LAION-5B is notoriously badly labeled that having a few poisoned data, even if it worked as advertised, would literally not matter at all.
Plus, it’s not doing anything to existing diffusion models that used LAION-5B, since many artists are under the mistaken impression that the models will constantly scrapes the Internet for new images and train on them automatically, when training a model to learn new information without catastrophic forgetting is almost impossible (hence, workarounds like LoRAs and such).
Again, a reminder that the creator of Nightshade and Glaze, Ben Zhao of UChicago, is literally a code thief who stole GPL code for his closed source product (warning: reddit link) to scam artists who doesn’t understand the tech behind ML models.
Hmm, looks like it would also mess up classification, recommendation, captioning, etc models using these images. Maybe image and duplicate search as well? Maybe could be used to get around automated copyright strikes?
I’m only vaguely familiar with ML datasets and have only trained on local data, but I’ve never heard of this? Can anyone provide some evidence this is the case?
Edit: Looking further I can still only find datasets containing the image files, ex.
“it takes all the running in the world just to stay in place.”
Normally refers to biology arms races, where a poisonous animal and a poison resistant predator play tit for tat, making stronger poisons and stronger resistances to try and outplay the other just to stay alive.
Now, artists and AI are doing the same. AI wants to steal art without paying for it, artists dont want their art stolen. Artists come up with little tricks to poison the data set if their art is used, AI comes up with little tricks to strip the poison from the data.
The dance continues, the dancers straining and struggling, all to stand still.
Ah, similar struggle my dad described when negotiating sales of complex systems, both parties start with unrealistic demands, just to have stuff to give away to the other side during negotiations.
He has told me several times that he just wishes that the process was way more streamlined and that the parties could start closer to the realistic goal.
He has since retired, so he no longer needs to deal with it…
FFS. People. It’s “stealing” as much as visiting a museum and going home to sketch/sculpt/compose is theft.
When did the chicken-little mindset of old fucks become default reaction set for the whole damn world? 🤮
Edit: Ah, yes. More uneducated armchair experts yelling rhetoric. How surprising. Please, tell me you came from Reddit without telling me you’re from Reddit. 🤣
It’s a little more nuanced as AI models keep spitting out verbatim training data when people figure out the right queries. E.g. the other day they banned you from asking chat GPT to repeat the same word forever as it’d do that for a while, then just spew out something it had been trained on. If someone reads thousands of articles on a subject, then writes the exact contents of one of them, that’s definitely plagiarism.
There’s also the issue that when a human reads a lot, they have to pay for a lot of books and view a lot of ads and pay taxes that fund a library system buying books, too. The human extracts value from what they’ve read and gives something to its author. Megacorporations training AI models are only extracting the value and aren’t paying for the privilege.
I said it was more nuanced than you said, which is pretty different to saying it’s not like you said. There are big similarities to a human who’s read lots of books, but the equivalent human is pirating all their media while being rich enough to pay for it, and sometimes passing off other people’s writing as their own. Neither of those things are allowed when humans do them.
Man, thats so far off base youve become the first human on pluto.
You think and reason. A math program running on a big computer doesnt. So, already, you are wrong.
But the theft and plagiarism isnt from the output from the large math program that chatgpts marketing team want you to think is artificial intelligence.
The primary problem is that they never had permissiom to feed other peoples things into their picture calculator in the first place.
It would be like if I illegally printed someone elses artwork on a canvas and used it at my job as an interior decorator to decorate an office. Or if I illegally printed a copy of a board game or card game and used those copies to host a pay-to-enter tournament.
Its a calculator, not a brain, and feeding it materials you do not have the permission to feed into it is theft.
There are highly degreed adults working in this field who disagree with you. But good job calling everyone who disagrees with you a child. That’s super mature brah. You’ve claimed the moral high ground.
Except you don’t put in the effort to go yourself and interpret the art or the effort to make it. The human element is completely removed; besides, people should do what they want with their art, including prevent AI from using it to the best of their abilities.
“AI” image “generation” has been known to spit back out more or less intact copyrighted works, complete with watermark. It doesn’t create anything it’s just an outright plagiarism machine.
I believe this is suggesting an AI poisoning edit, not removing the image entirely. It should be mostly imperceptible. Plus, you could update with newer methods as they come out.
Add comment