Painting with Gaussians

This coder taught a computer to paint, and the comments got wonderfully messy

TLDR: A developer built a tool that turns photos into painting-like images by following the picture’s outlines instead of blindly guessing. Commenters were split between cheering the clever approach, suggesting a simpler AI shortcut, and roasting the sample images for possibly making the results look better than they are.

A developer showed off a wild experiment: instead of having a computer "guess" its way toward a painted-looking image, they used the picture’s own outlines and shapes to decide where digital brush marks should go. In plain English, the program tries to follow the natural borders in a photo so the result looks more like a hand-made painting and less like a blurry filter. The creator also took a swipe at slower trial-and-error methods, basically saying, why let the machine fumble around when the image is already telling you where the details are?

And the comments? A tiny but delightfully chaotic art-school critique session. One camp was pure applause — "Very cool" and "Super impressive results" energy — the kind of supportive cheering that makes indie tech demos feel like mini movie premieres. But then came the spicy side-eye. One commenter essentially asked, why do all this hidden-art-teacher work at all? Why not just train an image model on paintings and be done with it? That instantly turned the post into a classic tech culture split: clever handcrafted approach versus "just use AI" shortcuts.

Then came the funniest jab of the thread: a deadpan complaint that the sample images were "mostly bokeh" — a brutally concise way of saying the examples might be doing the demo some favors. Another commenter rolled in with a humblebrag-meets-show-and-tell about their own brush-stroke project and a lone Mona Lisa gif, which is exactly the kind of nerdy one-upmanship the internet lives for. In other words: cool invention, but the real painting was in the comments

Key Points

  • The article builds on a prior edge-aware pixelation tool by reusing image edge information to guide digital painting.
  • Edges are presented as a source of information for stroke placement, scale, and orientation because they indicate boundaries and detail density.
  • The proposed brush-stroke representation is a 2D Gaussian splat with parameters for position, shape, rotation, color, and opacity.
  • The article contrasts its edge-guided method with prior Gaussian-painting systems that rely on gradient descent to fit splats to an image.
  • An initial attempt using a reference rasterizer with additive blending is reported to work poorly when thousands of splats are seeded directly rather than optimized.

Hottest takes

"Would it be easier to generate realistic images from paintings" — MeteorMarc
"pick images that are not mostly bokeh" — shen
"just the Mona Lisa one" — mkaic
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.