2y ago

PSA for artists

https://mastodon.social/@tess/111784134039294048

39 comments

If you replace all your online images then ai can't look at it. No one else can either, but you stop ai I guess.
- They're talking about using this: https://nightshade.cs.uchicago.edu/
  
  Isn't Nightshade defeated by just applying an anti aliasing filter to the image?
  
  I hope content hosting services start applying it by default.
- 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.
- You could generate a different temporary img url every time and nightshade it after the link expires.
- You could generate a different temporary img url every time and nightshade it after the link expires.
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.
https://www.lvisdataset.org/dataset
https://www.v7labs.com/open-datasets
- LAION is one of the big dogs (https://laion.ai/). Their datasets consist of urls and metadata.
  
  Ah perfect, thank you so much!
  https://github.com/rom1504/img2dataset
  Seems to be the main tool, I’ll have something new to explore this weekend.
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.
- Tell me you have no idea how it works without telling me.
  Edit: sorry that's unclear, I mean the OP on mastodon not you specifically
  
  "tell me without telling me" needs to fucking die
Could also Goatse them
- c/foundsatan
- This is the preferred method
- Fans of the actual content might have some issue with it
TIL about Glazing and Nightshade. Thanks!
- They don't work as advertised.
- I'm still confused.
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?

39 comments