunch of floating point numbers can pass the threshold for constituting infringement. This is applying an absolute standard to a fuzzy algorithm. It's like compressing a JPEG, at some level of compression on the scale a picture of Mickey Mouse becomes unintelligible. But with JPEGs it isn't really useful to have an unintelligible picture of Mickey Mouse. However, it can be extremely useful to have a LoRA with the weights underfit just enough to where the diffusion gives novel outputs.
However, the key thing here is: do you really think "avatar for saurik" is going to come up with something out of the blue--out of all the numerous random images of people that exist and who have avatar/profile photos... people wearing lots of different clothing and in different orientations--where we can even be talking about such silly things as props and gaze direction? I will assert that would be ridiculous. Clearly seeing at least one photo of me (and AFAIK it was only trained on thousands of copies of that single photo) was absolutely crucial to the construction of this image, and yet this website isn't finding any such to show as it isn't really doing what it is claiming to do (and it isn't clear to me how it could without the original prompt).
From there, SD is upscaling from memory a lot and filling in a ton of details (as almost all of the copies of my photo it was trained on are very small, embedded in screenshots of tweets), but the cornerstone of that construction is clearly my Twitter profile photo... and then, once you do that, you can go back and attempt to rationalize "these were the photos that were most similar to the photo that we generated, and so I guess those were what the algorithm used to generate this image", but--at least in this case--it is pretty obvious that that isn't how this worked as there's no way you start with "avatar of saurik" and arbitrarily pluck an image that is this similar to the one photo of saurik you have.
I think there are two separate ideas that are being conflated here. The first idea is that there is a mapping between text input and a joint text/image embedding space. For that mapping, yes, your profile photo (paired with some caption text that says saurik) would have to be the most important training input because, as you say, how else does SD know what “saurik” looks like?
But the second idea is the mapping from the shared latent space to an image output. It is NOT necessarily true that your profile photo is the most significant training example for the generated image. That’s because the text “saurik” can map to a neighborhood in the latent space that has all the photos this tool retrieved — photos which are indeed very similar to the generated image.
As an example, let’s say that I resemble Ben Affleck. And you type in “throwaway1851” and you get back an image that does look like me, but also looks an awful lot like Ben Affleck. In this case I really wouldn’t be surprised if my one profile photo didn’t contribute much to the generated image as compared to a dozen different photos of Ben Affleck. Perhaps the original mapping just pointed to Affleck anyway, because I wasn’t significant enough to end up taking space in the model’s parameters.
To take your example: let's say its knowledge that "throwaway1851" looks like Ben Affleck comes from your one profile picture--the only time the multi-modal embedding model was ever able to associate that word and Ben Affleck's photo into a similar location in the vector space--then if it wasn't allowed to see that photo during its training there is no way that would have happened. It simply doesn't matter if it seems to mostly rely on its knowledge of Ben Affleck to conjure up a photo of you: it has so many photos of Ben Affleck that none of them really matter anymore, but that single photo of you that made it even realize you looked like Ben Affleck in the first place is absolutely critical and probably deserves most of the attribution.
For that, if you wrote down the latent embedding for “saurik”, then retrained CLIP and Stable Diffusion from scratch without any of saurik’s profile pics in the training data, it is quite possible that you could generate an image from the embedding you wrote down and it would look the same.
Pure speculation on my part - perhaps your profile pic is the real source material and this website is junk. I just think it’s an interesting and worthy question the site is trying to answer, even if it’s not possible to answer it with much certainty.
https://rom1504.github.io/clip-retrieval/
https://github.com/rom1504/clip-retrieval
Ironically, the stableattribution.com authors didn't give any attribution or credit to clip-retrieval and its developer
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
Whether or not this tool works, expect it to make the rounds on any tech publications that have previously written “stable diffusion bad” -type stories.
This just needs to be done at the source. Ie whatever AI service is generating the image should generate attribution as part of the result
I was surprised that they even had a prototype. I thought this was just a really good CTA
(The silly story they have on the site doesn't really score any points either, it reminds me of RIAA et al)
I don't think such a solution can work just based on the resulting image though. It probably needs to input prompt to have any chance of working.
Wouldn't the attribution be basically all the dataset with various attribution probabilities? It's kind of like a reverse neural network.
More so how are these image generators any different from text generators like ChatGTP? I feel like if first tool out from these AI gates was a bot that wrote good-enough-to-use code, no one would have batted an eye. Everyone would just go "yeah we told you that those pesky programmers would eventually automate themselves out of jobs", but since it is rendering pictures which most of the populace can appreciate and it is a skill that is easy enough for literally any child to pick up, but requires a lot of dedication to master all of a sudden this is theft and should be illegal.
I really hope this lawsuit or whatever doesn't go anywhere, because it will not change anything. The genie is already out of the bottle as far as image generators are concerned - however it means that we the people won't get whatever comes next. Whatever comes next will be tightly held by big corporations and they alone will reap the benefits, whatever they may be.
I've spotted this pattern a couple of times and this sort of circular reasoning seems concerning. Whenever one of (Stable diffusion, Copilot, ChatGPT) comes up in a discussion, their legitimacy seems to be swiftly justified by existence of the other two, even though they're all uniquely problematic in how they wash away attribution and licensing.
This is not circular reasoning, more observation of what people value. Since everyone can "Google and gather information" ChatGPT isn't valued enough, but since most people don't know how to draw DALL-E and Stable Diffusion are seen as industry destroyers.
