Modifying Artwork With Glaze To Interfere With Art Generating Algorithms

With the rise of machine-generated art we have also seen a major discussion begin about the ethics of using existing, human-made art to train these art models. Their defenders will often claim that the original art cannot be reproduced by the generator, but this is belied by the fact that one possible query to these generators is to produce art in the style of a specific artist. This is where feature extraction comes into play, and the Glaze tool as a potential obfuscation tool.

Developed by researchers at the University of Chicago, the theory behind this tool is covered in their preprint paper. The essential concept is that an artist can pick a target ‘cloak style’, which is used by Glaze to calculate specific perturbations which are added to the original image. These perturbations are not easily detected by the human eye, but will be picked up by the feature extraction algorithms of current machine-generated art models.

As a result, when this model is then asked to generate art in the style of this artist who cloaked their art, the result will be art in the style of the cloak, not that of the artist. The tool is available for download as a Beta release for MacOS and Windows (10+). The FAQ does not detail the possibility of a Linux version, but does helpfully point out the answer to questions such as whether the cloaking is easy to defeat using filters or screenshots (in short: no).

Regardless of what your views are on whether the content in training data sets require consent from the original creator, it’s hard to argue with the rights of artists to protect their style using such cloaks when posting content online.

(Thanks to [Tina Belmont] for the tip)

48 thoughts on “Modifying Artwork With Glaze To Interfere With Art Generating Algorithms

    1. Yes, but you can still go to jail if you intentionally commit forgery or fraud. Why shouldn’t artists have a tool to protect their work? If they aren’t being paid or asked permission for the use of their art to train AI models, they should have the means to prevent that use. Simple as.

  1. Poor artists.. 😂 First, the invention of camera made them lose their job (traditional portrait/landscape painting, I mean).
    Now it’s the AI.

    The idea of manipulating original art in order to protect it from AI, reminds me of antivirus programs from the early 90s on DOS. They had the ability to “immunize” executables by “infecting” it with an antivirus program. ;)

    1. you’re out of your mind if you really think cameras put any artists out of being artists, and you’re double out of your mind if you really hope for and want artists to be made obsolete. Even if AI truly ended all future human creativity (it won’t) then you realize you will be stuck with the same training data forever, right? no new styles to steal.

      1. You seriously think the camera didn’t put any painters out of work? Instead of laborious expensive sitting sessions, you get a quick session with a photographer for a portrait. Unless you’re a President or a noble, nobody gets a portrait painting anymore.

        1. Unless you were a noble or a president you didn’t get a portrait even then. Portraiture was incredibly expensive. What photography did was make artists question the necessity for photorealism and released artists from the need to replicate reality.

  2. Using DRM to invoke DMCA section 1201 as a legal trump card is an incredibly scummy move.

    It doesn’t matter if it’s in the context of AI training, fair use of DVD content, or preventing Joe from refilling his inkjet cartridges.

    It’s an admission that “we don’t care what the courts decide regarding copyright, we’re just going to make it a federal crime now.”

    Their excuse that it’s a preemptive measure is utter BS because this will remain a crime even if courts were to find in favour of fair use. 1201 makes no allowance for any form of copyright defence.

    1. It’s not a matter of copyright, and it doesn’t make changes visible to humans who wish to view the work. The pictures can be released into the wild for free, for all humans to view and enjoy. The singular thing being stopped is the stealing of style. poor AI bros crying and pooping their pampers because they can’t steal work without permission. boo hoo

  3. training a model on publicly accessible images is clearly ethical and hurts nobody.

    taking payment for access to the resulting model is less clear.

    intentionally displacing living artists is even less clear.

    the technology in the article is silly.

    people deserve to self express.

    have a nice day!

    1. I really don’t know what to say about this whole thing. I don’t like displacing living artists, but also, this technology lets regular people make visual works other than stick figures, something that I suppose might in theory be possible to learn, but I’m not fully convinced you can do it without a decent spatial intelligence. I’m a fan of tech that lets you go beyond the limitations of you body and mind.

