It's a political solution that capitalized on fear, yet does not offer tangible protection. Even watermarking is more effective.
The point for a given artist is that your images aren't there so the AI can't imitate your style (at least not by using your name).
It'll work regardless. All it needs is X artist uses this medium, does these sorts of pictures, looks like artists Y & Z, etc. etc.
SD does okay on some artists that don't have works in the image dataset. All their name does is pick out a position in the embedding space.
I don't think this is true. Can you provide an example?
But i do feel that unless your style is very unique and no existing "roots" in existing styles, it would not be possible to "protect" it technologically.
That's the purpose of the training, no? To generalize the model so that it can produce images it has never seen before. That includes images in styles it has never seen before.
Whether or not models have generalized to that point is a different question, but if they do, (and let's be honest, alot of artistic styles aren't that unique or different) the only thing that would be different is that the model cannot conjure up the style by providing the artists name in the input, instead one would have to describe the style in other ways.
So CLIP can have picked up an association that a named artist usually is synonymous with for example "broad strokes, moody lighting" and then that is fed into the diffusion model, which doesn't know the artist but DOES know what broad strokes and moody lighting is.
But sure, if CLIP didn't know about the artist name either it won't work of course.
By the way you can still just enter the particulars of the artist you want to mimic by text as well. There is not THAT much information in a style and you won't need to feed an image into the system.
I guess all of this with artists trying to protect their online works by watermarking or glazing will only be a very short speedbump for better or worse. If a human can do a 1-shot style transfer by a single glance at a work, the next round of AIs will as well, and won't be hampered by adding noise to the works and you might have "style extraction" tools that could work like chatgpt in that you iteratively instruct by text commands what to do to get closer without ever letting the AI look at an image.
> There is not THAT much information in a style and you won't need to feed an image into the system.
I think it really depends on the style. Some styles are simple, others are more complex. How that translates to an AI understanding it though, I have no idea. But in terms of brush strokes I don't think it's fair to say every style can be described as "broad strokes, moody, dark". Some very simple styles, maybe.
Aren't you kind of jumping ahead and assuming copyright protection applies to ML training? I thought this remains undecided right now.
In short: "Section 103 (17 U.S.C Sec. 1201(a)(1)) of the DMCA states: No person shall circumvent a technological measure that effectively controls access to a work protected under this title."
So if Glaze is viewed as a protection measure, circumventing it would itself be a breach.
Hmm. That's a good point and now i am less sure.
I suspect it would come down to intentionality.
if f(x)1011 and 1010 both "look" the same shouldn't a well designed AI learn equally well to recreate that particular look. It doesn't need a "deglaze" circumvention step it just needs a continually improving learning process that automatically accommodates both glazed and un-glazed images.
It does, though.
> This is a hedge against courts deciding scraping data for training purposes is valid.
For that hedge to work, though, it needs:
(1) Not to substantially degrade the art it is used on, and
(2) [To protect anyone other than a major player against other major players] ] Not to be trivially bypassed by in a way that can be incorporated in automated workflows.
> now you are applying a filter (creating a derivative work) to defeat a copyright protection mechanism, both clearly prohibited in current law (US jurisdiction at least).
If the Fair Use exception applies to training an LLM (which is creating a derivative work, itself, before considering Fair Use), then its extremely clear that applying a filter to incoming works as part of that process (even if not if the filtered work was used for any other purpose) will also be protected. So the derivative work thing is useless.
The circumvention measure thing might technically work (in that it interjects a violation into a workflow that would otherwise be Fair Use), but as a practical matter that doesn’t matter for most users or against most violators. Moreover, to the extent its primary effect would be to adversely impact otherwise noninfringing use, it would be a textbook case of a justification for the Librarian of Congress issuing a DMCA exemption, which would then negate the legal utility entirely.
> For any serious player making tools, you’ll specifically avoid defeating these techniques. For any minor player you’ll now have to go to the backwaters of the internet for tools to do this that you hope won’t steal your bitcoins.
The proof of concept defeats already demonstrated use…the same tools that are used for AI image generation and training models on artist styles in the first place.
> Every notable artist will be only upload their art to sites that offer something like this
The samples I’ve seen of the damage this does to art suggests that this isn’t the case.
> This isn’t a technical solution to this problem, it’s a political solution that happens to use tech.
It’s neither, its just a non-solution.
That's a big if. Sure the librarian of congress could justifiably issue an exemption, but i think it is far from garunteed that they would.
I think you are underestimating what a cooling affect FUD related to anticircumvention could be. Just consider all the stuff related to dvds that went down back in the day e.g. Dmitry Sklyarov being arrested at defcon (even if it didn't stick). The uncertainty could definitely have a major cooling effect.
