Gallery of Unseen #23
Description
It's never meant to be a NaNoGenMo entry; I was just experimenting with a bunch of new tools and ideas. But, eventually, it converged into an imaginary guidebook for some strange, non-existing museum, so I decided to share it here.
The short story is:
- I ran the evolutionary search of the optimal prompt for the Stable Diffusion XL model to maximize the aesthetic scores of results measured by Google's pretrained NIMA model.
- I left it running overnight and got 50K+ of completely insane images; it was impossible to even check them all out.
- So I took the top 20K by scores, clusterized them stylistically to get pictures in the same visual style first.
- Then I deduplicated them semantically and ended with ~200 diverse strange images of non-existing art and architecture.
- Finally, I generated a short essay for each image and compiled them into the imaginary guidebook for non-existing museum.
The final PDF is here.
The code (pretty messy) is here.
I'm going to clean it up later, and also to write some post with more technical details; I will share the link here then.