    2. >training a model on publicly accessible images is clearly ethical
      I think that is still a grey area. As it rather depends on how/why it is publicly accessible – if it is accessible only because the real work was given a really really high quality photograph when it is put up for sale by an auction house or something the work was never intended to be public available. The original artist didn’t choose to make their work available for the AI or for anybody except the purchaser and their guests.

      For folks making art who really are making the effort to widely share their work intentionally I think it comes clearly under ethical “fair use” enough it is hard to argue against. But for the other type of artists I’d say this type of tech is quite likely exactly what they want, and entirely fair to use.

  4. Even outright copying is legitimate. I should be able to copy an artist’s work and sell it as my own. Originality has no value; good art is good art, no matter who is making it.

    1. Copying is not work. It may look like work when you’re doing it inefficiently, but it’s not the same.

      We pay artists for coming up with things that are not copied – because that’s something not everyone can do. The only thing that can have value is the original work; copies can be produced in any number so fixing any value to them is nonsense. You selling a copy of someone’s original work is basically cheating the customer.

      1. Selling a copy as an original is fraud. Allowing a customer to buy a copy os good sense not every can have the original but selling a copy allows the common man to enjoy art

      2. “You selling a copy of someone’s original work is basically cheating the customer.”

        Unless it’s market as a copy, maybe.
        There are Elvis imitators, for example. They act like Elvis and have their appearances on stages. They also get paid for their performance.

        Likewise, piano players play songs from Bach. And they get paid for it in a concert.

        Same goes for paintings. If a painting is sold/labeled as a copy explicitly, as a recreation of another artists work, it might be legal to sell. There are artists that dedicated their life to being very good counterfeiters.

        1. As you may notice in those cases, it’s the performance that costs something because not everyone can do it. Of course, if you want to go cheap, you can hire a bad Elvis imitator, or wear the suit yourself. It’s not the Elvis character you’re really selling, but your re-enactment and possibly your interpretation of it.

          In terms of the counterfeiter, the copied painting is only worth the minimum wage at which the painter would work. This is because what they’re really selling is the act of copying, and they’re not the only ones skilled in copying.

          1. To me that depends on what’s under the painting. If I was to make a painting and put Monet under it, it’s fraud. But if I put my name under it, it’s the same as an Elvis impersonator. As long as the “counterfeit” doesn’t have the name and/or signature of the original artist under it, why does it matter?

          2. With AI generated works it’s the same. You’re paying for the cost of the skill/computational time to make the AI model and run it. Not every computer do it (in reasonable time).

          3. The cost of the AI or computers diminishes by the unlimited copies they can make. They can churn out endless permutations of the stuff they’re copying, so there’s really no sense in fixing a price to any one.

          4. >In terms of the counterfeiter, the copied painting is only worth the minimum wage at which the painter would work. This is because what they’re really selling is the act of copying, and they’re not the only ones skilled in copying

            erm… wow. So materials that go into the recreations don’t have any cost and practically everyone can paint/draw in every style, so it is not a skill with any value?!?!

            It really isn’t easy to mimic the style of another artist convincingly, and the folks that are good at oils in the style of x are probably pretty awful at watercolour in the style of y and only alright at oils in some other very different artistic style…

            The creation of a good honest replica is just as much a performance as the Elvis impersonator, and rather similarly specific. You are almost certainly not a top grade Elvis AND really good Sid Vicious at the same time. If anything a replica is harder than creation of the original, as you have to work out what tools they had and how they used them to create the right effect, and almost certainly without damaging the original! While they just played around with the tools they liked or had on hand until they get an effect that meets their requirements.

      3. How are the customers being cheated if they enjoy the art? That’s the point, right? The more people that can produce enjoyable things, the better. We pay artists for the art they make, nothing more. If we paid them for “coming up with” things, they would never need to pick up a brush.