I think you are overestimating it. Because you overestimating the degree to which the present legal context is similar to when DeCSS was an issue, and because you are underestimating the degree to which the financial and political power imbalance between the xxAAs and the other side was key to the effect then, and because you are misunderstanding who is doing the thing this most effects, before countermeasures (its not the easily deterrable parties), and because you are (based on the circumvention POCs) overestimating the degree to which any controllable specialized circumvention tool is needed to circumvent this type of “protection”.
I agree that power imbalance is a significant factor that is different. Fair point.
Why would these parties not easily be deterred by legal concerns? Well there are a variety of groups in this space, many of them are corporate/start ups, and usually they are more adverse to legal risk than your average pseudo-anonoymous citizen.
The statute is essentially "we don't care if what you're doing is legal or not and we're not going to wait for a court to decide, we're playing this trump card to make what you're doing illegal regardless of it's copyright status."
It's equally scummy whether it's over AI training, preventing someone from using material from a DVD under fair use, or suing for Joe Blow for refilling his inkjet cartridges.
No, in none of them are the victims mainly large corporations.
In two out of the three, the scumbags who are victimizing are acting through a large corporation, but I’m not sure how that is material.
Do you not think that the response is justified in the face of the livelihood of artists being actually stolen from them as we speak?
Copyright maximalism makes the creative world worse for everyone in the long run.
Artists don't need to collaborate with the tool running their livelihoods either.
Ruining the lives of actual creatives is making the creative world a worse place right now, so spare me the thought for the long run.
Besides a diffuser trained just in public domain is going to be as dangerous to their job, they seem to generalize well enough, and if their style is anywhere close to anything existing pre 1920 they will have ai spitting out the exact same thing they do and no recourse.
And to get to the point, actual creations can use these tools to increase their output then fold or more. A reduction in commission price is to be expected, as is expected for them to use the tools to produce more.
They are not outcompeted by ai, they are outcompeted by a new crop of artists bracing ai tools to work at a greater volume and thus can sustain a lower price point. More art will be produced, not less, and those in true danger are the artisans that aren't embracing the ai industrial revolution.
Seems tech won’t be satisfied till its devalued the entire human experience of growth and effort and replaced it with quick answers and cheap results.
Edit: Then one day when the species goes extinct because we’ve got our robo-waifus we can exclaim in our final breathes that the luddites were wrong. All technology is great, no issues whatsoever.
A rising tide really does lift all boats, whether you believe it or not. History furnishes no exceptions to this rule.
To those looking at becoming politically active in this area, it will be much better to work to democratize AI than it will be to try to stop it. The former will be difficult, the latter impossible.
It doesn’t, though. The models that are transforming the markets now would continue to do so, in much the same way, if no additional art from working “traditional” artists was used to train either base models or fine tunes.
Sure, if this existed before the models – which it couldn’t, since it incorporates one of them directly – and had been applied to all existing (including public domain) art and it was completely effective and had no immediate countermeasure, it could have obstructed or delayed things, but now its trying to unring the bell, and it can’t, even if it worked flawlessly on a technical level.
> The robot can't train on its own output because it doesn't understand what it's supposed to be drawing.
Actually, a very significant source of training data for new training of models is…output of existing models curated by the trainer. The robot can be trained on its own output.
> Artists don’t need to collaborate with the tool running their livelihoods either.
Ironic statement, given that they do even to use Glaze, which incorporates exactly the tool they are blaming to do its work. In fact, the entire poisoning attack is done by injecting other artist’s styles into their work, in the hope of confusing style extraction. That is, the tool literally does exactly the thing it is intended to prevent, as its preventive measure.
To be fair, the economy is in a dump. Reduced commissions and positions being cut might have more to do with that than AI. That said, the market for “generic” art is probably going to be taken over by AI.
Nothing is taken over by AI, but just young clinical psychopath types are being enabled to be burned later at the expense of all. That's kind of wrong.
Also - if I want to get, say, cover art for my book - why should I hire a human if Stable Diffusion works just as good? I'm also not interested in ripping off a particular artist's style. Do those artists really think that I'm going to hire them after they opt out of being included into ML datasets? Nope - I'll just experiment with SD a bit more to get the style that I want.
Are you sure it works like this? Because SD training usually starts with resizing, cropping, and encoding to latent space. If applying a filter is clearly prohibited, the current approach of "resizing, cropping, encoding" is surely too, right?
b) The DMCA is a US legislation, what about training done in other countries?
Maybe? It certainly prevents reusing the work in that specific manner without authorisation. It doesn't prevent just seeing the work though, so it's a bit up to interpretation.
To b), US-like copyright law effectively applies in any WTO-conforming country. Anything that implements WIPO functiouns roughly the same way. Fun DMCA fact, there is no fair use provision. Any use is likely to be criminal. Another fun DMCA fact, apparently ripping out spyware and republishing is entirely legal. Huh.
> It certainly prevents reusing the work in that specific manner without authorisation.