The difference is I don't see book authors and news writers trying to sue OpenAI for "stealing" their articles and not providing attribution, or creating websites that try to divine which specific book chapters or news articles ChatGPT used to generate any particular response (as if that's at all representative of how GPT works).
By comparison, a “dog vs. cat” classifier that has 100% accuracy on the dog/cat task will, nonetheless, tell you that a slice of pizza is a dog... or a cat.
(You could possibly interpret the results as “if SD had generated this image, it would have drawn most from these sources in doing so”.)
Note that just because an image is similar (to human eyes) doesn't mean that it played a more significant role than a seemingly more dissimilar image. It could even return similar images that SD wasn't trained on at all. Even conditioned on providing an SD-generated image, it fails.
(Something doing what it claims to do, as opposed to naive image similarity, would actually be pretty cool and useful.)
It has more in common with using an AI language model as a "bullshit generator" than it has with human attribution. Which isn't a coincidence, since nearest neighbour similarity searches ARE a type of machine learning, just a much simpler one than SD.
Anti AI people who are upset about attribution, should learn the technology and try to create an actual attribution model: one that has a notion of causality, which could say who influenced who.
Causal model requires causal assumptions, but you can probably get far with the simple assumption that "works from the future do not influence works from the past".
It still finds ostensible "source images" for the art. So, yes, it's clear this service is pretty much bogus.
* https://en.wikipedia.org/wiki/Julie_Wilhelmine_Hagen-Schwarz
I will admit this is a nice tool for verifying the creations of SD aren't pure copies, so I think it will be useful for a time. But as AI-generated images start to taint future datasets, attribution is going to be significantly more complicated.
So now you can just use this tool to verify your output is safe.
They were already imperfect in it and now suddenly a new technology drags their work into the wilderness with no stuctures for compensation.
The images created by the AI are indeed novel but they feed on the work of people who spent decades building this style. Of course artist themselves feed from each other but they usually don't interfere with the business. So let's say, if an artist developed a particular style and someone wants to hire them for a business project like a game they can't feasibly just learn that style and use it so they hire the artist. Later other people catch on this style and develop over it. It only works because monetising through copying the style is not very feasible.
Suddenly you have a machine that makes it feasible. Instead of hiring the artist or licensing their works, you train your machine on it and start generating any number of images of that style or combination with other style without paying the people who come up with all that.
How is the artist supposed to be compensated for spending years of developing that style/method?
I'm fine with getting rid of all that copyright and license stuff but let's not pretend that what's happening now is a fair endeavour.
I have a ton of respect for calligraphers and believe they are artists, but at the same time I don't think that the millions of people who create custom fonts or use custom fonts are doing a bad thing.
That’s kind of the whole point of this debate. Should we change the laws and if so, how?
Style itself is not copyrighted, and that's a good thing.
I can create copyrighted content privately. I cannot share it. Law doesn't forbid imagination, wether human or ai and that's a good thing. But I cannot distribute these private rendition already without violating the content owner copyright.
Law seem pretty complete and well defined to me.
Are there cases where an ai and a human can generate the same media but which result in the media having different legality?
I don't think lawyers would agree with you at all.
> Just because an ai (or a human) comes out with a copyrighted character in a novel pose, it doesn't mean it's not copyrighted
My understanding is that there isn't a legal consensus on whether or not thats true. Copyright law wasn't written with AI generated art in mind. For a work to be copyrightable, my understanding is that it requires that you can make a "sweat off my brow" argument. Ie, you had to work to create something.
Does an AI count? We don't know. Or to put it in other terms, we haven't (collectively) decided as a society whether it should be copyrightable.
On the other side, "fair use" arguments are also at play here. If I train a 1bn parameter model on 5bn images, I could probably make an argument that 1/5th of a f16 probably constitutes fair use of copyrighted work. If that argument holds, I can train my model with impunity on any amount of copyrighted work so long as my model is small and the number of training examples is large.
Will that argument sway a court room? I have no idea. Is that fair? I don't know!
The law is decided by people. And we haven't had AIs like stable diffusion and ChatGPT before, so the laws haven't been written with this stuff in mind. There isn't even a legal precedent yet for how the current laws should apply to AI art. Speculating is fun, but speculating on how a judge will apply old laws to a totally novel problem is a fool's errand.
If you want to read about fun edge cases to copyright law, look up the history of copyright law for maps (can facts be copyrighted?) and how that interacts with trap streets.
This stuff is hairy and complicated even for lawyers. As an outsider, boldly claiming that copyright law is simple only demonstrates that you don't understand law. It'd be like a lawyer boldly arguing (with no knowledge) that compilers are simple.
The worst is that weight may be considered storage, but the point is... Law already covers that. Because it doesn't concern with technology, but with results of actions.
That is the whole problem with the argumentation, law is technology agnostic, just adding a layer of redirection doesn't matter, because tm it cares about the input and the output not what happens in between.
What makes you think that?
The law cares about whatever lawyers decide to care about. There was a case a few years ago where (if memory serves) a black woman sued an insurance company for discrimination after the insurance company refused to provide her cover. The company was using a neural net to decide whether to cover someone. The court demanded they explain the neural networks' decision - and of course, they couldn't. The insurance company lost the case.
In the aftermath they moved from a neural net to a decision tree based ML system. The decision tree made slightly worse decisions, but they figured if it lowered their legal exposure, it was worth it. With a decision tree, they can print out the decision tree if they were ever sued again and hand it to a judge.
> law is technology agnostic
Clearly not in this case.
There's plenty of other examples if you go looking. In criminal law, they care a great deal about the technology used in forensic analysis - both in its strengths and weaknesses.
