Spaghetti Splatting Fixes 3D AI's Biggest Blind Spot on Hair and Wires
A new splatting method extends 3D Gaussian Splatting from points to connected curves, producing strands that work for hair, wires, and physics simulation.
- Spaghetti Splatting generalizes 3D Gaussian Splatting by replacing point primitives with polyline curves convolved with Gaussian or Beta kernels.
- A zero-length spaghetti equals a Gaussian splat, making the representation a strict superset of 3DGS.
- Per-segment camera linearization enables closed-form rendering and fast training with non-Gaussian kernels.
- Outperforms specialized line methods and Gaussian Haircut on hair datasets while staying competitive on Mip-NeRF 360.
- Produces connected strands directly usable in physics simulation, with native anti-aliasing from the convolution formulation.
- Paper and project page available at spaghetti-splatting.github.io; code not yet released.
3D Gaussian Splatting, or 3DGS, reconstructs renderable scenes from posed photographs using collections of anisotropic Gaussian primitives. Each primitive remains independent, even when stretched into a strand-like shape, so wires, bicycle spokes, and hair can dissolve into disconnected elements outside the training views. Researchers from Simon Fraser University, Wayve, the University of British Columbia, and the University of Toronto address that limitation in Spaghetti Splatting, which replaces each point primitive with a connected polyline while retaining differentiable splatting.
The change gives the representation explicit continuity along thin structures. It also preserves the familiar 3DGS optimization loop: render the current scene, compare it with reference images, and backpropagate the image loss into geometry and appearance parameters.
A curve inside each splat
A conventional Gaussian splat can be described as a Dirac mass convolved with a Gaussian kernel. A Dirac mass is an idealized point that concentrates all its weight at one location; convolution spreads that weight into a renderable footprint. Related techniques, including elliptical weighted average splatting and Poisson reconstruction, use the same broad pattern.
Spaghetti Splatting changes the underlying mass from a point to a polyline, which is a sequence of connected straight segments. The method then convolves that geometry with either a Gaussian kernel or a beta kernel. When a segment’s two endpoints coincide, the construction reduces to the point-based 3DGS primitive, making the representation a strict generalization of Gaussian splatting.
Dense sampling would turn each curve into another point cloud and increase the primitive count. The authors instead derive segment-level convolution formulas and show that the polyline result converges to the convolution of a continuous curve as segment lengths shrink.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
