I saw this vox video and, before it was over, started playing around with prompting using an AI called Big Sleep. If you don’t know, prompting is the process of putting text into a text-to-image neural net. While there are not many options for software yet, I found big Sleep was a bit better than deep-daze. What’s crazy is how fast this stuff is improving.
While playing with this, I thought, why not set up image detection to automatically create new prompts. And that is what you are looking at here. I started with a prompt of “Frogs diving” and ran image recognition via detrectron2 every 200 images. As objects are found, they are added to a dictionary, and the least common terms are set as the prompt target. This lets the AIs find and amplify novel things in their own images. Because the image recognition AI does not always find something in the images, and because prompts tend to repeat, I added a few deceptive terms randomly picked from when a prompt stagnates.
I have seen others pipe in lines of a story, but this is a set of AIs telling their own story. If you prefer to watch without me talking, check out the link in the description.
As always my code is GPL and can be found here.