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

321–330 of 365 posts

Re: Stable Attribution

#321

Post by dahart »

pksebben wrote: post #271
dahart wrote: post #133

Earlier quoted context omitted.

Available for free online is not a valid justification for copying under copyright law, though, right? You can’t distribute something just because you can see it. True for museums and magazines as it is for online content. > They don’t have the right to not allow people to learn from it, that’s just never been a part of copyright. Yeah this is true. Stable Diffusion and other neural networks are not “learning” from i…

it's not remembering pixels. for it to do that, it would have to have the pixels stored somewhere. It does not. The laion 5b dataset is in the neighborhood of 220TB. (1) That is how much storage space you need to remember the pixels. The stable diffusion 1.5 checkpoint is 7gb. (2) 1 https://github.com/rom1504/img2dataset/blob/main/dataset_exa... 2 https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/m...

It is effectively remembering pixels, and we can prove it because it can regenerate some of the training images verbatim, close enough to violate copyright law. It doesn’t matter that it’s compressed.

Re: Stable Attribution

#322

Post by Ajedi32 »

wpietri wrote: post #151
josephg wrote: post #65

Earlier quoted context omitted.

The core of your concern (& argument) seems to be the problem of existing business models becoming disrupted by this technology. But the point of good law isn’t to protect an established business model. If that were the case, we would have outlawed the loom because it displaced weavers, and the camera (sorry portrait painters) and the iPhone. (How many telegraph operators are left? None!) When an artist learns to dra…

> But the point of good law isn’t to protect an established business model. Really? The Constitution specifically includes a bit about: "to promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries". That was quickly followed by the first copyright act. As I understand it, the copyright portion of this was e…

The stated goal is "to promote the progress of science and useful arts". Protecting an established business model via copyright was merely a means to that end, not the end in itself.

In this case, diffusion models are actually a perfect example of "the progress of science and useful arts". The law should be structured in such a way as to promote such progress, not hinder it.

Re: Stable Attribution

#323

Post by MacsHeadroom »

bioemerl wrote: post #99
hansihe wrote: post #70

Earlier quoted context omitted.

We have no idea if images by Stable Diffusion violates copyright, there is no legal precedent.

Technically we do have examples of stable diffusion violating copyright, it will generate some exact clone images if you give it the right prompt and that image exists a few hundred times in the training data.

For something like 11 images which were accidentally repeated hundreds of times in the training data this is true. They're more the exception that proves the rule.

Re: Stable Attribution

#324

Post by convexfunction »

beepbooptheory wrote: post #309
convexfunction wrote: post #242

Earlier quoted context omitted.

Yes, whenever I spend hours refining a prompt in my stable diffusion UI and select one of hundreds or thousands of outputs and spend even more hours refining it with inpainting and other non-AI tools, I've felt nothing whatsoever, and my choices aren't informed in any way by anything interesting going on in my inner life.

If you struggle with it, if you are always wrestling with something not quite right, if it makes you nervous to share it, you are on the right track for sure! I'm just saying, a guy like Van Gogh comes around just a few times a century, he is not somebody you can simply encode into a lot of different matrices, it takes a certain human life to make art like him, or even more simply, to simply see the world like him. I…

> That is my maybe my main issue with the AI art bots, do the artists who work with them even feel vulnerable about their work?

Sometimes, yes. John Q Nerd playing around with it might not, but the "good AI artists" I follow on Twitter certainly do.

> Can artistic sincerity and enthusiasm exist in AI art?

Unequivocally yes.

> Do AI artists feel the literally metaphysical stakes of what they are doing like Van Gogh did?

Some of them seem to, though I dunno precisely what Van Gogh felt so I guess I'm not sure.

> Is it all just different campaign posters for the cause of legitimizing itself as "real" art?

Not sure where you find this stuff, but it sounds really boring. Certainly not all of it is.

Your unsolicited advice falls flat (to me), sorry. My unsolicited advice to you would be to psychologize less.

Re: Stable Attribution

#325

Post by scarmig »

throwaway1851 wrote: post #297
XorNot wrote: post #253

Earlier quoted context omitted.

Yep straight up just uploaded a picture from my camera phone and it confidently found a bunch of strikingly similar images.

That in itself isn’t an indictment of the website. The tool is built on the assumption of being provided an image generated from Stable Diffusion. If you violate that assumption, it isn’t surprising that the tool fails your “test.” By comparison, a “dog vs. cat” classifier that has 100% accuracy on the dog/cat task will, nonetheless, tell you that a slice of pizza is a dog... or a cat. (You could possibly interpret t…

It's an image similarity search engine, slapping on the tag SD attribution for marketing purposes.

Note that just because an image is similar (to human eyes) doesn't mean that it played a more significant role than a seemingly more dissimilar image. It could even return similar images that SD wasn't trained on at all. Even conditioned on providing an SD-generated image, it fails.

(Something doing what it claims to do, as opposed to naive image similarity, would actually be pretty cool and useful.)