If you don't know much about law, being humble and wrong will serve you better than being confident and wrong.
Also, differential insurance coverage tends to be one of the worst ways that systemic racism is perpetuated.
And again that case is technology agnostic, discrimination law requires you to be able to provide proof that results are non discriminatory, law itself doesn't care that it was specifically a neural network, it only cares about the end result, the firm lost because it failed to provide required data about their decision process, not because it was using neural networks, that they used a neural network was irrelevant on its own, and it could have been fine if they baked explainability in it.
Yeah, software "doesn't seem too complicated" to non-programmers too.
No offense, but if you aren't a lawyer then your opinion on legal matters has about as much veracity as a dentist explaining how software is made.
I have a certain amount of scorn for non-engineers telling me how their app idea is a weekend job and I should do it for free. I'm sure lawyers feel the same way about us when we claim there are easy answers around stable diffusion and copyright law.
The AI was trained without permission on copyrighted data. If it was as cut and dry as you claim then why do both parties think its worth going to court?
I don't know much about the law, but I know enough to recognise when its a job for the lawyers to figure out.
It's a good thing I never mentioned right or wrong.
Just to be clear, I think art made with AI also can have value because it's a tool after all. My concern is, it breaks the business for the people that feeds on and that's not OK.
Why not? This is a website that's filled with programmers, almost all of our wealth is built on the skulls of jobs that once existed, and those programs were built without the consent of the people whose jobs were automated away, often through observing and rewriting the processes they used to do.
The standard, this has been going on for decades, the only reason it's getting backlash is because this is a group of people who never expected to have this happen to them.
continuing to transform things which in part centered around exploration, discovery and experimentation into something cold and kind of dumb
like we will make some interesting things but it's the general trend of modern society that upsets people everything just getting easier/worse faster
I think the pushback is around status and elitism, and that people with certain backgrounds are societally expected to be not making art.
Almost everything computers do was once done by a person, and the people who did those jobs laid the framework for which the processes were automated through.
This has happened dozens of times. My entire job is writing software that was written in the 80s, which replaced customer service reps who are needed in the '70s, and the computer does what they used to do. It took what they invented, their processes, and made the computer do it.
This is what automation is. It's always built on top of, and replaced, human beings who used to do those jobs.
The idea of a specific artist having a unique style they learn is dying right now: the faster and cheaper way to develop a style will be to train an AI model on public domain work, and tweak till you get something you like. Then use it as part of a process to go from sketch and composition to output. If you must, play with your own style and make a custom model - that technology has gotten very accessible very quickly.
This is a gnashing of teeth over an unexpected price shock: creatives have been feeling pretty assured that automation wasn't coming for their revenue because what they did was a unique human endeavor that a soulless machine could not possibly threaten. It was meant to be factory workers and retail staff first.
People aren't worried about copyright. People are worried that a skillset acquired at substantial sacrifice is about to become irrelevant and are desperately latching on to any perceived lifeline.
Doesn't presume such thing.
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 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.
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.
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.
They don't have the right to not allow people to learn from it, that's just never been a part of copyright.
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.
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.)
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.
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.
> 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.
Edited to add: There's another lawsuit (a class action - 2), and after a little light reading, I came across section 5: 'Do diffusion models copy?', and my stomach jumped.
What they're doing, to make a point at trial that stable diffusion copies images, is _training images into the model, then using that trained model to prove that stable diffusion is a compression algorithm_.
This is a patent fabrication. If you train a model hard enough, yeah, it will produce the image you trained it on. And become useless for all other images. Congrats, you've just compressed your 7kb image to a 7gb diffusion model.
What scares me about this, is that the average court in the US is absolutely dumb enough to fall for it.
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
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.
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.
I don't think copying an art style is that hard. Professional artists in your example of the game industry can absolutely imitate each other upon request. If using the same style as another artist were so difficult games with multiple artists wouldn't have a coherent art style, but they do because the studio will develop a style guide and use their senior artists to guide the juniors.
> How is the artist supposed to be compensated for spending years of developing that style?
They haven't been in the past and they shouldn't be. They're compensated for the specific works they produce which are granted copyright. This is a good thing because otherwise the artistic domain would be horrendously polluted with claims, and producing original art would be like navigating a legal minefield.
So what's your suggestion, free food and housing for the artistic types or mercy killings?
Food and housing should be a human right, and it is only because of greed and fear that society allows someone to become homeless.
Not only should we provide housing and food for the artists. We should do it for everyone. And then they can spend time doing what they want, and earn more money that way but never have to worry about housing or basic food for survival.
I don't think that's what's going to happen.
But if things get as bad as you guys are saying, maybe after enough people lose their jobs to AI society will change somehow to take care of the less fortunate among us in better ways, providing housing, food and healthcare to the people.
Our greatest failing as a society is viewing those three things as individual problems. It was not an accident, it is a very profitable social failure.
Clearly the argument can’t just be “because they should hire someone to do it”, heck what if that person is totally booked? It’s not taking away from them in the slightest to produce additional artwork using a similar style if they weren’t available to take the job, so that seems irrelevant.
As long as you’re not taking their work and saying it’s yours or selling it I don’t see why you can’t mimic the style.
I also think that the use of AI is a fair game, the problem is that the thing itself is made possible by the hard work of thousands of people. By hard work, I mean years or even decades of development. They didn't learn to draw perfect photorealistic paintings, they developed styles and methods that capture something about humans and the machine took it from them.
