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

41–50 of 365 posts

Re: Stable Attribution

#41

Post by Gigachad »

goldemerald wrote: post #8

This is a great website, but not in the way the authors intended. Based on some of the examples they explicitly provided, it is clear to me Stable Diffusion creates novel art. Here's a random example https://www.stableattribution.com/?image=a2666aee-0a1a-411b-... I will admit this is a nice tool for verifying the creations of SD aren't pure copies, so I think it will be useful for a time. But as AI-generated images s…

Yeah this is probably the future, it seems like the vast majority of the time the output is very unique, but if there is far too much source material for a particular prompt, it copies. Say a prompt like “Mona Lisa”.

So now you can just use this tool to verify your output is safe.

Re: Stable Attribution

#42

Post by GaggiX »

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 uploading actual photos and art just to see what the site tell me with confidence "These human-made source images were used by AI to generate this image".

Edit: https://rom1504.github.io/clip-retrieval, this site has always been there to explore the LAION dataset using CLIP image/text embeddings and without the need to mislead the user.

Edit 2: As it's showed in this tweet: https://twitter.com/kurumuz/status/1622379532958285824, they are just using CLIP-L/14 and find the most semantically similar images.

Re: Stable Attribution

#44

Post by mattbee »

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 the training pictures it identifies, and regenerate using the same prompt to see whether the output is remotely similar.

Obviously that would be hilariously expensive and slow for a casual web service like this, but not beyond the realms of possibility for a wealthy copyright-holder.

e.g. if an prompt for an image includes "in the style of Kincade", and you could subtract all of Kinkade's copyrighted images from the training data, would the model still be able to produce anything like his work? If not, Thomas Kinkade might have a copyright case against people who publish AI art "in the style of Kincade", because he could show that his input was the major contributor to any lucrative output, even if nobody could pin down the cause & effect.

Re: Stable Attribution

#45

Post by fwlr »

Uploading works by real human artists gives you a batch of results that resemble a reference board (mood board, inspiration board, etc) the artist could have been looking at while creating their original work of art. Obviously it’s not the actual reference board, the only way to get that is to ask the artist yourself, but it sure looks like what you’d expect their reference board to look like.

This site is grift, of course, and I doubt its creators expect it to sway anyone who knows it’s just doing a nearest neighbor search of image embedding vectors. But it’s oddly humanizing to see that the AI uses reference boards too.

Re: Stable Attribution

#46

Post by mrtksn »

goldemerald wrote: post #8

This is a great website, but not in the way the authors intended. Based on some of the examples they explicitly provided, it is clear to me Stable Diffusion creates novel art. Here's a random example https://www.stableattribution.com/?image=a2666aee-0a1a-411b-... I will admit this is a nice tool for verifying the creations of SD aren't pure copies, so I think it will be useful for a time. But as AI-generated images s…

The discussion over novelty is useless, all these legal structures like copyrights patents royalties and licenses are about creating business structures to allow artist or inventors be compensated.

They were already imperfect in it and now suddenly a new technology drags their work into the wilderness with no stuctures for compensation.

The images created by the AI are indeed novel but they feed on the work of people who spent decades building this style. Of course artist themselves feed from each other but they usually don't interfere with the business. So let's say, if an artist developed a particular style and someone wants to hire them for a business project like a game they can't feasibly just learn that style and use it so they hire the artist. Later other people catch on this style and develop over it. It only works because monetising through copying the style is not very feasible.

Suddenly you have a machine that makes it feasible. Instead of hiring the artist or licensing their works, you train your machine on it and start generating any number of images of that style or combination with other style without paying the people who come up with all that.

How is the artist supposed to be compensated for spending years of developing that style/method?

I'm fine with getting rid of all that copyright and license stuff but let's not pretend that what's happening now is a fair endeavour.

Re: Stable Attribution

#47

Post by ertian »

version_five wrote: post #2

This appears to be just looking for the nearest neighbors of the image in embedding space and calling those the source data. This by definition would find similar looking images, but it's not strictly correct to call it attribution. To some extent all of the training data is responsible for the result - as an example, the model is also learning from negative examples. The result here may feel satisfying, but it's ove…

Yeah, it seems like it'll just as happily 'attribute' human-made images as well, which calls the whole thing into question. If it's really showing the images that stable diffusion has 'stolen from', and it'll do the same for humans, does that not mean people are equally guilty?

Re: Stable Attribution

#48

Post by gfodor »

Ah, a nice visual proof that these AI systems are actually synthesizing images with a degree of inspiration from prior art similar to the way humans do.

Re: Stable Attribution

#49

Post by f38zf5vdt »

This is a company that allows you to search for images from a training dataset that have a high cosine similarity with a given image. It appears to be the same as the open source software published by LAION.

https://github.com/rom1504/clip-retrieval

It does not appear to actually show you how images were used to train a generative AI.

Re: Stable Attribution

#50

Post by mikeg8 »

mrtksn wrote: post #46
goldemerald wrote: post #8

This is a great website, but not in the way the authors intended. Based on some of the examples they explicitly provided, it is clear to me Stable Diffusion creates novel art. Here's a random example https://www.stableattribution.com/?image=a2666aee-0a1a-411b-... I will admit this is a nice tool for verifying the creations of SD aren't pure copies, so I think it will be useful for a time. But as AI-generated images s…

The discussion over novelty is useless, all these legal structures like copyrights patents royalties and licenses are about creating business structures to allow artist or inventors be compensated. They were already imperfect in it and now suddenly a new technology drags their work into the wilderness with no stuctures for compensation. The images created by the AI are indeed novel but they feed on the work of people…

I never thought of it from that perspective before, thanks for sparking that thought experiment.
Post reply on HN