Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Stable Attribution

stableattribution.com

Post reply on HN
Print view View original

351–360 of 365 posts

Re: Stable Attribution

#351

Post by throwaway1851 »

saurik wrote: post #349
throwaway1851 wrote: post #295

Earlier quoted context omitted.

> Clearly seeing at least one photo of me (and AFAIK it was only trained on thousands of copies of that single photo) was absolutely crucial to the construction of this image, and yet this website isn't finding any I think there are two separate ideas that are being conflated here. The first idea is that there is a mapping between text input and a joint text/image embedding space. For that mapping, yes, your profile…

I think a key question in attribution is whether the model would have been able to generate the same result without access to the input, and then how much it would have lost having been restricted from that input. If you remove from the mechanism all of the copies of my profile picture (and there are a lot of them...), I guess I am willing to believe that it might still have enough text descriptions of saurik to come…

That makes sense in a “but for” type of causality. However, I don’t think that’s what this website is aiming for (however flawed or misleading it might be so far as its methodology is concerned). I think the idea of attribution here is more of a visual concept: ie, “which images contributed visual features in the output image?”

For that, if you wrote down the latent embedding for “saurik”, then retrained CLIP and Stable Diffusion from scratch without any of saurik’s profile pics in the training data, it is quite possible that you could generate an image from the embedding you wrote down and it would look the same.

Pure speculation on my part - perhaps your profile pic is the real source material and this website is junk. I just think it’s an interesting and worthy question the site is trying to answer, even if it’s not possible to answer it with much certainty.

Re: Stable Attribution

#352

Post by friedman23 »

josephg wrote: post #73
friedman23 wrote: post #62

Earlier quoted context omitted.

It's completely different. The customizable fonts did not violate copyright.

That’s a sloppy answer. Right and wrong isn’t defined by how copyright law is written right now. That’s kind of the whole point of this debate. Should we change the laws and if so, how?

> Right and wrong isn’t defined by how copyright law is written right now

It's a good thing I never mentioned right or wrong.

Re: Stable Attribution

#353

Post by dahart »

pksebben wrote: post #339
temp512345 wrote: post #328

Earlier quoted context omitted.

It is exactly remembering the pixels. Just not all of them and it obviously fills in gaps (more hair as mentioned in a another post). You can consider the way it stores those pixels as a lossy compression format. If I copy a music sample but I store a compressed version of it (mp3 for example) you will not find the original bits in my database at all. I am still violating copyright.

But it's really not, though. It's remembering something related to the pixels, yeah, but that's like remembering the shape a line can take or the color of the sky. To extend your musical analogy, it's remembering that many songs are in 4/4 time, and that major chords sound appealing. Also, were you to compress anything, an mp3 or a picture, in a lossy fashion, to that degree of compression (~10^-5), you would no long…

In the case of an overfit image, which is the thing Stable Diffusion is being sued over, it is just compression, literally. The image data is stored in the network weights, and the image can be reconstructed. You’re drawing a distinction without a difference.

Re: Stable Attribution

#354

Post by nephanth »

Uh what? That's not how it works, SD doesn't just get inspiration from a few similar images. It uses the weights trained from every image, every time. If you want attribution, you need to give it to the whole dataset, or not at all

Re: Stable Attribution

#355

Post by romwell »

l33tman wrote: post #296
romwell wrote: post #182

Earlier quoted context omitted.

>except if you can tell that the novel GPTx "paraphrased" is a derivative work from yours, rather than transformative, you can either sue for loyalties, or take action. Respectfully, no. The top comment in this thread is saurik asking Stable Diffisuion to generate "an avatar of saurik" and getting a veritable likeness of themselves back. It would be laughable to think that this is feasible without the model having la…

"If it does, then that data was indispensable for generating the output for that prompt." It gets murkier than this, by far. The diffusion models are not just trained on images, but on text. It varies by implementation but Stable Diffusion for example used a pre-trained CLIP transformer network from OpenAI (and subsequently OpenCLIP). CLIP can have internal associations between words that in turn steer the diffusion…

Thanks for pointing this out!

I guess the test I described is the simplest case of detecting appropriation by AI. Still, I think it would be a good start.

Re: Stable Attribution

#356

Post by pksebben »

dahart wrote: post #353
pksebben wrote: post #339

Earlier quoted context omitted.

But it's really not, though. It's remembering something related to the pixels, yeah, but that's like remembering the shape a line can take or the color of the sky. To extend your musical analogy, it's remembering that many songs are in 4/4 time, and that major chords sound appealing. Also, were you to compress anything, an mp3 or a picture, in a lossy fashion, to that degree of compression (~10^-5), you would no long…

In the case of an overfit image, which is the thing Stable Diffusion is being sued over, it is just compression, literally. The image data is stored in the network weights, and the image can be reconstructed. You’re drawing a distinction without a difference.

is this (1) the lawsuit you're referring to?

'cause those images are not the same. Sports events are just easy to fake, because they're boring - all sports pictures look roughly the same.

Edited to add: There's another lawsuit (a class action - 2), and after a little light reading, I came across section 5: 'Do diffusion models copy?', and my stomach jumped.

What they're doing, to make a point at trial that stable diffusion copies images, is _training images into the model, then using that trained model to prove that stable diffusion is a compression algorithm_.

This is a patent fabrication. If you train a model hard enough, yeah, it will produce the image you trained it on. And become useless for all other images. Congrats, you've just compressed your 7kb image to a 7gb diffusion model.

What scares me about this, is that the average court in the US is absolutely dumb enough to fall for it.

1 - https://www.theverge.com/2023/2/6/23587393/ai-art-copyright-...

