Glaze and Nightshade are different methods. Glaze makes the AI incorrectly identify the art style and Nightshde makes the AI incorrectly tag the image content. The first one prevents the artist’s personal style to be learned by AI and the second one the specific image content (while also poisoning the model)
I really think the only difference between a religion and a cult is wether the founder is alive or not.
At which point we normalized listening to opinions of companies’ owners?
The ones that say that Nightshade is not working is because they don’t understand how it works. They “test” it by asking a LLM what image they see and it usually identifies it without any problems. The actual function is thay when the image is used to train data, it provokes errors in tagging the image. So a poisoned image of a car is correctly identified as a car by ChatGPT, but when is used to train the model, that car is used to train images of idk cakes.
For text there are a lot of interesting tarpits, like this one https://github.com/amenyxia/Sarracenia or the original one called Nepenthes
This is the easiest one https://rnsaffn.com/poison3/ but there are more advanced ones that you can self host that feed an infinite stream of poison, although LLM crawlers are hungry creatures and would keep a % of your servers doing that
All websites and services I managed are filled with poisoned data, text or images. At least 15% of my total processing power is spent on generating the poisoned text thata, but I’m glad to do it. Take more action people, not only “protest”
So, those are the onest floor, the twest floor, the threest floor…