4 ms·

> Stability AI has already announced that it is removing users’ ability to request images in a particular artist’s style and further, that future releases of St

> Stability AI has already announced that it is removing users’ ability to request images in a particular artist’s style and further, that future releases of Stable Diffusion will comply with any artist’s requests to remove their images from the training dataset. With that removal, the most outrage-inducing and troublesome output examples disappear from this case, leaving a much more complex and muddled set of facts for the jury to wade through. How can this possibly be a valid good faith argument? Either they're in breach of authors' copyright which extends to every piece of art that they included in the dataset without permission, or they're in the clear and aren't obligated to respond to removal requests. This reads like damage control to me in an effort to temporarily silence the loudest critics.

devwastaken 4y ago
They don't have to remove images from the training set, they're saying they're opting to do so, and using that as an argument as to why if there supposedly could be copyright infringement, they're not liable, because they allow it to be removed. They could just as well not do anything and continue on - it's likely this case will be in defendents favor. Same as how Google can crawl the net, cache data, transform it, etc.
hehdhdhkf 4y ago
Where is the form to remove my reddit comments from chat gpt training data? Or my blog posts from gpt training data? I have a paragraph on the Internet that someone read and got an idea - I want my royalties. These artists complaints are ridiculous, and are being made by people who don’t understand how things work. If some other person draws a picture in their “style”, no one has to ask permission. That’s not a thing. They either don’t understand how it works or they are just upset that a computer can make art as good as (or better than) they can in a fraction of the time. All knowledge workers and creatives are going to face this in the future. It’s going to suck, but it would be great if we all could try to understand reality first.
- yencabulator 4y ago
  > If some other person draws a picture in their “style”, no one has to ask permission. That’s not a thing. Try making a comic book with a character that looks like Mickey Mouse and see how well that goes.
- dns_snek 4y ago
  Do you want to live in a future where artists don't make original art, musicians don't make music, book writers don't write, and so on, all because AI companies can replicate 1000 different copies in their style or merely remix it for marginally $0 cost, washed of all original copyright? > All knowledge workers and creatives are going to face this in the future. It’s going to suck This is not a given. It's up to us and the copyright law. Real original work should be compensated appropriately unless you're proposing that we accelerate deployment of universal basic income and completely abolish copyright law. I have a feeling you might not like the violent outcome if you effectively strip original creators of their copyright, give corporations the right to effectively generate infinite profit off the backs of their work and tell the creators (and other people whose jobs will be automated away) to pound sand when they ask how they're supposed to pay rent from now on.
  - pixl97 4y ago
    Do you want to live in a future where anything 'original' an artist creates has now blocked anyone else on the planet for the next 99 years and you must pay them royalties? Because we already have Disney now and they suck quite a bit. I honestly want to live in a world where this 'worrying about paying for rent' is not a problem that we're concerned with, and a world with AI that can create and make we far more apt to achieve that than with the status quo we've been following so far.
    - dns_snek 4y ago
      > Do you want to live in a future where anything 'original' an artist creates has now blocked anyone else on the planet for the next 99 years and you must pay them royalties? No I don't want that but there's a very clear distinction between people who are truly inspired by each others' work and produce an average human output that maybe covers their bills and AI that can pump out billions of replicates per day to drown out all original work, with entirety of that value captured by some corporation. > I honestly want to live in a world where this 'worrying about paying for rent' is not a problem that we're concerned with, and a world with AI that can create and make we far more apt to achieve that than with the status quo we've been following so far. Hey, me too. But those human issues should be addressed first, not after automation is allowed to wipe out people's livelihoods. I'm not anti-AI, I'm anti-big tech that seeks to exploit billions of people's original work for their own benefit.
epistemer 4y ago
It seems highly foolish from a marketing perspective for the artist too. I don't see how more people copying the artist style would not increase the value of originals. A smart artist here should promote that their style is staying in the dataset. It is as good free publicity as they will ever get.
- jonhohle 4y ago
  It depends on whether or not the “style” is associated with the original artist or if the derivatives usurp the original. e.g. Ask 10 random people who wrote the song “Hurt” and 9 out of 10 will probably say Johnny Cash.
- sct202 4y ago
