Hacker News Reader

Comment by jrm4 | Hacker News Reader

I see why you're saying what you are about the "facts," but I'm not at all convinced that it's so clear-cut that the plaintiffs are all that wrong.

Which is to say, compare these models to e.g. simple image compression. If compressed images can violate, the argument that "this is merely VERY GOOD and VERSATILE image compression" probably has enough legal legs to do something with.

View on HN

It is a bit like saying a random number generator is a form of image compression since it is a small program the output of which contains every copyrightable image.

Compression has the property that almost all inputs are reconstructable. Generative models have the property that almost all inputs are not reconstructable.

Compression with extra steps is still compression. Most people don't understand the math involved in LZF-encoding, and lossy audio encoding seems even more strange, so "diffusion" just sounds like a somewhat-lossier form of compression if described in a certain way.

I was very dismissive of the anti-AI-image crowd until I saw that Anne Graham Lotz pic. Only a single known source image of her exists, although many copies of it are distributed far and wide across the internet, and using her name in a prompt results in an image that seems very much like a lossy reproduction of the "original." It's an edge case, to be sure, but that demonstration looks like lossy compression, even though it's not.

> I was very dismissive of the anti-AI-image crowd until I saw that Anne Graham Lotz pic. Only a single known source image of her exists, although many copies of it are distributed far and wide across the internet

https://duckduckgo.com/?q=Anne+Graham+Lotz&t=brave&iax=image...

I see dozens of images of this woman (who is apparently some kind of minister). Okay, I'm clearly missing something here. Can you explain further?

Also, besides the 20 or so copies of this specific pose I see on the first page, I also see numerous others of her either in a different pose or in a different (but similar) outfit -- she's apparently fond of that color combination.

I just put my local copy of Stable Diffusion to work cranking out images based on an "Anne Graham Lotz" prompt. So far, none of them look anything like the original image. It has generated one of her wearing a similar-colored shirt, but the pose is completely different, and as noted above, she seems to like that color.

My mistake! I mis-remembered the original Google paper[0], which stated only that the image was in the Stable Diffusion training set.

The complete set of "matches" they got were interesting, even while the number was very low.

0. https://arxiv.org/abs/2301.13188

From the article: researchers were able to reproduce training images (and bad versions of them at that) 0.03 percent of the time. Not 3 percent, mind you, but 109 out of 350,000 images could be reproduced. And if you go back to the paper, these tended to be images which were heavily duplicated in the training set, so the failure modes should be easy to guard against...

Sure, but I strongly suspect that in court the visceral impact of "Look, there's the image, it was in there," of even just ONE might just be enough to sway a judge or jury, percentages be darned.

Eh, there's a lot of space for argument and deliberation in a courtroom, moreso than a hacker news comment section... This is a rare issue with an obvious fix, so I wouldn't expect a ruling to hinge on it. (Additionally, the Lott image doesn't have any real harm; it's heavily reproduced already and freely enough licensed to appear on Wikipedia.)

My guess is that because there's a lot of these cases coming the courts are really going to want to come up with some tests for whether a model is ok or not. So even if they say SD did a bad thing with these few images, that's not going to be enough to rule out the broader class of algorithms - next week they'll have another Getty case against the algorithm with the technical mitigation for the rare copied images.

But a "rare issue with an obvious fix" sounds like a good Streisand effect type-target to me.

Technologically I know that this is a pretty honest statement of the problem, but I'd expect A LOT OF PEOPLE to read this as "oh, look, they're trying to tweak it to 'destroy evidence.'"

Reply on news.ycombinator.com