I’m not sure. How are they providing compensation to every other artist whose work they’ve seen? To all of the people who gave them feedback and criticism?
The arts have existed for far, far longer than IP law and copyright and royalties. This idea that artists have no influences and that human art is not derivative is just weird.
By claiming that style and all future works that look like it, barring humans and AI from ever drawing anything else that has a similar style. Copyright used to cover just the expression, now we want it to cover all possible expressions of a style.
Which means there will not ever be a Stable Diffusion 2.0 with new art styles. In a year or two the instantly recognizable style of Stable Diffusion will be seen as extremely lame and crass.
I am pretty convinced at this point that in a lot people's heads, the fight to claim pictures like this as "real" or novel art is unconsciously just an urge to claim the beautiful woman as their own! Just thousands of men every day, generating beautiful ladies, yelling at people on the internet that they are real and they need to be protected...
Maybe its all just coincidental or something, but if you just step back and look at it, its very... interesting.
You're jumping to a conclusion that the data doesn't warrant, I suspect because it's a conclusion that suits you.
They're doing something like a "reverse image search" from an AI generated image over the original dataset and returning a few examples that have a high degree of similarity. There's no guarantee that the images they return are actually the ones combined to create the AI image. In fact their results are dubious - see Saurik's comment:
https://news.ycombinator.com/item?id=34671149
I don't have an opinion regarding the question of whether SD creates art or not. To be honest, as someone who enjoys art or kind-of-art in many forms, it's not important to me (although, of course, I can see why it is important to someone who makes a living from art or AI generated images).
However, this website doesn't add anything except noise to either side of the debate, from what I can see.
Before this, there were anecdotal examples of SD outputting an image with some Shutterstock watermark or very similar to some artist's work, but the prompts also seemed highly specific or were asking for something in that artist's style.
This tool at least lets us start to trace back the average image, and so far it does seem SD is adding something novel to its outputs.
There are many better ways, in the sense that they actually do something like estimate the causal effect of a specific training datapoint, like leave-one-out cross-validation training or surrogates to Shapley value, or using nonparametric models to trace backwards. This is a whole subfield of ML research.* (The primary summary is: "it's hard and the easy approaches don't work." Which why he's not doing any of those but an easy incorrect thing.)
* I'm not entirely sure why anyone cared so much... Research topics can be kinda arbitrary. But in this case, I think there was something of a fad around 2017 that there were going to be 'data marketplaces' where you would be trying to estimate the value of each datapoint to price it. This turned out to not exist as a business model: you either used big public data for free for generic model capabilities, or you had small proprietary data you'd die rather than sell to a competitor.
I think this website is evidence SD is a novel image generator, and rarely creates infringing images (at least in the way we thought of infringing before these kinds of AI).
You seem to be taking their word that the examples they explicitly provided are truth.
That is, that they show you all the important sources for the AI, without which the output would be drastically different.
As other comments show, this is not the case.
This is my hangup about luddite hand wringing with claims that the existence of Stable Diffusion and tools like it constitute some violation of the rights of the artists whose data was included in the training set. Existing laws on copyright infringement and whatnot don't persecute the paintbrushes, they go after the artist who published infringing material. Why should this be any different?
Fiasco of a tool to detect the source of plagiarism does not mean that Stable Diffusion creates novel art. When you need a character with dyed hair tips, SD will occasionally spit out copyrighted Harley Quinn portrait from DC Comics, Stable Attribution won't mention Paul Dini and Bruce Timm as character authors, but tool may show you some other pictures of people with colored head. Good luck with lawyers, what I can say.
P.S. I'm having fun uploading actual photos and art just to see what the site tell me with confidence "These human-made source images were used by AI to generate this image".
Edit: https://rom1504.github.io/clip-retrieval, this site has always been there to explore the LAION dataset using CLIP image/text embeddings and without the need to mislead the user.
Edit 2: As it's showed in this tweet: https://twitter.com/kurumuz/status/1622379532958285824, they are just using CLIP-L/14 and find the most semantically similar images.
To me it actually seems like this language intentionally obscures that they are 'only' using similarity search.
Imo this does a huge disservice to AI communication, providing fake explanations where even cutting edge research has poor insight into these models.
This site is grift, of course, and I doubt its creators expect it to sway anyone who knows it’s just doing a nearest neighbor search of image embedding vectors. But it’s oddly humanizing to see that the AI uses reference boards too.
The only reference material it has is the training set. Most images are not actually retained in the model; but certain statistically significant or repeated images will be. So if the model is regurgitating a training set image, this will alert you to that fact. But you are correct that it is not strictly speaking "attribution". You still need a human to say whether or not the images are just "in the same style" or close enough to actually be a straight copy.
This is also only talking about the "type prompt, insert bacon" kind of AI art. There are plenty of people who are feeding other people's work into an AI as a sort of automated tracing tool, and this won't catch them if they're not tracing images that were in the training set. Unfortunately these are also the most egregious and awful abuses of AI art generators.
Mostly I was just tickled by the fact that you can now get something that looks like a reference board for the generated image. There are some parallels to the human process, but maybe there are not enough parallels or the parallels don’t run deep enough to make it sensible to call this “the AI is using a reference board”.
This isn't true. The distribution of images seen by the model is not the distribution of images in LAION; the base models were trained with deduplication (or tried to) and custom ones could've been forcibly overfitted. So "repeated" as seen in clip-front doesn't prove anything.
Minor technical clarification:
For SD 1.X, CLIPText encodes a prompt and passes a (77, 768) [edit: up to 77] matrix to the core UNet. For SD 2.X, OpenCLIP passes a (77, 1024) matrix.