  It's not foolish at all. There's no value prop for an artist to advertise their work in the model as 1 of 400+ million. SD doesn't tell you anything about the art that inspired the output, no one will ever know the artists work was ever used, so this 'exposure' is as good as $0.
  - rnjesus 4y ago
    this isn’t quite true. i was playing around with sd, and whilst browsing styles on a site where different artists’ names were compared using the same prompt (“[artist name] painting of flowers”, for example), a particular artist’s style stood out to me. i ended up buying a print from them. sd might not directly include this info in its outputs, but it really is free marketing for some (in fact, i suspect a lot of artists are probably flattered for being included by name in the dataset, but some are riding the waxing wave of outrage for the free media exposure)
counttheforks 4y ago
The models are already released. They can't retroactively censor released models.
benlivengood 4y ago
The LAION-5B dataset is metadata and URI pairs; all the images are publicly accessible on the Internet. Stable Diffusion's U-Net is trained to remove noise from images in latent space, which the variational autoencoder (VAE) converts to and from pixel space. CLIP embeddings are used to improve the denoising step of the U-Net by using the correlations between human language descriptions of the pixel image to reduce latent noise. Neither the U-Net nor the VAR are trained to interpolate or reproduce images from the training set; if that happened the model would be overfitted and loss would be terrible on the validation set. The VAE is trained to produce a latent space that can accurately encode and decode any pixel image, and the U-Net is trained to remove gaussian noise from the latent space. Stable Diffusion v2 16-bit is ~3GB of data. It was trained on hundreds of millions of images (minimum of 170M in the 512x512 step alone). That leaves a maximum of ~20 bytes per image that could conceivably be a copy, which is certainly not enough to directly reproduce either the style or contents of any individual image. There is no artwork included in Stable Diffusion. There is a semantic representation of how images are composed of varied subjects represented in the latent space and what pixel probabilities over those subjects relate to human language phrases during decoding, and finally a method to remove noise from the semantic representation, e.g. starting with a blank or random canvas and interpreting what may be there, iteratively guided by CLIP embeddings. If you give Stable Diffusion an empty CLIP embedding you get a random human-interpretable image obeying the distribution of the learned latent space.
- wokwokwok 4y ago
  It’s deeply disingenuous to say the U-Net is not trained on the images because it’s trained on the latent representations. Latents are a compressed representation of the source images that are fully recoverable. If you train a model on a compressed jpg of an image, or on any deterministic transformation of it, you’re still training it on that image. Any suggestion otherwise is only because someone is trying to put some spin on things. > Stable Diffusion v2 16-bit is ~3GB of data. It was trained on hundreds of millions of images… And yet! Remarkably! It can generate pictures of the Mona Lisa! Here’s a question for you: if you encode the process of drawing an exact copy of an image, does the pure code that implements that mean you have a copy of the image in it? Have you encoded pixels as code? Does that mean there’s no copy of the image? How about a zip file full of images? It’s just a high entropy binary blob right? Yet… remarkably!!! It can be transformed into images by applying an algorithm. I don’t know the answer, but this handwavy “it couldn’t possibly encode them it’s too small” is… Pure. Nonsense. Of course some part of some images is embedded in the model in some form. Stop trivialising the issue. The issue here is: Does an algorithm that generates content infringe copyright? Does a black box that takes the input “a picture of xxx” and a seed and outputs a copyrighted image infringe? You know that’s possible. Don’t dodge. Technical details about oh “it couldn’t possibly have…” are pure rubbish. Sure it could. It could have a full resolution copy of a photo of the image in that black box. Of all the training data? Probably not. But of some of it? In compressed latent form? Most definitely.
williamcotton 4y ago
> This reads like damage control to me in an effort to temporarily silence the loudest critics. I think it is to avoid any common law wrongs related to the publicity rights of the defendants. It seems like something that a legal team would flag as an unnecessary risk for the product. Removing their names and images from the training data doesn't impact the usefulness of the model while at the same time creating a much smaller surface area for collecting subpoenas.
pessimizer 4y ago
Sounds like an opt-out dark pattern. US law is unbelievably aggressive when it comes to issues of copyright, and makes copyright itself opt-out, i.e. everything you produce is copyrighted, and you have to license it in order to remove that automatic copyright. But when it comes to building these models to reproduce imitations of other people's work, suddenly copyright gets loosey-goosey. Notice that it's a legal posture that implicitly condemns Copilot, which ignores explicitly formulated opt-outs in the form of licensing.
- jacquesm 4y ago
  This is not just the US, this also applies to many, many other countries.

