4 ms·

> It seems like one of Stable Diffusion’s most popular use cases is give me <situation> in the style of <famous_artist>. You know what's the most popular <styl

> It seems like one of Stable Diffusion’s most popular use cases is give me <situation> in the style of <famous_artist>. You know what's the most popular <style> for Stable Diffusion? Photorealism. And the second most popular <style>? Anime. At the risk of stating the obvious, I will note that no artist can hold the honor of having invented generic "photography" and generic "anime". But sure, I will give you 2 internet points for accurately spotting one of the extremely rare use cases of Stable Diffusion where an artists' style is copied. I agree with you that is not all kosher. > I mean the whole point of using AI at all is to get something to give you high quality stuff for free, where the high quality has been lifted from a high quality training set. No, it's not. It's a failure mode. It's called "overfitting". We went over this. The vast majority of Stable Diffusion's usage is about producing something novel, and reproducing something that looks like a training image is considered a failure.

dahart 3y ago
> no artist can hold the honor of having invented generic “photography” That’s entirely beside the point. Copyright still applies to photographs, and training has been occurring on copyright-protected photographs. > It’s a failure mode It seems like maybe you didn’t fully parse what I said there, which wasn’t about reproducing works closely enough to violate copyright, it was about the broad goal of AI, which absolutely is to train on fixations (in Copyright parlance) and be able to produce mashup results that mimic the inputs as closely as possible without breaking the law. Because the goals here and the design intent here are to skirt the legal lines, it’s not at all surprising that sometimes we will cross it, whether it counts specifically as overfitting or not. When the broad design goal combined with users and prompts seeking specific outcomes is riding the line between mimicking style and clear copying, we’re basically priming the overall system to live in a legal gray area. This is outside the question of whether consuming copyrighted training material is legal, regardless of whether any reproductions occur. Watching copyrighted movies is not legal, I don’t have to share them with anyone else. Why should AI training that consumes copyrighted material even be legal in the first place? (Maybe it’s not.)
- baobabKoodaa 3y ago
  > That’s entirely beside the point. Copyright still applies to photographs, and training has been occurring on copyright-protected photographs. No, you missed the point. I was making a distinction between artist styles, like "in the style of Picasso", versus generic styles, like "photography". Furthermore, I was making the point that the generic styles, such as "photography", are way more popular in the Stable Diffusion community, compared to artist specific styles, like "in the style of Picasso". Just to spell it out as clearly as possible: the majority of Stable Diffusion usage is not trying to imitate the style of a specific artist. Whether or not training on copyright-protected photographs is legal is tangential to the point. > It seems like maybe you didn’t fully parse what I said there, which wasn’t about reproducing works closely enough to violate copyright, it was about the broad goal of AI, which absolutely is to train on fixations (in Copyright parlance) and be able to produce mashup results that mimic the inputs as closely as possible without breaking the law. I fully understood your point and expressed my disagreement of it. Specifically, what you described is typically not the intended outcome, it is a failure mode. I have repeated this many times now and you keep ignoring me, as if the problem goes away by ignoring it.
  - dahart 3y ago
    The very explicit intent of several high profile AI efforts to date has fundamentally been to distribute remixes of copyrighted works while skirting copyright law. This is precisely why OpenAI was requesting copyright exemption. They know how to fix overfitting. Their request is acknowledging that overfitting is not the core problem, and that they want the leeway to change or break existing copyright laws in order to consume the entire internet and be able to distribute the content that it contains. As far as copyright is concerned, there is no such “generic” category of photography, you’re trying to invent a term that has no legal status. All photos fall under copyright. Training on them may break copyright, reproducing them accidentally may break copyright. It is irrelevant what the majority use case is. (But how can you possibly claim to know what the “majority” of Stable Diffusion’s usage is?) It is a fact that examples of people asking to mimic specific artists are all over the internet, it’s quite common and easy to find. It doesn’t matter if this usage is not the majority, all that matters in terms of the legal question is whether the usage of AI meets the criteria established in the law. You’re splitting a hair trying to contradict me when I’m trying to point out that regardless of whether people are trying to imitate a single artist, the design and implementation of the system is expressly for the purpose of imitating multiple artists. My point doesn’t hinge on single vs 2, 3, or many. This isn’t a failure mode, this is the stated purpose of many of today’s neural networks. They are built to copy snippets of other people’s work.

