5 ms·

" In an interview with Forbes, Midjourney CEO David Holz said that the dataset that the company trained its AI model on was “just a big scrape of the internet”

" In an interview with Forbes, Midjourney CEO David Holz said that the dataset that the company trained its AI model on was “just a big scrape of the internet” and the company hadn’t sought consent from artists who owned the copyright to their work. “There isn’t really a way to get a hundred million images and know where they’re coming from,” Holz said." So you just go ahead and steal copyrighted images without ask and attribution? This attitude is far beyond the pale and I, for one, hope that the artists succeed with their lawsuit.

jaggs 4y ago
I think it might help if you understand how AI art works. They're not scraping images, they're scraping rules. Completely different thing. https://i.redd.it/2f00l6vsso6a1.jpg https://i.redd.it/2f00l6vsso6a1.jpg
- CaptainZapp 4y ago
  Does that really matter if you can easily generate something that comes pretty close (at least in a superficial way) to the artist by just entering the artists name with your description?
  - jaggs 4y ago
    That's a really good question, and it's key to the issue at hand I think. Any competent artist can create a lookalike style of another artist at will. It's just a matter of looking at the paint type, brushes, materials, strokes, composition, etc and copying that. In fact, here's Vincent Van Gogh copying other artists of his time - https://www.dailyartmagazine.com/van-gogh-copy/ https://www.dailyartmagazine.com/van-gogh-copy/ . How does this differ from a machine doing the output? Apart from speed and flexibility?
    - diemes1 4y ago
      Van Gogh's studies of existing works of the time still maintain his 'style', his brush strokes, use of color, rendering of form. No matter how many times he copies another's work he's still Van Gogh and his work expresses and reflects who he 'is'. There is no existing model from which we can accurately imitate the human creative process, so it begs the question, what is the nature of the imagery generated? Think of it like your hand writing and signature style. Even if it may appear similar to others, it's still uniquely identifiable as yours. If someone copies your writing style and signature to forge documents, it wouldn't just become ok as long as they forge new documents, in any case it would be immoral and illegal without your express consent. That is the nature of algorithmically generated imagery, complex forgery, but instead of the forgery being of an individual, it's modeled from billions of images scraped from the internet. This is entirely different from the human creative process.
  - williamcotton 4y ago
    Go type in any of those artists and/or the names of their works in to both Google Images and Stable Diffusion. Notice a difference? The image that is generated from SD is going to be so far from similar and something that someone else would choose instead of the plaintiff’s works it’s comical. Rather people are using SD to create music groove synths, remove people from the backgrounds of images, turn their dogs into Dali paintings, and many non-infringing uses (Sony v Universal). At some point VCRs, Google Images, Google Books and Stable Diffusion all make indiscriminate copies of many works. A factor that weighs heavy in both determining fair use or not is how useful the tool is for non-infringing cases, in favor for Google Books, against for TorrentFire-things. Since SD is basically incapable of recreating any of the plaintiff’s works and entirely capable of many things that has nothing to do with the plaintiff’s entire industry it weighs heavily in favor of fair use. The fact that the artist’s names come back is because the name is associated with a visual style. For Sarah Anderson you can get better results for her style by using “quirky, black and white cartoon, striped sweater, big eyes”… and then get back grotesque cartoons and sift through trash, inpainting for hours… and then end up with an entirely original cartoon looking thing that is noninfringing because styles cannot be copyrighted. Neither can ideas! (Seldon v Baker) Someone uploading “Sarah Anderson SD Tuned Model” to HuggingFace could be infringing on a trademark violation or a common-law publicity violation, but that’s already protected and has nothing to do with making copies for the sake of training an algorithm. I’m not sure it is currently illegal and should ever be considered wrong to have “Sarah Anderson is known for her quirky black and white cartoon characters with big eyes and black and white striped sweaters that she puts on daily calendars and sells on Amazon” be knowledge in the public domain as this is a fact about the world.
rose_ann_ 4y ago
You might want to learn how the models work before making inflammatory comments. To simplify the entire concept: Imagine you were an alien. You decide to work as an artist on earth. So you go on this thing the Humans call 'the internet', view millions of images & works of art, then start creating images & art pieces of your own based on the underlying rules you've gleaned from those millions of images you've viewed, without in any way keeping or saving a single copy of any of those images or artworks. Would you call the above copyright infringement? Or are you now claiming its 'copyright infringement' if anything or anyone apart from a human being views an image?