        1. >How are the customers being cheated if they enjoy the art? That’s the point, right?

          How much people “enjoy” something has nothing to do with how much they should pay for it. Things are really worth exactly as much as they cost to make, all things considered. If that is too much to pay, then you shall not make the thing. On the flip side, paying more than it costs to have makes no sense – that’s unearned profit. This profit exist mainly because demand exceeds supply, naturally or artificially, so people compete over who gets the product. It does not justify unearned profit – it merely means it happens.

          >If we paid them for “coming up with” things, they would never need to pick up a brush.

          How would they express that thing they came up with, without actually making something? We pay artists for two things: novelty or originality, and quality. If either originality or quality are totally lacking, then the price tends to zero because either everyone’s already seen it, or it’s so bad nobody wants to see it.

          1. The value of a thing has nothing to do with how much it cost to make.

            The value of something is how much someone/anyone is willing to pay for it.
            A thing’s value is determined by what someone is willing to trade for it.

        2. More to the point, the fact is that you’re selling your WORK, not the ART – the art is merely the end result of said work. The idea that you would copy and then sell it as your work means you’re misrepresenting yourself to the customer – it wasn’t your work that made the art: you took false credit.

          Such false advertising means that people are willing to pay you more – because they believe you spent far more effort and talent into it than what actually took place or what was necessary. Instead, they could have gone to the original guy to have a copy, possibly for free. Why pay you? You didn’t have anything to do with the case.

          1. The free market will take care of this, though. If I can copy artwork and sell it for $$$, someone else can do the same and sell it for $. In the long run, art gets cheaper and more people can have it. This is a good thing.

  5. Hmmm. I think if I were an artist and wanted to use “Glaze” to prevent AIs from copying my stuff, I’d train it to use pictures of middle fingers as the “obfuscation style.” Steal my stuff, get the bird.

  6. To be fair, reading Hack A Day’s comment section is an odd place…

    I guess I should generally stop caring about all forms of copyright protection, it is after all only there to ensure that those who make content has means to support themselves for providing future value in what they do.

    All this machine learning content generation seems to much better. I can’t afford to train any such machine learning systems, and will simply have to rely on those who can to act in good faith and be respectful towards the users of their systems. Just like they have been respectful to those who’s source material their machine learning systems are built on.

    Now. Personally I don’t see content generating machine learning systems relying on public domain content for their training dataset as unethical. Nor do I see any ethical issues with using content that has been explicitly authorized by their creator/owner.

    Copyright is mainly about ensuring that content creators can sustain the means to support their livelihood and provide future content of a similar theme. (that the court system favors the rich, and that proving copyright infringement likewise is a paper work nightmare is however a major flaw in the legal system.)

    The discussion about machine learning and art is frankly lacking foresight. The same arguments can be made about source code licenses and large language models.

    Is it okay for chatGPT and the like to be trained on GPL/BSD/Apache licensed code to then go on and make proprietary solutions. If all that GPL/BSD/Apache source code weren’t part of the training data, how good would chatGPT and the like be at spitting out working code? (my guess, not all that good.)

    I frankly think that the free software foundation should ask OpenAI and the like to test that theory. If there is a noticeable/major difference. Then there is grounds for arguing that the LLM’s output in regards to code should potentially be licensed appropriately to whatever source material it used. (some argue that “clean room practices” prevail here, but then I suggest reading about clean room practices. An LLM wouldn’t suffice. The one studying the source can’t also write the “copy”, they can only describe it for someone else to write. Yes, even humans have to regard/respect the license of their learning material when making similar solutions of their own.)

    1. In short, yes, it might be worth stepping away from the assumption of being entitled to prevent the self expression of others when it is peaceful. And no, imagining a missed sale does not count.

      Someone’s expectation of income does not supersede my right to self expression at all.

      Not everyone who “creates” things has this machiavellian trait. Many people put hard work into software and release it under MIT license.

