Painting with Gaussians

I investigated a gradient-descent-based concept adjacent to this back in like 2023 [0] inspired by the original 3DGS paper, I was pretty happy with the results I got with back-to-front alpha-blending and polar-coordinate gaussians for variable-curvature brush strokes. One of these days I should add another example gif to the repo besides just the Mona Lisa one that's currently there, but that's the only one I've got at the moment unfortunately.

[0] https://github.com/mkaic/painterbot

The effect looks decent in the foreground of some pictures, particularly the wolf and the cat. It really exaggerates the depth of field in the background, though, and those areas feel like someone ran a posterise after a blur, rather than brush strokes. Might just be a side effect of the author's predilection for bokeh, I guess. The Tokyo shot looks good throughout.

I’d suggest for the examples to pick images that are not mostly bokeh.

Intriguing and nice writeup. This however tries to encode all the hidden painter's knowledge. Would it be easier to generate realistic images from paintings and use these pairs to finetune an image generation model?

Super impressive results, and an awesome write up. Thanks :)

Very cool