The black-box interaction between the text-encoded matrices and the UNet is how Stable Diffusion "learns" text associations.
Also, I am slightly wrong in that the first dimension will many not always be 77, since apparently there is no padding in the tokenizer. Test notebook here: https://colab.research.google.com/drive/192PDIbc2XiI1HgJQSdN...
Interestingly, even if it did build reference boards, that doesn't mean that the generated image would 'copy' or 'look like' the references all that much. We know this because there are diffusion generative models which do exactly that: they do retrievals first, and include the 'examples' along with the prompt before generating. Here is a visualization: https://arxiv.org/pdf/2204.02849.pdf#page=19 (the generated sample on the left is clearly semantically correct and high-quality but also very different from all the exemplars it was generated using).
You can already run exactly this kind of image similarity search against the Stable Diffusion training set using existing tools - https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2... for example
It attempts to do as you suggest in a tractable way, to understand which training data is most influential.
However, I'm not aware of anyone actively trying to adapt these techniques for large generative AI models (though would be great to see).
Now I go and create a painting called Sunflowers. Does Van Gogh's estate own some of my work?
Should every HN comment I write include a full list of everything I've ever read? What about a commercial work like a book, should that include a list of everything I've read or heard in the past 35 years of my life? That's a lot of attribution to keep track of ...
edit: I guess my question is where does derivativity end and creativity begin?
What Stable Diffusion does is effectively say "it doesn't matter how this was made" for every single artifact it creates, providing neither attribution nor acknowledgement. This happens even when it's clearly been influenced by, and in some cases is directly copying, elements or a whole of a very small set of highly influential inputs.
If a human author did that, people would probably be upset, if not outright accuse them of plagiarism.
But the controversy over Stable Diffusion is whether they owe the human artists a cut or not, right?
> What they _are_ saying is not to claim that you didn't train on Van Gogh
Stable Diffusion is definitely not claiming that it didn't train on human artists.
The controversy of Stable Diffusion is attribution. I could not care less about receiving a cut. Rather, I want my life's work in pushing particular style further than anyone has before to be recognized primarily so the next artist can trace the lineage and improve upon it.
> Stable Diffusion is definitely not claiming that it didn't train on human artists.
Stable diffusion claims that attribution is not important because "all artists copy each other".
Otherwise, it depends what the sunflowers you painted look like. You’ve seen the work that it is alleged you copied, so stuff that looks a lot like copying is assumed to be copying, roughly speaking. If you produced a work that reproduces substantial elements that were original in Van Gogh’s work, then Van Gogh has a case in copyright infringement. That can include the overall composition of the painting, some particular brushwork, etc. Techniques are not protected by copyright (that is generally the domain of patents, but afaik there are no extant patents on how to use a brush), so you can copy the techniques as much as you like, as long as you are not copying actual brushstrokes to reproduce the painting.
Copyright is concerned fundamentally with the degree to which you are stealing the market for the original work. It was created to allow authors etc to profit off their hard work, rather than labour for years writing a book only for someone else to make all the money from selling it. The existence of copyright makes it possible to make money despite the fact that copying someone else’s work has always been easier and cheaper than making something new. You could spend ten years writing a book, and it would take someone a day to print a copy, and twenty other printers could do it and none of them would pay you… so why would you write a book in the first place unless it were a political/religious pamphlet? To guide your interpretation of copyright, think about the degree to which you are stealing a market for Van Gogh paintings. If you produce a very close replica, then yeah, you are stealing some of the market. Photographing the painting and selling prints does the same thing. Producing something close from scratch without reference to the original is absolutely fine, if there is no copying then there is no infringement. The copyright statute just doesn’t want you stealing Van Gogh’s market for his own works using the crutch of copying his stuff, and through no original effort of your own.
Obviously AI art is a huge problem for this fundamental goal, because it hides the copying that it does. Someone selling a copy of your book without a license is easy to identify. AI art lets you take some art you like, launder it a bit, and then use it in your marketing materials without the original artists being able to notice the infringement and sue you for it. Fundamentally it is like money laundering, the same way AI data analysis has often been used for laundering human biases. That’s a huge problem and really subverts the entire intention of copyright. It steals the market and hides it at the same time.
You can also see how using an AI tool to launder images is just copying with a small amount of original human labour curating the input works. You can have some credit for your labour picking them, but you can’t really pretend the source images’ copyrights just disappear. The only issue is which humans are owed royalties. There’s a lot of debate going on about whether computers doing art is just like painters studying, just like you said, but if you analyse it as the laws do only in terms of the humans involved, you have a bunch of artists, and someone operating a lever labelled “mix these paintings so nobody can recognise my IP theft” that will vastly outpace their production and steal 99% of the market very quickly. It’s the invention of the printing press all over again.
If the copyright laws are too difficult to apply (eg an AI painting is such that it has laundered away any ability to pick out which bits infringe a human’s original painting used as input) then rather than everyone throwing up their hands and admitting defeat, the laws will probably change to better serve their original purpose. For example, AI paintings may have to pay royalties to every single artist used as input regardless of whether individually identifiable instances of copying are visible in the output.
Its not simply that he had good technique. Its that he felt something deep down. He felt it ever since he was a child, and made constant and frustrating (to him) attempts to then manifest that feeling. He would sometimes linger by a roadside too long and become his own little philosopher, seeing a new treatise in the ebb and flow of landscape from the sky. He had loves and heartbreak and fantasies, and found no other place to recover from those things than in patient work with the canvas. His whole life was a sacrifice, a sacred devotion to the spirit of humanity only art can capture so thoroughly and universally. You can't just do that.
