Spaghetti Splatting Replaces Points in 3DGS with Curves: Hair and Wires Ready for Simulation

If you have ever tried to reconstruct a head with hair, a bundle of wires, or a bush from photographs, you know how it ends. The picture looks decent, but inside you have a point cloud, not strands. Moving such a model or applying physics to it is almost pointless. The Spaghetti Splatting paper targets exactly this weak spot.

The problem: a strand is not a set of points

3D Gaussian Splatting describes a scene with millions of small blurry "blobs." For walls, faces, and furniture this works great. For thin, long objects it does not. A single hair has to be assembled from dozens of blobs, and there is no connection between them. For rendering this is tolerable, but for everything else it is bad. You cannot ask where a hair begins and ends. You cannot bend it, because the system does not know which points belong to the same strand.

Until now, such scenes usually called for specialized methods for hair or vegetation. They perform worse in other cases, and the pipeline has to be assembled from several pieces.

The idea: a curve instead of a point

The authors propose changing the basic primitive. In classic 3DGS, each Gaussian describes a point mass. In Spaghetti Splatting it is stretched into a continuous line, that is, a curved segment or a spatial fiber. Standard 3DGS emerges as a degenerate case where the curve length is zero. That is why the authors call the method a strict generalization: everything 3DGS could do is preserved, and strands become a native data type.

The post mentions three technical details:

• linearization of the camera projection, so that curves are projected onto the screen quickly and correctly;
• support for beta kernels, which, according to the authors, speed up rendering;
• built-in anti-aliasing, which solves the perennial problem of thin lines: they flicker and break up at a distance.

According to the authors, on scenes with strand-like geometry (hair, wires, vegetation, meshes) the method outperforms specialized approaches.

What this means in practice

The main practical effect is not a prettier picture but the structure of the result. Strands come out long and connected, and they can be fed straight into a simulator of deformable strand and cloth dynamics. Previously, a manual or semi-automatic step lay between reconstruction and simulation: strands had to be extracted from the point cloud. Now this step, judging by the description, can be thrown away.

Who should care:

• teams building digital humans and avatars: hair is their main pain point;
• VFX and game studios that need to quickly digitize a real object and immediately make it move;
• those working in robotics and infrastructure scanning: cables and wires are everywhere there;
• 3D tool developers who want one format instead of a zoo of methods.

Limitations and what to check

For now this is research work. The post gives no concrete quality or speed figures and no data on production readiness, and comparisons with alternatives should be checked in the paper itself on arXiv (2611.12716) and on the project page. It is unclear how the method behaves on dense bundles where strands overlap and tangle, and how hard it is to integrate into existing engines and renderers.

The practical takeaway is simple. If you already work with 3DGS and keep running into hair, cables, or grass, look at the code and examples on the project page and run them on your own scenes. If you are only choosing a digitization technology, keep this work on your watch list. The idea of storing strands as strands rather than as a cloud of blobs looks like a natural step, and it may become the standard for such tasks.