- sleeplessworld 4y ago
  Perhaps we are in new legal territory, where existing copyright laws and legal theory doesn't cover the issues completely? I wonder how things will go if (or when) similar AI technology enters the realm of production industries e.g. manufacturing or chemical. One can imagine that data of actual production plants and processes are obtained using drone cameras, pictures, schematics, text books etc. This data is then analyzed and synthesized into improved production flows, using the same method, which basically claims (it seems to me): we did not use any of your actual production knowledge, we just let the AI learn the principles. None of the original IP is in the AI working data. No patents are violated etc. If this technology is sold to the manufacturing companies - that can choose to opt out or opt in by buying or not bying - there probably isn't a problem. If this technology is used to establish competing manufacturing companies, where the economy can be significanty improved, and thus drive the existing companies out of business, then what? I think in the latter case, the AI company would be obliterated using lawsuits, without any thought given to the merit or legality of the AI tech. Or at least the AI company would be bought out using an "you can't say no" offer. And thus effectively shut down. Now I think of it, this seems to boil down to some philosphical issues where these thought experiments are needed and have value in order to work out what the issues really are. Perhaps an issue here is that there isn't much value on artists and works of art in general (with some economically strong exceptions like movies), but there is very strong protection of other industries - industries with vested interests, strong economies and monetary resources to defend against attackers.
Spivak 4y ago
Does this apply to GPT too? They also scraped massive amounts of copyrighted text. The situation with art is super weird because it seems the difference between them is how threatened they are not whether they actually have a compelling argument which is why there was a whole kerfufle about Copilot because devs felt threatened. When you pick people who aren’t in the situation where their livelihood depends on them holding a specific opinion on AI, like authors, the copyright issue suddenly doesn’t matter.
williamcotton 4y ago
> > " In an interview with Forbes, Midjourney CEO David Holz said that the dataset that the company trained its AI model on was “just a big scrape of the internet” and the company hadn’t sought consent from artists who owned the copyright to their work. “There isn’t really a way to get a hundred million images and know where they’re coming from,” Holz said." > So you just go ahead and steal copyrighted images without ask and attribution? It's not ever "stealing". It's a copyright infringement and only if it isn't in fair use. > This attitude is far beyond the pale and I, for one, hope that the artists succeed with their lawsuit. The fact that the model is based on scanning a hundred million images in order to function and that less images means it would not function as well and the fact that it is actually impossible to ask for a hundred million permissive uses is all highly relevant to judgements related to fair use. And of course that there are many non-infringing uses factors in heavily. In the case of Sony Corp. of America v. Universal City Studios, Inc. (1984), also known as the "Betamax case," the Supreme Court found that the sale of home video recording equipment (such as VCRs) was not copyright infringement, because the equipment had "commercially significant noninfringing uses." This principle, known as the "Betamax defense," was relied on by Google in the later case of Authors Guild, Inc. v. Google, Inc. (2015). In the Google Books case, the court found that Google's scanning of millions of books to create a searchable database was a fair use of the copyrighted works under copyright law. The court cited the Betamax case in finding that Google's use of the books had a transformative purpose, and that the creation of a searchable database had significant public benefits, including the ability to find and identify books, and the ability to create new forms of research and scholarship. The court also found that Google's use did not harm the market for the original books, as it did not serve as a substitute for the original books and that the use of the copyrighted material was minimal. Similarly, in the case of Stable Diffusion, it could be argued that the tool is being used for a transformative purpose, such as creating new and unique images (for inpainting new background imagery in place of removed people from vacation photos, or generating spectrographs of "70s disco rock", or converting a photo of a dog into a painting of a dog in the style of Dali) , and that it is not harming the market for the original works. Additionally, the fact that Stable Diffusion is free for the public to use may further support the argument that it is fair use, as it does not compete with the original images and does not generate commercial gain for the defendants.
poulpy123 4y ago
You don't ask permission to look at an image that is publicly on display. AFAIK there is no law saying that machine learning bots are forbidden to look at an image that is publicly on display.

