The sketch perturbs the generation, but that’s about it. Going with the kitten prompt, I’m unable to get any given pose or positioning of the kitty. Drawing a second cat only changes the pose of the only one again. Mirror image the sketch: the tail curve changes. Add a random dot: get a whole new pose.
Probably related tweet: https://twitter.com/steveruizok/status/1729190769708376295
I wasn't able to get perceptibly better results than I usually do with simply prompt engineering with stable diffusion. Maybe my art skills aren't good enough.
I drew a face with a beard and a hat and I got someone off in the distance. I had to really enforce common stable diffusion terms to get it to be a portrait and not a woman. You'd think a long beard and a hat would have been enough of an indication without me having to express it in text
I drew a cat laying on its back and got back a head with a donut-shaped tail on top.
It looks like this model only knows how to draw specifically kittens in very specific positions?
Another very similar project: https://github.com/flowtyone/flowty-realtime-lcm-canvas
I really can't figure out what relation the doodle has to what's generated. I draw a dot, I get a cat as per the prompt. I add another dot, or a line; the newly generated image changes, but there's no predictability. We can call this 'AI', but is there really any intelligence here?
Seems like a call to stable diffusion's API without any real intent.