Well, technically speaking, it doesn't prevent it. It just messes up the results of the work being used in that manner. And what if someone builds a training workflow that can just ingest such changed images and use them without being negatively affected by the changes?
> so it's a bit up to interpretation.
Interpretation that would likely have to be decided in court, and likely in a very drawn out and very very very expensive manner, with uncertain outcome.
A machine-readable tagging that simply says "noone is allowed to use this for training AI" sounds way easier to argue in court to me.
That is what encryption does as well. You can certainly attempt to watch DVDs without permission. It won't be very enjoyable. And what if someone builds a viewing application which can just watch those DVDs anyway? You see, if this is legally protected, building that workflow is circumvention and very, very illegal. Defining "effectively restricts" is left intentionally up to interpretation because there is no clear line between messing up the result and preventing access.
Encryption prevents usage of the data. This doesn't. The data can still be viewed without any special software, device, password, key, etc. A sufficiently robust ingestion engine could still use it for training. In fact, even an unprepared engine can train on it, it only messes up the outcome.
Honest question: Is it harder to argue legally that I DRM-protected my work if I publish them in a form that needs an encryption key/software/device, or if I publish them for everyone to see after changing some pixels around?
> if this is legally protected
"If" is the important term here. It was mentioned above that "maybe" this counts in the same way as a Copyright protection measure.
I don't argue against that. Maybe it does. That is for lawmakers, courts and similar legal experts to decide.
My opinion as someone who isn't a lawyer, is that it would be EASIER to get courts to agree on that machine-readable tagging, simply disallowing usage of works for training, is similar to DRM measures, and ignoring them should be punished the same way as circumventing copyright mechanisms.
The added bonus for users: Such tagging is easier to implement, easier to update, there exists prior law already covering it (see several european countries) providing legal guidelines. Plus, artists wouldn't have to mangle their works to implement them, and it is useable with all forms of data, not just images.
It should be. It'd be really great if it was. Sadly, lawyers wrote the DMCA. It has to be a measure which actually restricts access in an effective enough way, just saying "don't touch this" isn't a measure because it doesn't effectively prevent the usage. 17 USC § 1201(a)(3) and all that.
If training on image sets isn't copyright infringement, "don't use this" doesn't count. It's a license, and you're not infringing on it. If it is copyright infringement, "don't use this" is the default and you can't use anything without explicit permission, effectively requiring datasets to only include CC0 images. Since the first one is way more likely to be true, you instead use the 17 USC 1201(a) which prevents circumvention of technical protection measures, by creating what is hopefully a technical protection measure. Is it foolproof? It was never meant to be. It's an attempt at best. But it's better than relying on a law which is incredibly likely to never apply to dataset scraping.
Conversations aside, laws can adapt to include terms regarding tagging. As said before, there are legal examples for this, eg. in the EU:
https://discoverdigitallaw.com/is-web-scraping-legal-short-g...
Quote:
the new law, that must be applied by all EU countries until 7 June 2021 (Directive (EU) 2019/790 on copyright and related rights in the Digital Single Market or ‘DSM Directive’), in its Article 4 provides an exception from the rights of the database owner mentioned above in case of ‘reproductions and extractions of lawfully accessible works and other subject matter for the purposes of text and data mining’ unless ‘the use of works and other subject matter referred to in that paragraph has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online’.
Again, I'm not a lawer, but to me that seems like it's up to lawmakers to do their homework, and update existing laws to deal with the reality that
a) data mining exists and is useful for lots of things
b) people want to make their works available publicly, and therefore ...
c) people publishing works need a workable, stable and reliable way to tell others whether they are okay with their work being scraped and used for analysis/training/etc. or not
And as I said above, ideally such a solution doesn't require mangling the published data in some way, and works for all kinds of data.
The cloak is absolutely meant to prevent copying of the image. Thd cloak is only effective against ML, not humans viewers.
When CSS encrypts a DVD, it scramble the bits so your unauthorized computer can't decode it. You could decrypt it by hand.
This is not a problem from the start as you are reading it but not coping it.
And as far as I know an artistic style is not just protected because someone uses a style.
There are million artists out there. Style references a certain amount of rules.
You can't just protect styles as this would also hurt artists
This is a hedge against courts deciding scraping data for training purposes is valid.
Maybe you're allowed to scrape data, but with this, now you are applying a filter (creating a derivative work) to defeat a copyright protection mechanism, both clearly prohibited in current law (US jurisdiction at least). For any serious player scraping this opens your buisness up to huge lawsuits. For any serious player making tools, you'll specifically avoid defeating these techniques. For any minor player you'll now have to go to the backwaters of the internet for tools to do this that you hope won't steal your bitcoins.
Every notable artist will be only upload their art to sites that offer something like this, paid at first, but when the cost are low enough, pretty much every site that wants art content will offer it.
This isn't a technical solution to this problem, it's a political solution that happens to use tech.