deglazer

program to remove glaze / nightshade / glaze2 from images

requirements

pip install -r requirements.txt

usage

python run.py -h

python run.py clean image.png

python run.py clean_folder path/to/images --newfolder path/to/cleaned/images

methods

• vaeloop [broken since i forgot how to make vae code work lol]

+ most consistent
~ passes image through VAE, which removes imperfections (e.g. glaze)
~ resizes images to a mult of 8
- needs a gpu

• glaze1

~ classical glaze remover (has it even worked at all?)
- could remove softer details

• glaze2 [broken since i forgot how to make vae code work lol]

+ both vaeloop and glaze1 combined
~ passes image through VAE, which removes imperfections (e.g. glaze)
~ resizes images to a mult of 8
- could remove softer details
- needs a gpu

effectiveness

the following claims are from my testing and not backed by any papers, take with a grain of salt (or just ignore idk)

considering that glaze/nightshade doesnt work at all, this is extremely effective (100% captioning accuracy after deglazing)

training effectiveness decreases after deglazing since glaze actually helps training (the glaze/shade acts as noise offset)

faq

arent the filters destructive and remove data?

yes, the filters inherently remove data, but, so does glaze; in terms of this removing stylistic / defining attributes, it could potentially remove some.

would this affect ai training? no (at least not enough to cause severe damage). this is due to SD / FLUX / most models using a VAE, which compresses the image and decompresses it, effectively "ignoring" the artifacts. this is also why i say that "glaze and nightshade have no effect on training", for captioning, they are targeting the completely wrong architecture (not targeting eva / vit)

what does glaze actually do?

basically nothing; it adds adversarial noise which supposedly makes the model think a dog is a cat

the issue is SD doesnt really care about that since we pass it through a VAE, but it does confuse the CLIP (see: nightshade)

vaeloop will fix the 'glaze' and make it CLIP taggable (even though most people dont use CLIP) need to fix vaeloop

why did you make this

glaze n co think that they've made a miracle cure for ai training on other art; they have not

is it possible to actually make images untrainable

not without making it terrible to watch/private

pick 2 of 3 things:

• untrainable
• good to watch
• publicly viewable