Meet Spaghetti Splatting, a new 3D reconstruction and differentiable rendering technique that extends 3D Gaussian Splatting from individual points to continuous curves. And it could be particularly useful for rendering and simulating hair, wires, vegetation, and other complex strand-like geometry. Traditional 3D Gaussian Splatting represents scenes using collections of individual 3D Gaussians. It delivers impressive photorealistic rendering, but reconstructing long, connected structures remains challenging. Spaghetti Splatting takes a fundamentally different approach. Instead of isolated point primitives, it uses connected, learnable curve segments that can reconstruct thin geometry while maintaining structural continuity. Here's what makes the research interesting: • Beyond 3DGS: mathematically generalizes Gaussian Splatting from point-based Dirac masses to continuous curvilinear representations. • Connected 3D geometry: reconstructs hair strands, cables, wires, meshes, and vegetation as continuous structures rather than scattered points. • Differentiable rendering: supports optimization of curved geometry directly from images. • Gaussian and Beta kernels: supports different convolution kernels for flexible geometric representation. • Efficient rasterization: linearizes camera projection per segment to accelerate training and rendering. • Native anti-aliasing: improves visual stability across different viewing distances. • Competitive reconstruction quality: the authors report improvements over specialized line-based reconstruction methods. • Physics-ready structures: connected strands can be used in simulations involving deformable curves, including hair dynamics. The important distinction is that Spaghetti Splatting doesn't merely make thin objects look better. It reconstructs connected geometry that can potentially become part of a physical simulation pipeline. That could help bridge the gap between neural rendering, 3D reconstruction, and traditional computer graphics workflows. Physics compatibility doesn't mean the method includes a complete cloth or hair simulation engine. Its contribution is producing connected geometry suitable for downstream simulation. Research paper: https://arxiv.org/abs/2611.12716 Project page and interactive examples: https://spaghetti-splatting.github.io/ Research by George Shramko and collaborators from Simon Fraser University, Wayve Labs, the University of British Columbia, and the University of Toronto. Could connected neural primitives become the next major step beyond point-based 3D Gaussian Splatting? #SpaghettiSplatting #GaussianSplatting #ComputerGraphics