Hacker News Reader

Comment by kweingar | Hacker News Reader

This is going to be an arms race. Inevitably Glaze will be thwarted, and a new version will need to be made.

Generative art models will always have the advantage in this fight. For them, they just need to crack the protection and update their model. For artists who want to protect their work, they’ll have to continually update their whole portfolio of images and republish them whenever a new version of Glaze comes out (and attempt to ensure that their images with old versions of Glaze don’t end up in a training set, which will be difficult if not impossible).

View on HN

Right, so off the bat, if you're going into a situation with your generative art model, and you're first thought is, "we're totally going to break through this attempt at someone trying to protect their artwork"....aren't you automatically in the wrong?

Not if you don't believe in Imaginary Property.

All property is imaginary.

It also turns out to be useful in often helping human beings coexist relatively peacefully and sometimes negotiating mutual interests productively, so like a number of things that don't exist until enough people agree to conduct themselves as if they do, it's probably a good idea in general.

Intellectual property has some distinct characteristics, but it's fundamentally the same.

And assuming entitlement to any available data for ingestion into a training set without a moment's regard to the labor and agency of the people who produced it has things in common with the ethic of a thief who, coming upon your stuff, decides that if they can take it, it's theirs, society's construction of property be damned.

Except that intellectual property can be copied while leaving the original behind, but this is a centuries-old debate and I doubt we'll break new ground going 'round the mulberry bush about it again here.

(... I'm literally wearing a "YOU WOULDN'T REIMPLEMENT AN API" t-shirt as I type this.)

You'll note on reviewing my comment that I didn't claim there are no differences. That's intentional, having tread the relevant ground plenty of times myself.

The similarities matter more than the differences here.

And the problem really isn't breaking new ground in general. People saw the similarities at least as far back as the printing press. The problem is getting people to care about the claims others have on the fruits of their labor, and why incentives might matter for everyone.

If technology causes the value of the fruits of that labor to crash, then... The value of the fruit of that labor crashes. No more and no less.

The invention of the printing press didn't stop writing, the invention of the camera didn't stop painting, and the invention of Stable Diffusion won't stop creation of novel art. But it definitely upends a known methodology for extracting value from non-fungibility of labor product.

In any case, the labor-to-value question is moot because Stable Diffusion will enable companies to generate unlimited visuals off the single-time-compensated manual labor of a handful of artists (for no other reason than some artists will take that deal). So whether any one artist is in the training set will become irrelevant as people choose nearly-free good-enough product over far-more-expensive handcrafted product most of the time.

Glaze is useless when an advertising firm is fairly-compensating consenting artists to toss their work into the meat grinder and that grinder then still churns out work that swamps the individuality of traditional-methods artists.

> If technology causes the value of the fruits of that labor to crash

Well then, if the fruits are worthless, then presumably there's no loss from excluding it from a training set, and creating norms or even laws which allow people the privilege of negotiating the basis on which their work can be so used, right?

Of course, nobody believes that. Much like no one really believed that unauthorized copies of 18th century print works were valueless just because the production could be industrialized. The enterprises producing unauthorized copies did so because they knew full well they could capture the value... without any of the pesky pro-social obligation to respect the labor and interests of those who created it.

The interest in the works as training data betrays the position. This work is valuable, and works derived from it are valuable.

> Glaze is useless when an advertising firm is fairly-compensating consenting artists

I imagine there will be artists who opt in after negotiating the basis on which their work can be used -- although who knows what kind of surprises there could be there, some might even demand equity or collectively bargain.

But deciding there's no need for a framework of negotiation and consent just because some people will get to yes sounds like a really bad precedent.

And if the argument is that this isn't sufficient to protect interests, well, sure. Necessary but insufficient points exist all the time, usually the thing to do is pair them with other axioms / initiatives.

Or simply don't, and let the old method wither.

The printing press didn't end copyright, but it obviated the need for hand-illuminated manuscripts. I expect automatic image generation will do something similar to hand-created from-scratch art. And yes, if there's a framework to be created to allow artists to indicate that they don't want to be involved and wish to lead the charge into irrelevance, it should be created. It's best to have a revolution be tidy not messy.

