xormetric: wow so glaze doesn't work and never did and the whole idea of "protecting" art from training is fundamentally flawed, who could've guessed! arxiv.org/abs/2406.12027

wow so glaze doesn't work and never did and the whole idea of "protecting" art from training is fundamentally flawed, who could've guessed! arxiv.org/abs/2406.12027

Adversarial Perturbations Cannot Reliably Protect Artists From...arxiv.org Artists are increasingly concerned about advancements in image generation models that can closely replicate their unique artistic styles. In response, several protection tools against style...

Wow who could have known.

sad that so many people were duped into spending their time and compute on this. saw many posts like "wow this really makes my computer hot"

… are you saying everyone knew? (I've barely heard about it)

Ive been saying this basically since it came out

bsky.app/profile/why....

If enough people start using glaze so that it becomes common enough in the training sets, the decoder will (i claim) learn to accurately extract the style again.

bsky.app/profile/why....

We actually have been talking through some ideas here, unfortunately Nightshade/Glaze don't really work (its a neat attack on specific existing models but wont actually protect against any new models or clever plagiarists). More to come soon once we nail down our approach!

bsky.app/profile/why....

I'm not the dev who would be touching anything related to a feature like this, but I do do a decent amount of ML stuff. I'm very pessimistic that glazing could actually work here, model training already expends a lot of energy 'cleaning' up the picture and making it more accessible to the model

it's hilarious to consider that even this skepticism is *overselling* glaze. it's not even safe from naive mimicry, like in the paper there's: "We mimic the style [...] using: an alternative off-the-shelf finetuning script from HuggingFace [...] with no effort to bypass Glaze’s protections."

I had a hunch, heh

I wonder what percentage of their operational costs come from the glaze-ing. IMHO it's a waste of GPU time, but it sounds like its a feature that users want regardless

Bad Day for DRM lovers

I kept saying this over and over again, yet many in the art community will get angry at me for saying this. I’m glad I’m not actually crazy this time. Thank you for sharing this.

quelle surprise

Ah, my ignorance was what gave me brief hope. As usual

afaict glaze is a targeted attack against clip. clip is actually one of the shittiest things in diffusion modeling right now and we absolutely need a better replacement. clip had _very_ undiverse training data (eg. very little digital art), and didn't learn to encode positioning information.

their noisy upscaler technique works extremely well, from what I can tell almost all of the visibly glazed portions of the image is gone after what's essentially adding noise then upscaling the image (tbf they tested v2.0, v2.1 claims to have fixed sth?). also, someone tell cara.app about this

My doubt (and I am asking without knowing how this works) is if it gives some protection against scrapping bots and the general algorithm. If someone says "you" and inputs a specific image, I assume they would find a way. But at least I am ok with a way to be protected from mass scrapping and imput.

The paper finds that glaze overfits (lol) to protect against one specific algorithm, which isn't the only one. "We mimic the style [...] using: an alternative off-the-shelf finetuning script from HuggingFace [...] with no effort to bypass Glaze’s protections." So no, even mass scraping will work.

I guess that this will be a race of constant updates. Which is a drag because it means reglazing old stuff. But it's that or throwing the towel. In a worldwide scenario, copyright can only do so much in certain regions.

This reminds me of that ship that scrapes plastic from the ocean. It has lots of critics because it does very little. But at least something is better than nothing.

Post