via https://twitter.com/DigThatData/status/1584826328704221184
train a diffusion model to invert a "paint splatter" corruption process, so the noise latent would basically be jackson pollock paintings.
- https://github.com/Huage001/PaintTransformer
- https://github.com/arpitbansal297/Cold-Diffusion-Models
- if the corruption process includes deleting brushstrokes, that would give opportunity for the inverse processes to include predicting new brushstrokes
- My hope here is that we could train a "touchup" model that iteratively predicts new brushstrokes or paint splatter or whatever atomic unit of a visual artists labor
- using the PaintTransformer brushstrokes as the final target of the inverse process, maybe we could shuffle or randomize strokes for the forward corruption?
- pretrain on PaintTransformer, finetune on real art