HN Explorer

13 comments

[ 3.8 ms ] story [ 46.1 ms ] thread
19h
If you're afraid of crawlers that will feed into AI datasets, why not just render images into canvas elements?

This kind of watermark makes images bad for everyone, not just the AI. Do you really want to see this on every website you visit?

wongarsu
Most artists don't have control over the way their image is rendered. They just use Deviantart or Instagram, because that's where the communities are.
nomel
Sounds like a good feature to attract artists, with relatively little effort.
creata
If a site as big as DeviantArt tries something like "draw it to a canvas", people will just build automatic workarounds.
XorNot
No people just won't be able to find anything on DeviantArt.

Let's not pretend there is not a quid-pro-quo here. If you render to canvas, then you can't be indexed by Google Image Search. If you can't indexed by GIS, then how are most people going to find anything they're looking for?

You could entirely prevent direct indexing by crawlers by simply throwing authentication in front of the image requiring people to agree to a TOS before viewing it but that's not done because in reality 99% of what's there is of no particular value to anyone to be worth the hassle.

EDIT: Although I'm also not seeing how this would protect against anything? The AI generators are de-noising algorithms, and the effect of this is simply to add noise to an image. But their whole mechanism of action is based on whether they roughly know what the things in an image should look like, and then to denoise the image towards that.

dymk
Protects the art from being enjoyed by the human, too
rnkn
As always, the best way to protect your house is with the law, not a better lock.

Regulation is constructive, deregulation is destructive.

Paul-Craft
And yet my car got stolen yesterday.

Got anything else?

Loocid
I would strongly argue you need both. There will always be people willing to break the law.
dymk
so is regulatory capture constructive or destructive?
jahewson
This is delightfully misguided because diffusion models work by learning to remove noise.
ballenvironment
Diffusion models learn how to denoise noise into images that are similar to training data. But what if the training data itself is noisy? Then the model will learn to produce noisy images also.

There's a reason why they try to remove noisy/blurry/bad q images, because simply put, what you put in is what you get out. While I don't agree with intentionally destroying the quality of images (ruins it for humans as well as AI), I don't see why this wouldn't work.