devwastaken 4y ago

They don't have to remove images from the training set, they're saying they're opting to do so, and using that as an argument as to why if there supposedly could be copyright infringement, they're not liable, because they allow it to be removed. They could just as well not do anything and continue on - it's likely this case will be in defendents favor. Same as how Google can crawl the net, cache data, transform it, etc.

hehdhdhkf 4y ago

Where is the form to remove my reddit comments from chat gpt training data? Or my blog posts from gpt training data? I have a paragraph on the Internet that someone read and got an idea - I want my royalties. These artists complaints are ridiculous, and are being made by people who don’t understand how things work. If some other person draws a picture in their “style”, no one has to ask permission. That’s not a thing. They either don’t understand how it works or they are just upset that a computer can make art as good as (or better than) they can in a fraction of the time. All knowledge workers and creatives are going to face this in the future. It’s going to suck, but it would be great if we all could try to understand reality first.

yencabulator 4y ago
> If some other person draws a picture in their “style”, no one has to ask permission. That’s not a thing. Try making a comic book with a character that looks like Mickey Mouse and see how well that goes.
dns_snek 4y ago
Do you want to live in a future where artists don't make original art, musicians don't make music, book writers don't write, and so on, all because AI companies can replicate 1000 different copies in their style or merely remix it for marginally $0 cost, washed of all original copyright? > All knowledge workers and creatives are going to face this in the future. It’s going to suck This is not a given. It's up to us and the copyright law. Real original work should be compensated appropriately unless you're proposing that we accelerate deployment of universal basic income and completely abolish copyright law. I have a feeling you might not like the violent outcome if you effectively strip original creators of their copyright, give corporations the right to effectively generate infinite profit off the backs of their work and tell the creators (and other people whose jobs will be automated away) to pound sand when they ask how they're supposed to pay rent from now on.
- pixl97 4y ago
  Do you want to live in a future where anything 'original' an artist creates has now blocked anyone else on the planet for the next 99 years and you must pay them royalties? Because we already have Disney now and they suck quite a bit. I honestly want to live in a world where this 'worrying about paying for rent' is not a problem that we're concerned with, and a world with AI that can create and make we far more apt to achieve that than with the status quo we've been following so far.
  - dns_snek 4y ago
    > Do you want to live in a future where anything 'original' an artist creates has now blocked anyone else on the planet for the next 99 years and you must pay them royalties? No I don't want that but there's a very clear distinction between people who are truly inspired by each others' work and produce an average human output that maybe covers their bills and AI that can pump out billions of replicates per day to drown out all original work, with entirety of that value captured by some corporation. > I honestly want to live in a world where this 'worrying about paying for rent' is not a problem that we're concerned with, and a world with AI that can create and make we far more apt to achieve that than with the status quo we've been following so far. Hey, me too. But those human issues should be addressed first, not after automation is allowed to wipe out people's livelihoods. I'm not anti-AI, I'm anti-big tech that seeks to exploit billions of people's original work for their own benefit.

epistemer 4y ago

It seems highly foolish from a marketing perspective for the artist too. I don't see how more people copying the artist style would not increase the value of originals. A smart artist here should promote that their style is staying in the dataset. It is as good free publicity as they will ever get.

jonhohle 4y ago
It depends on whether or not the “style” is associated with the original artist or if the derivatives usurp the original. e.g. Ask 10 random people who wrote the song “Hurt” and 9 out of 10 will probably say Johnny Cash.
sct202 4y ago
It's not foolish at all. There's no value prop for an artist to advertise their work in the model as 1 of 400+ million. SD doesn't tell you anything about the art that inspired the output, no one will ever know the artists work was ever used, so this 'exposure' is as good as $0.
- rnjesus 4y ago
  this isn’t quite true. i was playing around with sd, and whilst browsing styles on a site where different artists’ names were compared using the same prompt (“[artist name] painting of flowers”, for example), a particular artist’s style stood out to me. i ended up buying a print from them. sd might not directly include this info in its outputs, but it really is free marketing for some (in fact, i suspect a lot of artists are probably flattered for being included by name in the dataset, but some are riding the waxing wave of outrage for the free media exposure)

counttheforks 4y ago

The models are already released. They can't retroactively censor released models.