I'm just saying, a guy like Van Gogh comes around just a few times a century, he is not somebody you can simply encode into a lot of different matrices, it takes a certain human life to make art like him, or even more simply, to simply see the world like him. Its not just about the way the sunflowers look, its about the choice in the first place to paint them, and making that choice in the particular context he was in, at the particular time in history he did. Its important that it was this painting that was painted after all the other paintings that had come before, both in particular for Van Gogh and for the world. And you can't ever recreate it, because its already done, and you can never step in the same river twice.
In general, and no offense to you in particular, but you will be forgotten as an artist. If this wasn't a certainty before, it is one now with all art generation and such. But don't take that as a bad thing, it is more than anything liberating. Try to remove all expectation and ego from what you are trying to make, and that will bring you closer to Van Gogh's work than any particular technique he used--your artwork should be completely personal, entirely internal, right up to the point that someone else see's it for the first time. It should only be judged on a sui generis rubric, more like a dream than a product. Stop, in general, trying to prove yourself on the basis of the tools you use. You could use sticks-and-twine or a million-GPU-DALL-E thing, whatever, its all orthogonal to the effect the final product might have, and that effect itself is only influenced by your attitude, your sincerity, and your vulnerability, as it is applied to the work.
(That is my maybe my main issue with the AI art bots, do the artists who work with them even feel vulnerable about their work? Can artistic sincerity and enthusiasm exist in AI art? Do AI artists feel the literally metaphysical stakes of what they are doing like Van Gogh did? Or is it all so constantly folded back into the idea of AI art itself? Is it all just different campaign posters for the cause of legitimizing itself as "real" art? Maybe just some time needs to pass, but I hope artists can still get nervous in the future, can still struggle at all to bring something out. I worry that people using this technology are too worried about proving themselves, they are not opening themselves up the possibility of failure that all artists need. They are not in a silent enough room to be able to hear what there brain is actually saying, only trying to respond respond respond. It is usually the struggle the artist has with themselves that is transmuted into the work such that it makes us cry or feel profound joy or sadness or whatever, we still need that psychic energy in the world, it doesn't need to be solved.)
To be like Van Gogh, you need to make art as if you would die and go to literal hell if you didn't, or go there if you made the "wrong" art. If you feel something like that, you are on the right track, but you still have 0.001% chance of dealing with that illness as productively as Van Gogh did.
The next Van Gogh, or the next Francis Bacon, or Rothko, or Cezanne, or whoever, will probably not have the machine-capturable-"style" of any of those people, but she will share in the particular mental complex they had, the same spiritual curse which drives one to reach past a scientific or practical world, to tap into something pure, almost completely formal, that cannot simply be taught or transmitted, much less encoded.
Sometimes, yes. John Q Nerd playing around with it might not, but the "good AI artists" I follow on Twitter certainly do.
> Can artistic sincerity and enthusiasm exist in AI art?
Unequivocally yes.
> Do AI artists feel the literally metaphysical stakes of what they are doing like Van Gogh did?
Some of them seem to, though I dunno precisely what Van Gogh felt so I guess I'm not sure.
> Is it all just different campaign posters for the cause of legitimizing itself as "real" art?
Not sure where you find this stuff, but it sounds really boring. Certainly not all of it is.
Your unsolicited advice falls flat (to me), sorry. My unsolicited advice to you would be to psychologize less.
But I really really do wish you the best with this stuff, the future is bright and interesting either way for art!
Have you studied all masters and eclipsed them, without acknowledging them so they can share in your fame? Will you use your untiring, undying skill to flood the market with your work so "artist" ceases to be a profession?
That's categorically new, and worth preventing.
This would be bad for present-day artists who've invested their lives in developing the skills of putting brush strokes on the page; those skills would become worthless. But the flip side is that Stable Diffusion lets anyone make art easily, generating whatever image they can dream up, without having to spend years developing the skills first.
So even though Stable Diffusion is bad for present-day _artists_, I suspect it's actually good for _art_ in the long term. "Anyone can easily create an endless supply of art, based on any idea they can dream up" seems like a good thing to me.
Put another way, it is the belief that the 2020s are the peak of manual human art.
Imagine if we'd felt that way a thousand years ago, or even a hundred. Is it worth being wrong?
Also, I think we have to weigh that tradeoff against all the new things that will be created if art becomes accessible to 100x as many people. On net, I think Stable Diffusion opens many more doors for art than it closes.
As far as I know, nobody is even thinking about doing the very expensive experiments needed to get ground truth data for formal attribution techniques in the generative AI context (for a given prompt, retrain your model so you can see how the output changes when a particular training example or group of examples is omitted or added), so we're nowhere near building true attribution systems for these very large models. Centering the training data will be net good for public discourse on the topic.
That said, I see why people want to push back on some of the language used here.
Even if it only finds similar matches and not true attribution, I actually think that is better. Say I come up with a neat design but I'm not very famous, and later someone more famous comes up with the same design on their own. I don't deserve attribution, but I would argue I deserve recognition. Regardless of whether or not the popular design was inspired by or derived from the original; having a model like this match the popular design with original, see that the original was created earlier, and give it recognition would be vindicating.
In fact, what if we create a neural network like this one to trace out huge DAGs linking every media with its similar-but-earlier and similar-but-later counterparts? It would show the evolution of culture on a large scale, how various memes and pieces of culture get created, where "artistic geniuses" likely get their inspirations from; and it would function as a great recommendation engine.