Re: Stable Attribution

#326

Post by MacsHeadroom »

dzink wrote: post #212

From the beginning of using Stable Diffusion in local and cloud instances, I’ve been promoting SD to generate objects I know nobody has ever drawn before. “Airplane by Tesla”, “Taylor Swift flying in the clouds”, “Little girl riding on an ira descent unicorn and chasing butterflies in the clouds”, “Turkey as a Judge” etc. I highly encourage everyone to try doing that. The results are absolutely atrocious in the begin…

> The results are absolutely atrocious in the beginning and it takes many many runs short and long, with seeds guiding the model to get closer and closer to what I ask. It took a long time to get one instance of SD to make the invention look plausible, and then trying on a new model copy/instance takes the results back to crap.

This is simply not how any of this works and is only your imagination. Stable Diffusion is deterministic and has no instance memory. The seeds are random and length of session or starting a new instance has no effect on the randomness of seeds. Every seed is as random as the last, regardless of how long an instance has been running.

Re: Stable Attribution

#327

Post by GaggiX »

ausbah wrote: post #320
GaggiX wrote: post #42

Calling the nearest neighbors of the CLIP embeddings of an image "attribution" feels really misleading, the model has been influenced by the entire dataset it was trained on, just by finding the most semantically similar images does not mean the AI is just using that speficific group of images as references, they probably have almost no influence compared to the entire size of the dataset. P.S. I'm having fun uploadi…

I don't think this is right. it's possible to fine tune these models with a few pieces from a select artist such that when you say "[art] is the style of [new artist]" you will get new pieces in the style of that artist. those few select pieces have clearly had disproportionate influence on the generated images, even if just via conditioning in the prompt

Okay but here we're are talking about the standard SD 1/2 model trained on the LAION dataset, the site only do CLIP retrieval on the LAION dataset, the only thing that makes this website different from Google/Yandex/etc image reverse search.

Re: Stable Attribution

#328

Post by temp512345 »

pksebben wrote: post #271
dahart wrote: post #133

Earlier quoted context omitted.

Available for free online is not a valid justification for copying under copyright law, though, right? You can’t distribute something just because you can see it. True for museums and magazines as it is for online content. > They don’t have the right to not allow people to learn from it, that’s just never been a part of copyright. Yeah this is true. Stable Diffusion and other neural networks are not “learning” from i…

it's not remembering pixels. for it to do that, it would have to have the pixels stored somewhere. It does not. The laion 5b dataset is in the neighborhood of 220TB. (1) That is how much storage space you need to remember the pixels. The stable diffusion 1.5 checkpoint is 7gb. (2) 1 https://github.com/rom1504/img2dataset/blob/main/dataset_exa... 2 https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/m...

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.

Re: Stable Attribution

#329

Post by beepbooptheory »

convexfunction wrote: post #324
beepbooptheory wrote: post #309

Earlier quoted context omitted.

If you struggle with it, if you are always wrestling with something not quite right, if it makes you nervous to share it, you are on the right track for sure! I'm just saying, a guy like Van Gogh comes around just a few times a century, he is not somebody you can simply encode into a lot of different matrices, it takes a certain human life to make art like him, or even more simply, to simply see the world like him. I…

> That is my maybe my main issue with the AI art bots, do the artists who work with them even feel vulnerable about their work? Sometimes, yes. John Q Nerd playing around with it might not, but the "good AI artists" I follow on Twitter certainly do. > Can artistic sincerity and enthusiasm exist in AI art? Unequivocally yes. > Do AI artists feel the literally metaphysical stakes of what they are doing like Van Gogh di…

Sorry as well! Thanks for answering all my parenthetical rhetorical questions, definitely shows your not too defensive or anything about this stuff we are all, honestly, still navigating.

But I really really do wish you the best with this stuff, the future is bright and interesting either way for art!

Re: Stable Attribution

#330

Post by convexfunction »

mattbee wrote: post #44

I like this because they are trying to show how AI is a copyright laundry. I can see other commenters picking apart its method of heuristically guessing at source images from training data. That obviously won't be accurate, or a full picture, but I wonder if it would convince a judge. An interesting challenge for these heuristics would be to take the picture under test along with its prompt, retrain the model without…

It's not quite the same, and might not be possible to make rigorous enough that it really proves anything, but something sort of similar would actually be practical to at least attempt in many cases. Stable Diffusion checkpoints of the same major version, along with other families of model weights, have the IMO fascinating property that you can do element-wise arithmetic with them, and the resulting model will actually sort of function like you'd naively expect. Recent paper on the topic (in LLMs, not diffusion models) here: https://arxiv.org/abs/2212.04089

So, if you take a Stable Diffusion checkpoint (call it "A") which is only lightly trained on some subset of an artist's work, then fine tune it on the full corpus of that artist's work to a point where it's still coherent/"good" and just shy of actually memorizing the fine tuning data (call the resulting model "B"), then define model "C" as 2A-B (i.e. A + (A-B), where A-B is the artist's task vector multiplied by -1), can you still produce qualitatively similar images with model C? Whether with the exact same prompt, or the same prompt with "in the style of Kinkade" removed (which doesn't mean as much if Kinkade's task vector was subtracted), or with any prompt whatsoever?

Lots of issues with this as laid out -- it's definitely not quite the same as "forgetting" Kinkade from the training data, and "any prompt whatsoever" introduces tons of leeway, and most good AI-assisted art is not just an unmodified single text-to-image output anyway -- but it might be a promising direction to explore.

(Strongly disagree with the "copyright laundry" characterization, by the way.)

Post reply on HN