      Enhancing average people’s ability to self express is a net benefit. Both in images, text, and code. The meme of the suffering artist is counterproductive, doubly so when invoked for large businesses.

      Separating analysis of creative tools and economic systems can facilitate a rational discussion. If one has a qualm with capitalist type system, then it is helpful to delineate that from preventing self expression of others.

      In summary, just because someone spent a lot of time working on something does not mean they can reasonably expect compensation. That is absurd. I could spend ten years writing a poem, that does not make it valuable.

      I would, and have downloaded an ai model, and would also download a car.

      The cat is not going back in the bag.

      Have a nice day!

      1. My main point is that the rapid development of machine learning systems is mainly thanks to a somewhat unethical use of source data for the training of these systems.

        I am not saying that one should outlaw these systems.
        Nor am I saying that using machine learning systems is unethical.

        What I am saying is that the source data for the machine learning systems should be gathered with respect to those who has created the source material.

        Depending on the end use of the system, different levels of interaction with the creator/owner will be needed. If the end goal of the system is analysis of input data, like object recognition, then one really don’t need to ask for permission. If one however has the intent of creating similar content as what were used to train the system, then asking for permission for such use would be respectable.

        In the end.
        The situation that unfolds isn’t favorable to anyone as far as copyright protection goes. But the development of tools allowing for this “creative freedom” is effectively in the hands of the upper echelons of society due to associated costs. So mentioning machiavellian is rather foreshadowing.

        1. The cost of training a model is certainly still high; this prevents many people from creating their own models with their own data. Not to mention the frequent need for a large dataset.

          Running models on personal computers is becoming more common, and that seems like at least a bit of an improvement vs running it on “cloud” or someone else’s computer.

          EXIF seems like a decent way for images to embed the “intent of file creator”. Like robots.txt or something. It would still be up to each actor/programmer whether or not to check for or even honor such a metadata flag.

          The point about application vs permission sounds nice. Recognition task vs generating content task. Recognition task is less displacing of the “original expressing entity”

          Either way, it’s unclear to me how the concept of copyright is going to translate into practice over the near future.

          Your thoughtful responses are greatly appreciated.

    2. “copyright protection, it is after all only there to ensure that those who make content has means to support themselves for providing future value in what they do”

      Respectfully, in practice, this is a complete fallacy.

      1. To continue the quote you cut short: “(that the court system favors the rich, and that proving copyright infringement likewise is a paper work nightmare is however a major flaw in the legal system.)”

        So yes, in practice copyright doesn’t strictly help creators in a lot of situations.

        But that largely doesn’t change the fact that copyright is still there to protect the ability for a creator sustaining a livelihood on their work if said work is suitably popular for such.

        1. The spirit of copyright does indeed seem to be to protect “creators” to promote an ability to gain compensation for said “creations”. It is unfortunate that the paper work obstacle results in a preferential protection of larger entities. Disney seems like a decent example of things going sideways.

          It’s unclear to me how this is going to work in a present/future where there are so many models that the barrier to “creation” is lowered significantly. Or even people using language models to automate applying for things, or even litigating claims.

          Looking forward to see how people interpret the concept of “fair use” in the context of these tools.

  7. Good God the number of brain-broken AI bros who both believe this is putting artists “out of a job” AND believe that is a good thing. You just want your toys, screw everyone else. Gross

    1. In creating with AI, you are not the artist. You are a client, and the AI has become the artist. Previously you would be requesting a piece from the artist and they would likely change you.

      People always say that because the images are online they should be able to do whatever they want with them but when artists put those images up this technology did not exist. I know a lot of artists who are just exhausted now with people mocking their livelihood saying they are obsolete, and those artists are looking elsewhere for work. Just because people have been displaced before by technology without any kinds of employment protections doesn’t make it any better that we are doing the same now.

Leave a Reply

Please be kind and respectful to help make the comments section excellent. (Comment Policy)

This site uses Akismet to reduce spam. Learn how your comment data is processed.