jaggs 4y ago

I think it might help if you understand how AI art works. They're not scraping images, they're scraping rules. Completely different thing. https://i.redd.it/2f00l6vsso6a1.jpg https://i.redd.it/2f00l6vsso6a1.jpg

CaptainZapp 4y ago
Does that really matter if you can easily generate something that comes pretty close (at least in a superficial way) to the artist by just entering the artists name with your description?
- jaggs 4y ago
  That's a really good question, and it's key to the issue at hand I think. Any competent artist can create a lookalike style of another artist at will. It's just a matter of looking at the paint type, brushes, materials, strokes, composition, etc and copying that. In fact, here's Vincent Van Gogh copying other artists of his time - https://www.dailyartmagazine.com/van-gogh-copy/ https://www.dailyartmagazine.com/van-gogh-copy/ . How does this differ from a machine doing the output? Apart from speed and flexibility?
  - diemes1 4y ago
    Van Gogh's studies of existing works of the time still maintain his 'style', his brush strokes, use of color, rendering of form. No matter how many times he copies another's work he's still Van Gogh and his work expresses and reflects who he 'is'. There is no existing model from which we can accurately imitate the human creative process, so it begs the question, what is the nature of the imagery generated? Think of it like your hand writing and signature style. Even if it may appear similar to others, it's still uniquely identifiable as yours. If someone copies your writing style and signature to forge documents, it wouldn't just become ok as long as they forge new documents, in any case it would be immoral and illegal without your express consent. That is the nature of algorithmically generated imagery, complex forgery, but instead of the forgery being of an individual, it's modeled from billions of images scraped from the internet. This is entirely different from the human creative process.
- williamcotton 4y ago
  Go type in any of those artists and/or the names of their works in to both Google Images and Stable Diffusion. Notice a difference? The image that is generated from SD is going to be so far from similar and something that someone else would choose instead of the plaintiff’s works it’s comical. Rather people are using SD to create music groove synths, remove people from the backgrounds of images, turn their dogs into Dali paintings, and many non-infringing uses (Sony v Universal). At some point VCRs, Google Images, Google Books and Stable Diffusion all make indiscriminate copies of many works. A factor that weighs heavy in both determining fair use or not is how useful the tool is for non-infringing cases, in favor for Google Books, against for TorrentFire-things. Since SD is basically incapable of recreating any of the plaintiff’s works and entirely capable of many things that has nothing to do with the plaintiff’s entire industry it weighs heavily in favor of fair use. The fact that the artist’s names come back is because the name is associated with a visual style. For Sarah Anderson you can get better results for her style by using “quirky, black and white cartoon, striped sweater, big eyes”… and then get back grotesque cartoons and sift through trash, inpainting for hours… and then end up with an entirely original cartoon looking thing that is noninfringing because styles cannot be copyrighted. Neither can ideas! (Seldon v Baker) Someone uploading “Sarah Anderson SD Tuned Model” to HuggingFace could be infringing on a trademark violation or a common-law publicity violation, but that’s already protected and has nothing to do with making copies for the sake of training an algorithm. I’m not sure it is currently illegal and should ever be considered wrong to have “Sarah Anderson is known for her quirky black and white cartoon characters with big eyes and black and white striped sweaters that she puts on daily calendars and sells on Amazon” be knowledge in the public domain as this is a fact about the world.

rose_ann_ 4y ago

You might want to learn how the models work before making inflammatory comments. To simplify the entire concept: Imagine you were an alien. You decide to work as an artist on earth. So you go on this thing the Humans call 'the internet', view millions of images & works of art, then start creating images & art pieces of your own based on the underlying rules you've gleaned from those millions of images you've viewed, without in any way keeping or saving a single copy of any of those images or artworks. Would you call the above copyright infringement? Or are you now claiming its 'copyright infringement' if anything or anyone apart from a human being views an image?