2 - https://arxiv.org/pdf/2212.03860.pdf

Re: Stable Attribution

#357

Post by dahart »

pksebben wrote: post #356
dahart wrote: post #353

Earlier quoted context omitted.

In the case of an overfit image, which is the thing Stable Diffusion is being sued over, it is just compression, literally. The image data is stored in the network weights, and the image can be reconstructed. You’re drawing a distinction without a difference.

is this (1) the lawsuit you're referring to? 'cause those images are not the same. Sports events are just easy to fake, because they're boring - all sports pictures look roughly the same. Edited to add: There's another lawsuit (a class action - 2), and after a little light reading, I came across section 5: 'Do diffusion models copy?', and my stomach jumped. What they're doing, to make a point at trial that stable dif…

This is dismissive in the face of increasing evidence that a bunch of NN models have already been caught reproducing accidentally overfit data. Many examples have popped up with Stable Diffusion, not just one you disagree with. Same goes for ChatGPT, for GitHub Copilot, for Imagen, and a bunch of models.

Calling people dumb is to be willfully ignorant to the fact that neural networks actually can and really do remember images, not just when overfitting, but also when examples are in a low-density area of the latent space, when it doesn’t have enough neighbors to average with. The machine really is technically a machine intentionally and specifically built to reproduce a weighted combination of it’s inputs, and it really is possible for that weight vector to spike on some specific training examples. This won’t go away by pretending it doesn’t happen, it will go away when people curate training data that is legal to use, and/or when people write software that detects and rejects outputs that are too similar to a training sample, or otherwise guarantee no individual examples can be reconstructed. This is precisely why the project we’re commenting on is interesting, because it takes a step in that direction.

Re: Stable Attribution

#358

Post by wilimitis »

timmaxw wrote: post #337
wilimitis wrote: post #313

Earlier quoted context omitted.

> But the controversy over Stable Diffusion is whether they owe the human artists a cut or not, right? The controversy of Stable Diffusion is attribution. I could not care less about receiving a cut. Rather, I want my life's work in pushing particular style further than anyone has before to be recognized primarily so the next artist can trace the lineage and improve upon it. > Stable Diffusion is definitely not claim…

Maybe I'm missing something, but I don't see human artists doing the form of attribution you're describing. When people post artwork online, or when artwork is displayed in museums, it's normal to credit the artist who did the work, but not the artists who inspired them. (Unless the artist is directly copying another artist's work, or drawing fanart of a character from another work.) Can you clarify what this kind of…

I can understand how you would be missing that context without being a professional artist yourself. 1. Art historians carefully note the chain of inspiration, mentorship, etc in their works through primary sources (books, wikipedia, documentaries). 2. Modern living artists purchase the art of, study the art of, and share the art of their inspirations.

Re: Stable Attribution

#359

Post by wilimitis »

vintermann wrote: post #348
wilimitis wrote: post #311

Earlier quoted context omitted.

Why do you close the space for people who are pro-ai AND pro-attribution? The entire ai space wreaks of this divisiveness, and is likely why it will continue to die out as another "art-fad". There is seemingly little willingness to integrate into the existing art world in good faith.

I close the space for liars, whether they are pro or anti, and this service lies about attribution. Lying about attribution will just make everything worse.

I carefully did not mention this service, but rather pointed it to your sentiment which AGAIN seems to be binary in this pro/anti idea which is unnecessarily rigid and resistant to furthering discussion on the topic.

Re: Stable Attribution

#360

Post by pksebben »

dahart wrote: post #357
pksebben wrote: post #356

Earlier quoted context omitted.

is this (1) the lawsuit you're referring to? 'cause those images are not the same. Sports events are just easy to fake, because they're boring - all sports pictures look roughly the same. Edited to add: There's another lawsuit (a class action - 2), and after a little light reading, I came across section 5: 'Do diffusion models copy?', and my stomach jumped. What they're doing, to make a point at trial that stable dif…

This is dismissive in the face of increasing evidence that a bunch of NN models have already been caught reproducing accidentally overfit data. Many examples have popped up with Stable Diffusion, not just one you disagree with. Same goes for ChatGPT, for GitHub Copilot, for Imagen, and a bunch of models. Calling people dumb is to be willfully ignorant to the fact that neural networks actually can and really do rememb…

I agree with you that they have the capacity to remember an image - but they're not compressing them. That's a fundamentally different thing. The argument being made by that class action lawsuit is that "this thing can reproduce image X so it's a compression algorithm and nothing more", which they are predicating on an exercise that is sneaky and dishonest, and only likely to hold water with someone who has a limited understanding of the tech and isn't paying very close attention.

I think it does go without saying that our legal system has made some pretty dumb decisions regarding tech in the past - we read here all the time about the patent system, which is damn close in spirit to copyright.

Again, yes, they can remember an image, but they are not remembering pixels, and it's not compression. The vectors you're referring to are not a smaller version of the data, nor are they a pixel representation or even a close derivative thereof. Sure, there's a connection between the latent space and the pixels, but I don't see how that's the same thing.

For those following along, (1) is the best paper I could find talking about extracting images from SD. I'm open to more resources, and I'm even open to being convinced I'm wrong, but not by intentionally overtraining a model and calling it 'compression'. That's a lie.

To take a step back here, is it really the incidental occasional regurgitating of an existing image that's got everyone on edge, or is that just an easier target than "this is disruptive so I want to make it go away"? I'm not saying it doesn't suck that this is gonna put a ton of people out of jobs; both my parents were professional photographers in the 80s. I get it. But like, let's talk about that. Not some orthogonal strawman.

And hey, just to get it out there. We might disagree but I'm not calling you dumb. I do appreciate your willingness to engage an opposing view - it's part of what keeps me coming back to HN.

1 - https://arxiv.org/pdf/2301.13188.pdf

Post reply on HN