From the archives:
"Over the last year I've looked into bringing various AI techniques into more traditional workflows to see if I can smush them together and make something new. So this time I wanted to try to bring it into something realtime. One thing I always wanted to do in Touchdesigner is to be able to call up random images that gave a sense of complexity, texture and color to this generative image, but had some kind of structure (ie not tooo random that you can't guide the form). Enter some guy who made a stable diffusion API thing for Touch (I'll credit him if this ever gets shared), and away we go.
Kinda.
It was still really complicated to set this up; I can maybe do a share one day. In a nutshell, the problem is that these txt to image things take a long time to process - even if it's just 10s, that doesn't really jive with a realtime piece. So this is constantly ringing up stable diffusion, generating images (based on the webcam input) under the hood, and what you see on the screen is the previous image being drawn with points.
Anyway I'm waffling, this generative poetry version of it does roughly this:

• I manually write in a list of text prompts, showing things that are easy to visually read (in the case below: clouds, sunrise, geometric shapes, water reflections, stormy waves, underwater, forest)

• Off to chat GPT to generate poetry with the prompts as a guide (btw, chatgpt seems to be not great at writing non-cheeseballs poetry).

• for every poetry line, it asks Stable Diffusion to generate an image based on the prompt, and loosely on the webcam image (good example around 20s when I hold up that pink post-it).

• A bunch of complicated touchdesigner stuff happens to it to "draw" the images.

• Oh, and it's loosely audiovisual (not very well atm, just based on audio signal), cos why not. This is all generative music so in theory this could be infinite too."