so this is a bit of a complicated topic! i will do my best to summarize in a way that will make sense to people with 0 experience w/ AI, but it's a little difficult so bear with me. i'm also sorry for how long this is
basically, the kind of art theft that GLAZE is made to prevent is people using AI text-to-image generators to imitate an artist's art style. they can feed a relatively small number of artpieces into the generator, and it can spit out work that looks like it was produced by the artist in question:
GLAZE is intended to disrupt this process by preventing the text-to-image generator from recognizing the art style that's used. the way it predicts how much it can modify a photo before AI can no longer recognize it is.... hard to understand if you don't have a background in this stuff, it's a bunch of fancy code basically. but the VERY simple explanation is this: it takes the art and creates a different version of it, which is in a completely different style. below you can see:
on the left you have the original art. the third column, titled "GLAZE target style", are the altered versions - as you can see, they have been intentionally turned into different art styles. Ortiz's art style is realistic, but GLAZE turned it into a Gogh-style oil painting. it will then, basically, blend this altered version with the original image (called 'perturbation'). by doing this, the AI will read the image as a Gogh-style painting instead of a realistic portrait, so it will just produce Gogh-style paintings - completely ruining the mimicker's efforts.
"invisible to the human eye" varies depending on the person and ALSO on the settings used. you can add more perturbation and it will be more obvious that the art was altered, at least to the artist themselves. here's some examples, with increasing levels of perturbation:
these changes are still very small to us, but perceptible to text-to-image generators, which rely on these very small differences to fine-tune the reproduced artwork.
all of this information was pulled from the associated paper, which is currently under peer review. i'm not sure if GLAZE will be able to outsmart text-to-image generators for very long (and rn it doesn't work on all of them) but it's a very cool strategy.
vaporwave
turnnoffyourmind
dailybridgerton
izzy806
prguitarman
kianamaiart
baratheonpilled
ellievsbear
discoholicmusic
