Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Nightshade: An offensive tool for artists against AI art generators

nightshade.cs.uchicago.edu

Post reply on HN
Print view View original

501–510 of 710 posts

Re: Nightshade: An offensive tool for artists against AI art generators

#501

Post by ittseta »

jerbear4328 wrote: post #483
kmeisthax wrote: post #378

Earlier quoted context omitted.

The screenshots you sent in [1] are inference, not training. You need to get a Nightshaded image into the training set of an image generator in order for this to have any effect. When you give an image to GPT-4V, Stable Diffusion img2img, or anything else, you're not training the AI - the model is completely frozen and does not change at all[0]. I don't know if anyone else is still scraping new images into the genera…

[3] sounds really interesting - do you have a link?

https://www.nature.com/articles/s41467-023-40499-0 https://deepmind.google/discover/blog/images-altered-to-tric...

Study on the Influence of Adversarial Images on Human Perception

Re: Nightshade: An offensive tool for artists against AI art generators

#503

Post by ptdn »

kmeisthax wrote: post #378
542458 wrote: post #48

This seems to introduce levels of artifacts that many artists would find unacceptable: https://twitter.com/sini4ka111/status/1748378223291912567 The rumblings I'm hearing are that this a) barely works with last-gen training processes b) does not work at all with more modern training processes (GPT-4V, LLaVA, even BLIP2 labelling [1]) and c) would not be especially challenging to mitigate against even should it become…

The screenshots you sent in [1] are inference, not training. You need to get a Nightshaded image into the training set of an image generator in order for this to have any effect. When you give an image to GPT-4V, Stable Diffusion img2img, or anything else, you're not training the AI - the model is completely frozen and does not change at all[0]. I don't know if anyone else is still scraping new images into the genera…

The context windows of LLMs are now significantly larger than 2048 tokens, and there are clever ways to autopopulate context window to remind it of things.

Re: Nightshade: An offensive tool for artists against AI art generators

#504

Post by scheeseman486 »

kmeisthax wrote: post #378
542458 wrote: post #48

This seems to introduce levels of artifacts that many artists would find unacceptable: https://twitter.com/sini4ka111/status/1748378223291912567 The rumblings I'm hearing are that this a) barely works with last-gen training processes b) does not work at all with more modern training processes (GPT-4V, LLaVA, even BLIP2 labelling [1]) and c) would not be especially challenging to mitigate against even should it become…

The screenshots you sent in [1] are inference, not training. You need to get a Nightshaded image into the training set of an image generator in order for this to have any effect. When you give an image to GPT-4V, Stable Diffusion img2img, or anything else, you're not training the AI - the model is completely frozen and does not change at all[0]. I don't know if anyone else is still scraping new images into the genera…

Hey you know what might not be AI generated post-2021? Almost everything run through Nightshade. So given it's defeated, which is pretty likely, artists have effectively tagged their own work for inclusion.

Re: Nightshade: An offensive tool for artists against AI art generators

#505

Post by webmaven »

kmeisthax wrote: post #378
542458 wrote: post #48

This seems to introduce levels of artifacts that many artists would find unacceptable: https://twitter.com/sini4ka111/status/1748378223291912567 The rumblings I'm hearing are that this a) barely works with last-gen training processes b) does not work at all with more modern training processes (GPT-4V, LLaVA, even BLIP2 labelling [1]) and c) would not be especially challenging to mitigate against even should it become…

The screenshots you sent in [1] are inference, not training. You need to get a Nightshaded image into the training set of an image generator in order for this to have any effect. When you give an image to GPT-4V, Stable Diffusion img2img, or anything else, you're not training the AI - the model is completely frozen and does not change at all[0]. I don't know if anyone else is still scraping new images into the genera…

Even if no new images are being scraped to train the foundation text-to-image models, you can be certain that there is a small horde of folk still scraping to create datasets for training fine-tuned models, LoRAs, Textual Inversions, and all the new hotness training methods still being created each day.

Re: Nightshade: An offensive tool for artists against AI art generators

#506

Post by battles »

SirMaster wrote: post #436
vhcr wrote: post #186

Earlier quoted context omitted.

The animation when you change images makes it harder to see the difference, I opened the three images each in its own tab and the differences are more apparent when you change between each other instantly.

But that’s not realistic? If you have to have both and instantly toggle between them to notice the difference, then it sounds like it’s doing its job well and is hard to notice the difference.

The person who drew it would definitely notice.

Re: Nightshade: An offensive tool for artists against AI art generators

#507

Post by xpe »

jazzyjackson wrote: post #415
xpe wrote: post #363

Earlier quoted context omitted.

Trying to make sense of the above comment is difficult. > Copyright law demands that everyone play along with the lie that is intellectual monopoly. Saying "lie" suggests willful deception. Perhaps you mean "socially constructed"? Combined with "playing pretend" makes it read a bit like a rant. > Then let's use a more precise term that is also present in law: monopoly. Ok, in law and economics, the core idea of monop…

> Saying "lie" suggests willful deception. Consider instead the term "legal fiction", it's not so derogatory.

Legal fiction is a technical term used by legal scholars. To be clear, any legal system is built from legal constructs; I'm not talking about these. A legal fiction has a markedly different meaning than lie as used in the comment two levels above (which seems to me more like a rant than a clear or convincing argument) Are you familiar with specific expert writings about claimed legal fictions in U.S. copyright law?

Re: Nightshade: An offensive tool for artists against AI art generators

#508

Post by dartharva »

password54321 wrote: post #368
r3trohack3r wrote: post #166

The number of people who are going to be able to produce high fidelity art with off the shelf tools in the near future is unbelievable. It’s pretty exciting. Being able to find a mix of styles you like and apply them to new subjects to make your own unique, personalized, artwork sounds like a wickedly cool power to give to billions of people.

Not really. There is a reason why we find realistic painting to be more fascinating than a photo and why some still practice it. The effort put in by another artist does affect our enjoyment.

The word "we" is doing a lot of heavy lifting here. A large majority of consumers can't even tell apart AI-generated from handmade, let alone care who or what made the thing.

Re: Nightshade: An offensive tool for artists against AI art generators

#509

Post by petesergeant »

Theodores wrote: post #469
kredd wrote: post #308

Earlier quoted context omitted.

In terms of art, population tends to put value not on the result, but origin and process. People will just look down on any art that’s AI generated in a couple of years when it becomes ubiquitous.

https://en.wikipedia.org/wiki/Labor_theory_of_value According to Marx, value is only created with human labour. This is not just a Marxist theory, it is an observation. There may be lots of over-priced junk that makes you want to question this idea. But let's not nit-pick on that. In two years time people will not see any value in AI art, quite correctly because there is not much human labour in creating it.

> This is not just a Marxist theory, it is an observation.

Yeah? Well, you know, that's just like uh, your opinion, man

Re: Nightshade: An offensive tool for artists against AI art generators

#510

Post by jedberg »

gfodor wrote: post #175

Huge market for snake oil here. There is no way that such tools will ever win, given the requirements the art remain viewable to human perception, so even if you made something that worked (which this sounds like it doesn’t) from first principles it will be worked around immediately. The only real way for artists or anyone really to try to hold back models from training on human outputs is through the law, ie, levera…

Everything old is new again. It's the same thing with any DRM that happens on the client side. As long as it's viewable by humans, someone will figure out a way to feed that into a machine.
Post reply on HN