This is my problem. The tech itself is fine, no one is arguing about training data and making art from trained data.
But the source of all of that data was ripped without artists consent. They did not agree to take part in this. (And no, I don’t think clicking “I Accept” 15 years ago on DeviantArt should count, we had no concept of this back then). Then on top of that people are profiting off of the stolen art.
I’m pretty sure this whole issue has to end either in some catastrophe or the complete abolishon of interlectual property rights. Which I already don’t have any love for so I’m fairly convinced we should see artists and inventors get their needs met and being able to realise their projects as a separate issue from them effectively owning ideas.
Being inspired on vs copying is what I had in mind when I created my comment. I came to the conclusion that AI can’t be creative and can’t be inspired because it takes a 1:1 copy of the picture and stores it into a weighted neuronal network. Therefore it can also 1:1 recreate the picture and manipulate/change it or combine with other images with patterns that it learned. At the end the picture is stored on a silicon device but instead of a ordered structure its stored in a for us chaotic structure which could easily reassmble it back to the original.
because it takes a 1:1 copy of the picture and stores it
What makes you think that? This is wrong. Sure you can try and train a neuronal network to remember something exactly. But this would waste gigabytes of memory and lots of computing for some photo that you could just store on the smallest thumbdrive as a jpg and clone it with the digital precision, computers are made for. You don’t need a neural net for that. And once you start feeding it the third or fourth photo, the first one will deteriorate and it will become difficult to reproduce each of them exactly. I’m not an expert on machine learning, but i think the fact that floating point arithmetic has a certain, finite precision and we’re talking about statistics and hundreds of thousands to millions of pixels per photo makes it even more difficult to store things exactly.
Actually the way machine learning models work is: It has a look at lots of photos and each time adapts its weights a tiny bit. Nothing gets copied 1:1. A small amount if information is transferred from the item into the weights. And that is the way you want it to work to be useful. It should not memorise each of van gogh’s paintings 1:1 because this wouldn’t allow you to create a new fake van gogh. You want it to understand how van gogh’s style looks. You want it to learn concepts and store more abstract knowledge, that it can then apply to new tasks. I hope i explained this well enough. If machine learning worked the way you described, it would be nothing more than expensive storage. It could reproduce things 1:1 but you obviously can’t tell your thumbdrive or harddisk to create a Mona Lisa in a new, previously unseen way.
Just take for example Stable Diffusion and tell it to recreate the Mona Lisa. Maybe re-genrate a few times. You’ll see it doesn’t have the exact pixel values of the original image and you won’t be able to get a 1:1 copy. If you look at a few outputs, you’ll see it draws it from memory, with some variation. It also reproduces the painting being photographed from slightly different angles and with and without the golden frame around it. Once you tell it to draw it frowning or in anime style, you’ll see that the neural network has learned the names of facial expressions and painting styles, and which one is present in the Mona Lisa. So much that it can even swap them without effort.
And even if neural networks can remember things very precisely… What about people with eidetic memory? What about the painters in the 19th century who painted very photorealistic landscape images or small towns. Do we now say this isn’t original because they portrayed an existing village? No, of course it’s art and we’re happy we get to know exactly how things looked back then.
Well, you mentioned that it could reproduce the imagine 1:1 which is just my entire point. It doesn’t matter what your thumb drive can’t do.
And I guess the main point is that every pixel is used and trained without changes, making it kinda a copyright issue as some images don’t even allow to be used somewhere and edited.
I think what i was trying to say is something different: It can not do that. only theoretically under specific circumstances, more a maths homework assignment, but not really in practice.
This is my problem. The tech itself is fine, no one is arguing about training data and making art from trained data.
But the source of all of that data was ripped without artists consent. They did not agree to take part in this. (And no, I don’t think clicking “I Accept” 15 years ago on DeviantArt should count, we had no concept of this back then). Then on top of that people are profiting off of the stolen art.
I’m pretty sure this whole issue has to end either in some catastrophe or the complete abolishon of interlectual property rights. Which I already don’t have any love for so I’m fairly convinced we should see artists and inventors get their needs met and being able to realise their projects as a separate issue from them effectively owning ideas.
Isn’t that a bit like someone faking a painting? Let’s say by Monet? This can be everything from 100% alright to illegal.
In addition to that, there’s also a difference between being inspired by, or copying something.
I think all of that is just a variation of an old and well known problem.
Being inspired on vs copying is what I had in mind when I created my comment. I came to the conclusion that AI can’t be creative and can’t be inspired because it takes a 1:1 copy of the picture and stores it into a weighted neuronal network. Therefore it can also 1:1 recreate the picture and manipulate/change it or combine with other images with patterns that it learned. At the end the picture is stored on a silicon device but instead of a ordered structure its stored in a for us chaotic structure which could easily reassmble it back to the original.
What makes you think that? This is wrong. Sure you can try and train a neuronal network to remember something exactly. But this would waste gigabytes of memory and lots of computing for some photo that you could just store on the smallest thumbdrive as a jpg and clone it with the digital precision, computers are made for. You don’t need a neural net for that. And once you start feeding it the third or fourth photo, the first one will deteriorate and it will become difficult to reproduce each of them exactly. I’m not an expert on machine learning, but i think the fact that floating point arithmetic has a certain, finite precision and we’re talking about statistics and hundreds of thousands to millions of pixels per photo makes it even more difficult to store things exactly.
Actually the way machine learning models work is: It has a look at lots of photos and each time adapts its weights a tiny bit. Nothing gets copied 1:1. A small amount if information is transferred from the item into the weights. And that is the way you want it to work to be useful. It should not memorise each of van gogh’s paintings 1:1 because this wouldn’t allow you to create a new fake van gogh. You want it to understand how van gogh’s style looks. You want it to learn concepts and store more abstract knowledge, that it can then apply to new tasks. I hope i explained this well enough. If machine learning worked the way you described, it would be nothing more than expensive storage. It could reproduce things 1:1 but you obviously can’t tell your thumbdrive or harddisk to create a Mona Lisa in a new, previously unseen way.
Just take for example Stable Diffusion and tell it to recreate the Mona Lisa. Maybe re-genrate a few times. You’ll see it doesn’t have the exact pixel values of the original image and you won’t be able to get a 1:1 copy. If you look at a few outputs, you’ll see it draws it from memory, with some variation. It also reproduces the painting being photographed from slightly different angles and with and without the golden frame around it. Once you tell it to draw it frowning or in anime style, you’ll see that the neural network has learned the names of facial expressions and painting styles, and which one is present in the Mona Lisa. So much that it can even swap them without effort.
And even if neural networks can remember things very precisely… What about people with eidetic memory? What about the painters in the 19th century who painted very photorealistic landscape images or small towns. Do we now say this isn’t original because they portrayed an existing village? No, of course it’s art and we’re happy we get to know exactly how things looked back then.
Well, you mentioned that it could reproduce the imagine 1:1 which is just my entire point. It doesn’t matter what your thumb drive can’t do.
And I guess the main point is that every pixel is used and trained without changes, making it kinda a copyright issue as some images don’t even allow to be used somewhere and edited.
I think what i was trying to say is something different: It can not do that. only theoretically under specific circumstances, more a maths homework assignment, but not really in practice.