dahart 3y ago

> no artist can hold the honor of having invented generic “photography” That’s entirely beside the point. Copyright still applies to photographs, and training has been occurring on copyright-protected photographs. > It’s a failure mode It seems like maybe you didn’t fully parse what I said there, which wasn’t about reproducing works closely enough to violate copyright, it was about the broad goal of AI, which absolutely is to train on fixations (in Copyright parlance) and be able to produce mashup results that mimic the inputs as closely as possible without breaking the law. Because the goals here and the design intent here are to skirt the legal lines, it’s not at all surprising that sometimes we will cross it, whether it counts specifically as overfitting or not. When the broad design goal combined with users and prompts seeking specific outcomes is riding the line between mimicking style and clear copying, we’re basically priming the overall system to live in a legal gray area. This is outside the question of whether consuming copyrighted training material is legal, regardless of whether any reproductions occur. Watching copyrighted movies is not legal, I don’t have to share them with anyone else. Why should AI training that consumes copyrighted material even be legal in the first place? (Maybe it’s not.)

baobabKoodaa 3y ago
> That’s entirely beside the point. Copyright still applies to photographs, and training has been occurring on copyright-protected photographs. No, you missed the point. I was making a distinction between artist styles, like "in the style of Picasso", versus generic styles, like "photography". Furthermore, I was making the point that the generic styles, such as "photography", are way more popular in the Stable Diffusion community, compared to artist specific styles, like "in the style of Picasso". Just to spell it out as clearly as possible: the majority of Stable Diffusion usage is not trying to imitate the style of a specific artist. Whether or not training on copyright-protected photographs is legal is tangential to the point. > It seems like maybe you didn’t fully parse what I said there, which wasn’t about reproducing works closely enough to violate copyright, it was about the broad goal of AI, which absolutely is to train on fixations (in Copyright parlance) and be able to produce mashup results that mimic the inputs as closely as possible without breaking the law. I fully understood your point and expressed my disagreement of it. Specifically, what you described is typically not the intended outcome, it is a failure mode. I have repeated this many times now and you keep ignoring me, as if the problem goes away by ignoring it.
- dahart 3y ago
  The very explicit intent of several high profile AI efforts to date has fundamentally been to distribute remixes of copyrighted works while skirting copyright law. This is precisely why OpenAI was requesting copyright exemption. They know how to fix overfitting. Their request is acknowledging that overfitting is not the core problem, and that they want the leeway to change or break existing copyright laws in order to consume the entire internet and be able to distribute the content that it contains. As far as copyright is concerned, there is no such “generic” category of photography, you’re trying to invent a term that has no legal status. All photos fall under copyright. Training on them may break copyright, reproducing them accidentally may break copyright. It is irrelevant what the majority use case is. (But how can you possibly claim to know what the “majority” of Stable Diffusion’s usage is?) It is a fact that examples of people asking to mimic specific artists are all over the internet, it’s quite common and easy to find. It doesn’t matter if this usage is not the majority, all that matters in terms of the legal question is whether the usage of AI meets the criteria established in the law. You’re splitting a hair trying to contradict me when I’m trying to point out that regardless of whether people are trying to imitate a single artist, the design and implementation of the system is expressly for the purpose of imitating multiple artists. My point doesn’t hinge on single vs 2, 3, or many. This isn’t a failure mode, this is the stated purpose of many of today’s neural networks. They are built to copy snippets of other people’s work.