benlivengood 4y ago

The LAION-5B dataset is metadata and URI pairs; all the images are publicly accessible on the Internet. Stable Diffusion's U-Net is trained to remove noise from images in latent space, which the variational autoencoder (VAE) converts to and from pixel space. CLIP embeddings are used to improve the denoising step of the U-Net by using the correlations between human language descriptions of the pixel image to reduce latent noise. Neither the U-Net nor the VAR are trained to interpolate or reproduce images from the training set; if that happened the model would be overfitted and loss would be terrible on the validation set. The VAE is trained to produce a latent space that can accurately encode and decode any pixel image, and the U-Net is trained to remove gaussian noise from the latent space. Stable Diffusion v2 16-bit is ~3GB of data. It was trained on hundreds of millions of images (minimum of 170M in the 512x512 step alone). That leaves a maximum of ~20 bytes per image that could conceivably be a copy, which is certainly not enough to directly reproduce either the style or contents of any individual image. There is no artwork included in Stable Diffusion. There is a semantic representation of how images are composed of varied subjects represented in the latent space and what pixel probabilities over those subjects relate to human language phrases during decoding, and finally a method to remove noise from the semantic representation, e.g. starting with a blank or random canvas and interpreting what may be there, iteratively guided by CLIP embeddings. If you give Stable Diffusion an empty CLIP embedding you get a random human-interpretable image obeying the distribution of the learned latent space.

wokwokwok 4y ago
It’s deeply disingenuous to say the U-Net is not trained on the images because it’s trained on the latent representations. Latents are a compressed representation of the source images that are fully recoverable. If you train a model on a compressed jpg of an image, or on any deterministic transformation of it, you’re still training it on that image. Any suggestion otherwise is only because someone is trying to put some spin on things. > Stable Diffusion v2 16-bit is ~3GB of data. It was trained on hundreds of millions of images… And yet! Remarkably! It can generate pictures of the Mona Lisa! Here’s a question for you: if you encode the process of drawing an exact copy of an image, does the pure code that implements that mean you have a copy of the image in it? Have you encoded pixels as code? Does that mean there’s no copy of the image? How about a zip file full of images? It’s just a high entropy binary blob right? Yet… remarkably!!! It can be transformed into images by applying an algorithm. I don’t know the answer, but this handwavy “it couldn’t possibly encode them it’s too small” is… Pure. Nonsense. Of course some part of some images is embedded in the model in some form. Stop trivialising the issue. The issue here is: Does an algorithm that generates content infringe copyright? Does a black box that takes the input “a picture of xxx” and a seed and outputs a copyrighted image infringe? You know that’s possible. Don’t dodge. Technical details about oh “it couldn’t possibly have…” are pure rubbish. Sure it could. It could have a full resolution copy of a photo of the image in that black box. Of all the training data? Probably not. But of some of it? In compressed latent form? Most definitely.

williamcotton 4y ago

> This reads like damage control to me in an effort to temporarily silence the loudest critics. I think it is to avoid any common law wrongs related to the publicity rights of the defendants. It seems like something that a legal team would flag as an unnecessary risk for the product. Removing their names and images from the training data doesn't impact the usefulness of the model while at the same time creating a much smaller surface area for collecting subpoenas.

pessimizer 4y ago

Sounds like an opt-out dark pattern. US law is unbelievably aggressive when it comes to issues of copyright, and makes copyright itself opt-out, i.e. everything you produce is copyrighted, and you have to license it in order to remove that automatic copyright. But when it comes to building these models to reproduce imitations of other people's work, suddenly copyright gets loosey-goosey. Notice that it's a legal posture that implicitly condemns Copilot, which ignores explicitly formulated opt-outs in the form of licensing.

jacquesm 4y ago
This is not just the US, this also applies to many, many other countries.