Gaussian splatting works because it never optimizes a point directly. Each point is blurred by a soft kernel, which gives training something to push against, and the same recipe has since been carried over to bounded Beta kernels and to triangles.
Spaghetti Splatting applies it to curves. The paper comes from George Shramko, Shrisudhan Govindarajan, Rudy Lam, Yizheng Xie, Daniel Rebain, David B. Lindell and Andrea Tagliasacchi, across Simon Fraser University, Wayve Labs, the University of British Columbia and the University of Toronto. It represents a scene as chains of short straight segments, each blurred by a kernel, and shows that the result approaches a smooth curve as the segments shrink.
Each segment is rendered on its own and composited along the ray, so connected segments meet without doubling up at their shared joints, and splitting a segment in two leaves the image exactly unchanged when both halves keep the parent's density. The camera projection is simplified once per segment to keep training fast, which also lets the method use Beta kernels alongside Gaussians. A segment of zero length is an ordinary 3D Gaussian, so the representation contains 3DGS as a special case. Anti-aliasing is built in.
On wires, spokes and netting, point methods scatter points along each stroke, while Spaghetti Splatting fits continuous strands that follow the edges.
Against Gaussian Haircut, which leans on hair-specific priors, the recovered strands score 61.2 against 46.9 on the paper's strand accuracy measure at a 2 mm threshold, with no learned hair prior. Those strands pass directly into a physics simulation.
Fine texture still takes many short segments, and the authors list real hair capture as an open problem.
The project page lists code as coming November 23, and the paper is available on arXiv.
