3 ms·

Alright - what if I take Stable Diffusion and further train it on my own art, and then use that to generate things? Or what if I don’t do that, but generate se

Alright - what if I take Stable Diffusion and further train it on my own art, and then use that to generate things? Or what if I don’t do that, but generate several images and spend hours making a composite? What if I start with my own photograph and generate things to use in a composite on top of that? It’s not so clear cut.

CRConrad 3y ago
> Alright - what if I take Stable Diffusion and further train it on my own art, and then use that to generate things? Operative word there: "further". If it's already trained on everyone else's input, then without that there wouldn't even be a Stable Diffusion for you to "further train on your own art", so YTF should you get sole ownership of the sum of it all? What you're saying is like "Alright - what if I take Linux and add in my own code, and then sell that as closed source?" Nope, it's already GPLed, and so should AIs trained on public data be.
jongjong 3y ago
It is clear cut because the Stable Diffusion AI you refer to was nevertheless trained on public data to get to that point were it could then be 'fine-tuned' with your own art. Your own art is just a small addition to its much, much larger input/training set. The size of these public training sets are truly massive. If you start with a Stable Diffusion program from scratch with all weights initialized at random (or whatever weights its initialization algorithm requires) and you train it on only your own photograph(s), then you will almost certainly find that it's not effective at all. In that case though, the low-quality output would be 100% your own. The reason why generative AIs are so good is precisely because they have access to a large body of public data. Without the public data, they would not be good no matter how good the underlying software/algorithm is.

3 ms·

Alright - what if I take Stable Diffusion and further train it on my own art, and then use that to generate things? Or what if I don’t do that, but generate se

by Auracle 3y ago

CRConrad 3y ago
> Alright - what if I take Stable Diffusion and further train it on my own art, and then use that to generate things? Operative word there: "further". If it's already trained on everyone else's input, then without that there wouldn't even be a Stable Diffusion for you to "further train on your own art", so YTF should you get sole ownership of the sum of it all? What you're saying is like "Alright - what if I take Linux and add in my own code, and then sell that as closed source?" Nope, it's already GPLed, and so should AIs trained on public data be.
jongjong 3y ago
It is clear cut because the Stable Diffusion AI you refer to was nevertheless trained on public data to get to that point were it could then be 'fine-tuned' with your own art. Your own art is just a small addition to its much, much larger input/training set. The size of these public training sets are truly massive. If you start with a Stable Diffusion program from scratch with all weights initialized at random (or whatever weights its initialization algorithm requires) and you train it on only your own photograph(s), then you will almost certainly find that it's not effective at all. In that case though, the low-quality output would be 100% your own. The reason why generative AIs are so good is precisely because they have access to a large body of public data. Without the public data, they would not be good no matter how good the underlying software/algorithm is.