Wrong? It's just a fact that generative AI can simply train against these "protected" images. I don't think the parent said anything about ethics, if this is even an ethical issue in the first place.

Unfortunately as we have learned from open source software, it doesn't matter if it's write or wrong. The only way to enforce licensing is to litigate.

The decades of GPL violations have taught us the only way to get commercial interests to not abuse copyleft licenses is to force them via the courts, and in doing so dis-incentivize other businesses from violating the licenses.

But it isn't 100% effective by any means, and usually relies on a large commercially successful organization to already have aligned interests in enforcing their GPL/copyleft licenses.

Anything else is just ethics paper fodder.

Maybe so, but that’s not going to stop people.

The tech industry is full of examples.

You wouldn't even need to try to break it. The goal of the model is to learn patterns and pass tests against the training data. Given that these images are expected to ned up in training sets, then eventually whatever technique is being used here will fail to thwart the training process.

Unless it’s specifically trained not to, I’m fairly certain AI will bypass this without even trying. Such is the way of the unsupervised hill climb on huge unstructured datasets. To put it another way, in order for an artist’s style to truly be invisible to AI, it would have to be invisible to humans.

If bypassing "protection" like this makes you automatically in the wrong, does that also make it immoral to try to refill ink cartridges?

Imagine a GUI-less version of Glaze designed to be plugged into other systems. Stuff it into a plugin for your portfolio CMS. Add it to the popular gallery site you run. Add it to the tool you use to post art to a half-dozen galleries with one click. When a new version of Glaze comes out, update it, and push the button that starts to re-glaze the original copies of the uploaded images, and replaces the publicly-displayed stuff.

Imagine a CMS that takes 40 minutes to return a single image.

More seriously, reglazing old images isn't very helpful, since they will amost always be already in training datasets by that point.

Well...

1) That's not how it works.

LAION is a set of urls, not images.

Actually distributing the images would be blatant copyright infringement and, even if you don't care, is prohibitively massive.

So, yes, modifying the source images will be effective against new players who want to scrape the dataset... and yes, as the tech to 'build your own stable diffusion' becomes more commoditized, you can be 100% assured that multiple groups will be doing it in the future.

...this also has a FAQ that states even a limited number of poisoned images (eg. most recent art by Greg) can cause artifacting in outputs.

2) ...but more importantly, are you joking?

Artists complain and artstation does what, update their TOS?

That obviously does nothing.

So, companies start inventing this anti-ai watermarking tech that shops like deviantart and artstation can pickup and put into their pipelines and make into a pro-feature.

If you think it's not going to happen, you are 100% deluding yourself. The playbook is so obvious, I'd be amazed if it's not already a WIP for these places.

Artist friendly. Make money. Big players can work around it and small players are screwed out of any good images to use for training.

It's going to happen.

3) '40 minutes to return a single image...'

Oh come on. Are you objecting because you think the process has no technical merits, or because you don't like it?

Technically, this isn't a limitation. These places already have massive image processing pipelines for resizing, thumbnailing, etc.

LAION is currently a set of URLs.

That's because AI art was a mere research toy back then.

Now it has massive economic potential, do you really think downloading say 100TB of images is a big deal to any company? Less than $100000 to get an invaluable database that's safe and sound. Its also not prohibitive at all to just clone this data into a NAS array, and just literally ship the hard drives to whomever needing it.

Google has already downloaded terabytes of images and they store resized thumbnails of those images on their own servers. If google can do it why can't everyone else?

> More seriously, reglazing old images isn’t very helpful, since they will amost always be already in training datasets by that point.

And trained into models.

I think the issue here is that the old versions will still have potentially been crawled.

On the “plus” side, glaze will cost commercial AI model creators both time and money. And they won’t really hurt academic research use cases.

Kind of a David and Goliath situation (yes, David is an acknowledged sniper).

> And they won’t really hurt academic research use cases.

How do you figure?

It’s just additional data to academics - something that helps them identify issues and hallucinations in their models.

They’re not trying to produce sellable images; free adversarial images are useful to them.

> ; free adversarial images are useful to them.

Only if they were correctly tagged/ labelled. Of course, academics always have access to cheap labour, aka undergrads.

How is that not also true to commercial AI model creators who will train against these adversarial images?

Reply on news.ycombinator.com