As for copyright and royalties - the site's intro never mentioned them, just "attribution" and "people's identities". And honestly, I don't think people deserve a cut from art generated from AI using their art unless the art is extremely similar. Because most of the time they are not that similar: the AI takes one artist's work (which would not be enough training data on its own) and mixes it with many others, like humans do, and I don't believe the two are different in a way that makes the AI mixer preserve copyright.
"These human-made source images were used by AI ... to generate this image."
The genie isn’t going back on the bottle.
I'd like to believe this isn't one of those things where we can only move on by everyone who believes the various correlated falsities dying, but I don't think I can.
> Usually, the image the model creates doesn’t exist in its training data - it’s new - but because of the training process, the most influential images are the most visually similar ones, especially in the details.
Would be cool if this were true, but I don't think it is, because the prompt you used and the captions on the training images are being completely ignored. If two different words tend to be used in captions for very visually similar images, and you use just one of those words in your inference prompt, I'm pretty sure the images that were captioned with the word you used are much more "influential" on your output than the images that were captioned with the word you didn't use. (Like, "equestrian" vs "mountie" or "cowboy" or something.)
1. Take the prompt you used, and use it with a model checkpoint that was trained identically to whatever model you're using, except that the top 21 images this website shows you are removed. In most cases, while your outputs won't be identical (I assume), you can probably get something pretty similar.
2. Now, take that same prompt, and use it with a model checkpoint that was only trained on the top 21 images this website shows you. (AFAIK you can't really do this because Stability hasn't released a "completely untrained" version of any of their models... though maybe they have and nobody cares because it's useless for most purposes.) I'm not completely sure what you'd get, but my bet would be that you get either nonsense or a memorized replica of one of the training images, not the same output image you got previously.
Some prompts I've come up with generate excellent and definitely novel results (without necessarily much work put into refining the prompt), others are extremely hard to get working well with hours of work even if I know it's something that isn't novel.
This is simply not how any of this works and is only your imagination. Stable Diffusion is deterministic and has no instance memory. The seeds are random and length of session or starting a new instance has no effect on the randomness of seeds. Every seed is as random as the last, regardless of how long an instance has been running.
However, I have definitely found at least a handful of generated images over the last few months that were almost 100% the same as the training image. I don't have the references handy, but it shouldn't be too hard to replicate. I was looking at different artists I liked on Artstation but that had few examples of their work. I then used a fairly standard prompt and only changed references to the artists I was testing out. In several of those instances, the generated images in SD were near 1-1 with one of the source images the AI was trained on.
https://github.com/rom1504/clip-retrieval
It does not appear to actually show you how images were used to train a generative AI.
In Who Owns the Future, Jerron Lanier proposes that the only path to an economy with a sustainable middle class not ruled by Google and Facebook is through this kind of attribution and subsequent micro-royalties. It's a fascinating read
I can see other commenters picking apart its method of heuristically guessing at source images from training data. That obviously won't be accurate, or a full picture, but I wonder if it would convince a judge.
An interesting challenge for these heuristics would be to take the picture under test along with its prompt, retrain the model without the training pictures it identifies, and regenerate using the same prompt to see whether the output is remotely similar.
Obviously that would be hilariously expensive and slow for a casual web service like this, but not beyond the realms of possibility for a wealthy copyright-holder.
e.g. if an prompt for an image includes "in the style of Kincade", and you could subtract all of Kinkade's copyrighted images from the training data, would the model still be able to produce anything like his work? If not, Thomas Kinkade might have a copyright case against people who publish AI art "in the style of Kincade", because he could show that his input was the major contributor to any lucrative output, even if nobody could pin down the cause & effect.
It’s a strange situation. Certainly a human painter is free to study copywritten works closely and then produce art in the style of the other artist, provided the resulting work is “in the style of” but does not copy a specific painting. So why is a human SD-artist not allowed to do the same with their preferred tool? Perhaps because we can draw a causal line (a very fuzzy one) between the original and the knock off? But wait… in the human painter case the same fuzzy line exists, it’s just through wet-ware not hardware.
Perhaps passing first into a LLM trained to generate good prompts but never output specific artists’ names could get close? Has someone tried this?
The line will surely get a lot less fuzzy with AI models.
So, if you take a Stable Diffusion checkpoint (call it "A") which is only lightly trained on some subset of an artist's work, then fine tune it on the full corpus of that artist's work to a point where it's still coherent/"good" and just shy of actually memorizing the fine tuning data (call the resulting model "B"), then define model "C" as 2A-B (i.e. A + (A-B), where A-B is the artist's task vector multiplied by -1), can you still produce qualitatively similar images with model C? Whether with the exact same prompt, or the same prompt with "in the style of Kinkade" removed (which doesn't mean as much if Kinkade's task vector was subtracted), or with any prompt whatsoever?
Lots of issues with this as laid out -- it's definitely not quite the same as "forgetting" Kinkade from the training data, and "any prompt whatsoever" introduces tons of leeway, and most good AI-assisted art is not just an unmodified single text-to-image output anyway -- but it might be a promising direction to explore.
(Strongly disagree with the "copyright laundry" characterization, by the way.)
All the negative examples are also instructing the AI how to make an image, not just the most similar images.
This is a bad joke that reinforces poor understanding of how image generation works.
