Re:Illustrating Gutenberg library using Stable Diffusion
by spyder on Sunday September 04, 2022 @02:56PM
(#32716441)
The creators of stable diffusion and others are already doing that.
https://paperswithcode.com/dataset/simulacra-aesthetic-capti...
https://github.com/LAION-AI/laion-datasets/blob/main/laion-a...
Re:Illustrating Gutenberg library using Stable Diffusion
by nuclearnice3 on Sunday September 04, 2022 @08:15PM
(#32719075)
I like it. Your methodology seems analogous to a GAN [1]
https://en.wikipedia.org/wiki/Generative_adversarial_network
Edit: In other words...
1. Train your stable diffusion model using a corpus of normal human images.
2. Use the model to generate thousands of images.
3. Use human analysts to train another AI how to find "good" images in that set.
4. Generate ten million stable diffusion images.
5. Use the other AI to find the "good" ones.
6. Use those to train another stable diffusion model.
7. Goto 4.
(You lose some information each iteration, though, so it would probably eventually converge to grey goo or millions of pictures of Agent Smith.)