Magnus Petersen Profile picture
Deep Learning methods dev for creativity 👾🖼️| Ph.D. student researching deep learning for molecular dynamic simulations @fias_science @CovinoLab

Aug 30, 2022, 5 tweets

Given the speed of #stablediffusion 's generation, automated image/prompt curation is becoming a necessity.

I've been using @RiversHaveWings aesthetics rating model to build an evolutionary algorithm that evolves a random prompt population to become more aesthetic. 1/x

A population is initialized randomly from lists like a dictionary, artist names, etc. Images are generated and given a score with the aesthetics model. Based on these scores prompts are selected to seed the next generation via crossover, mutation, etc of words. 2/x

The aesthetics model does not have the best taste but it filters out a lot of the chaff that comes from random prompts. This leads to the discovery of a lot of cool results from prompts that are gibberish that people won't usually come up with! 3/x

For example: "trireme Fantasy Art nondefiling rhinopharyngitis canaliculization cricotracheotomy conure atheology beret Aestheticism Vicente Juan Masip Ashley Bickerton aplustria pelodytoid"

All images in this post were found this way. Code coming soon(tm)! Till then have a look at @rivershavewings aesthetics rating model alone for filtering. It's on her Github. 4/x

Share this Scrolly Tale with your friends.

A Scrolly Tale is a new way to read Twitter threads with a more visually immersive experience.
Discover more beautiful Scrolly Tales like this.

Keep scrolling