sleeplessworld 4y ago
Perhaps we are in new legal territory, where existing copyright laws and legal theory doesn't cover the issues completely? I wonder how things will go if (or when) similar AI technology enters the realm of production industries e.g. manufacturing or chemical. One can imagine that data of actual production plants and processes are obtained using drone cameras, pictures, schematics, text books etc. This data is then analyzed and synthesized into improved production flows, using the same method, which basically claims (it seems to me): we did not use any of your actual production knowledge, we just let the AI learn the principles. None of the original IP is in the AI working data. No patents are violated etc. If this technology is sold to the manufacturing companies - that can choose to opt out or opt in by buying or not bying - there probably isn't a problem. If this technology is used to establish competing manufacturing companies, where the economy can be significanty improved, and thus drive the existing companies out of business, then what? I think in the latter case, the AI company would be obliterated using lawsuits, without any thought given to the merit or legality of the AI tech. Or at least the AI company would be bought out using an "you can't say no" offer. And thus effectively shut down. Now I think of it, this seems to boil down to some philosphical issues where these thought experiments are needed and have value in order to work out what the issues really are. Perhaps an issue here is that there isn't much value on artists and works of art in general (with some economically strong exceptions like movies), but there is very strong protection of other industries - industries with vested interests, strong economies and monetary resources to defend against attackers.

Spivak 4y ago

Does this apply to GPT too? They also scraped massive amounts of copyrighted text. The situation with art is super weird because it seems the difference between them is how threatened they are not whether they actually have a compelling argument which is why there was a whole kerfufle about Copilot because devs felt threatened. When you pick people who aren’t in the situation where their livelihood depends on them holding a specific opinion on AI, like authors, the copyright issue suddenly doesn’t matter.

williamcotton 4y ago

> > " In an interview with Forbes, Midjourney CEO David Holz said that the dataset that the company trained its AI model on was “just a big scrape of the internet” and the company hadn’t sought consent from artists who owned the copyright to their work. “There isn’t really a way to get a hundred million images and know where they’re coming from,” Holz said." > So you just go ahead and steal copyrighted images without ask and attribution? It's not ever "stealing". It's a copyright infringement and only if it isn't in fair use. > This attitude is far beyond the pale and I, for one, hope that the artists succeed with their lawsuit. The fact that the model is based on scanning a hundred million images in order to function and that less images means it would not function as well and the fact that it is actually impossible to ask for a hundred million permissive uses is all highly relevant to judgements related to fair use. And of course that there are many non-infringing uses factors in heavily. In the case of Sony Corp. of America v. Universal City Studios, Inc. (1984), also known as the "Betamax case," the Supreme Court found that the sale of home video recording equipment (such as VCRs) was not copyright infringement, because the equipment had "commercially significant noninfringing uses." This principle, known as the "Betamax defense," was relied on by Google in the later case of Authors Guild, Inc. v. Google, Inc. (2015). In the Google Books case, the court found that Google's scanning of millions of books to create a searchable database was a fair use of the copyrighted works under copyright law. The court cited the Betamax case in finding that Google's use of the books had a transformative purpose, and that the creation of a searchable database had significant public benefits, including the ability to find and identify books, and the ability to create new forms of research and scholarship. The court also found that Google's use did not harm the market for the original books, as it did not serve as a substitute for the original books and that the use of the copyrighted material was minimal. Similarly, in the case of Stable Diffusion, it could be argued that the tool is being used for a transformative purpose, such as creating new and unique images (for inpainting new background imagery in place of removed people from vacation photos, or generating spectrographs of "70s disco rock", or converting a photo of a dog into a painting of a dog in the style of Dali) , and that it is not harming the market for the original works. Additionally, the fact that Stable Diffusion is free for the public to use may further support the argument that it is fair use, as it does not compete with the original images and does not generate commercial gain for the defendants.

poulpy123 4y ago

You don't ask permission to look at an image that is publicly on display. AFAIK there is no law saying that machine learning bots are forbidden to look at an image that is publicly on display.