At least from a legal perspective, protection (i.e., indemnification) should be offered to those who are clearly attributing their sources (which means they shouldn't be violating copyrights in the first place) and if they don't use clear attribution of sources, they should hold the burden of proof to show that they are not violating copyrights.
If this is just CLIP retrieval, that raises some ethical problems with the pretense of this site. It could make artists look silly for depending on an overstated claim of provenance, or worse still have artists pursue AI artists because an image looked kind of like their own artwork, with nothing more behind it.
Underlying this whole push to give attribution via AIs, there seems to be the general understanding that these AIs like SD are doing something very different to "creating" art, but are merely "combining art". I agree with this view, but it doesn't seem to be explicitly said very much.
I have uploaded my image made by SD: https://ibb.co/3NxPdNw The list of proposed sources seems to be very random.
EDIT: https://www.stableattribution.com/?image=541edf3c-6281-4177-...
Bullshit.
People learning every day to draw they are learning from other images. They using others styles and some time they are coming up with a unique style which was never used before. SD does the same. You can't recover original images. You can have similar style or composition, but you will never recover original image.
Getty Images v. Stability AI – Complaint - https://news.ycombinator.com/item?id=34668565 - (194 comments and counting)
I suspect the humanity is omniscient and it's only problem is lack of possibility to keep everything it knows in a mind simultaneously and connect the dots. An "AI" seems to be a solution to this problem.
Even an individual human could be almost omniscient if he could simultaneously put everything he has ever knew/thought/seen into his working memory.
So, even if it fails at correctly attributing source data. I'm wondering if it doesn't also fail at the concept of attribution. So far it just shows you some pictures with no attribution, and saying that whoever made those, made that.
Am I missing something? Why doesn't it know who the "human made sources" were made by?
Download the first input picture, search for it, and somehow the picture is AI generated ripping off itself, and somehow influenced by other images despite the first input and the output being bit-identical.
Yeah for a small subset of insiders to a small subset of creators. Really that was already a fiction prior to stable diffusion.
And then why human-made artwork would not need attribution of prior work that influenced it but AI made work would require some. We could extend the right to cite there and if we can't make it obvious that more than 10% of a given image was from another one, attribution makes no sense.
I say : my code as in "the text of my program", not "the idea of my code" (I have no problem with people re-using my ideas as they are mostly ideas that I have learnt from someone else).
My point is that it doesn't seem to prove that any of the images it finds are actually the source images.
Maybe just do reverse image search in google for the first approximation of attribution?
Have not seen Artists doing the same thing before very much. Perhaps an inspipred by but not all the images they have seen which lead them to be able to draw a new enough image.
I don't know what it's trying to achieve other than to waste everyone's time.
Those human artists saw art from other artists who saw art from other artists who saw art from other artists etc.
It's a nice, but as I say, flawed concept.
Ironically they don't give attribution to where this technique comes from https://rom1504.github.io/clip-retrieval/
I have yet to hear firsthand from any professional artist a single incident of Stable Diffusion causing them harm or lost revenue, but I have heard from a lot of armchair lawyers salivating about the idea of demonizing/criminalizing anybody that uses a piece of novel software.
1) have a vested business interest in not automating this kind of stuff (e.g. they churn out low-effort quickturnaround copy / design), so the automation is coming for yet another mind-numbing half-automated (sorry non-stop CMD+C / CMD+V isn't really "creating" anything novel or of value). Yeah those SEO spam gigs will be written by robots now, how quaint.
2) are copyright justice warriors who are raging for the sake of it, meaning they aren't even designers, artists, writers or anything like that.
3) do create some kind of art / text etc., but of a quality that is not putting them at risk of getting the badge of honor that is "in style of <copyright justice warrior>".
At the core of all this rage is just envy at someone being smarter and more successful than you are - in computer science, in statistics, in business, in art, in writing. Yeah someone did something so good it made it into Stable Diffision as a "named prompt". Yeah someone was actually capable of creating Stable Diffusion.
Protectionists are disgusting.
Finding similar pictures in an embedding space does not mean any of those pictures are part of the attribution chain any more than anything else.
The page in this thread doesn't do attribution, as I said in my other comment. But if it was possible, it would be very important to the field
I mean it’s technically impressive, I guess. But why would anyone care or pay for this product?
Second, a rant.
Look, if you are a photographer or artist or writer, your individual contributions to civilization are zero. I'm sorry I have to be the one to break the bad news to you. It's just the mathematical truth.
We are still in the childish (c)opywrong era of civilization where people born in privilege were brainwashed into thinking they were special snowflakes and their creative contributions to humanity are far more valuable than they actually are.
Your contributions, I don't care if you are a modern day Da Vinci, are but a grain of sand on a Mount Everest of human creation.
That photo you took? Far far far easier than the immense amount of effort it took to build that camera and get it into your hands. That book you wrote? Far far far easier than the thousands of years it took to evolve the letters and words you used to write it. That song you sang? Trivial compared to the collective efforts of hundreds of millions of people who pioneered music theory and instruments.
People who clamor for attribution ironically spend relatively little time digging into the details of all histories of all the ideas they are building upon.
I'm not saying go out and plagiarize. I'm not saying stop creating. I am saying to wake up and think from root principles about ideas and the absolute stupidity of the (c)opywrong regime. All these big AI models are ignoring (c)opywrong law, and you should too. Even better, contribute to the fight to pass and amendment to abolish (c)opywrong once and for all.
</endrant>
Yes, AI is still a bit crude now, but in 10 years this is going to like an old man yelling at the wind.
https://www.gettyimages.co.uk/detail/news-photo/an-unrestrai...