SpaghettiSplats: A Rigorous Generalization of 3DGS That Excels at Thin Structures

Author taiyasaki released a new rendering method, SpaghettiSplats: it treats 3D Gaussian Splatting as a convolution of Dirac masses and Gaussian kernels and derives a rigorous generalization, extending Dirac masses to lines and curves—naturally recovering thin structures that traditional 3DGS struggles to represent.

Confirmed

• The 3D convolution has a closed-form solution, further approximated with EWA (elliptically weighted averaging); rendering frame rate is on par with 3DGS-MCMC + SV8
• Primitives can be subdivided at arbitrary locations with rendering unchanged
• Because geometry is decoupled from convolution, the method gets covariance alignment for free and can simulate physical effects directly on primitives
• The project is led by George Shramko, a first-year PhD student at SFU advised by Andrea Tagliasacchi, researching differentiable 3D representations suitable for editing and motion

Why it matters

• Thin structures (lines, filaments) are a recognized weakness of traditional 3DGS; this method provides a mathematically rigorous path to generalization
• The covariance alignment and physics simulation enabled by decoupling geometry from convolution lay new groundwork for editable, animatable 3D representations

2026-11-11 ~ 2026-11-11 · 5 related posts

• [source] SpaghettiSplats: a strict generalization of 3DGS extending splats to lines and curves — taiyasaki · 2026-11-11
• SpaghettiSplats: closed-form 3D convolution with EWA matches 3DGS-MCMC frame rates — taiyasaki · 2026-11-11
• [source] SpaghettiSplats primitives subdivide anywhere with rendering unchanged — taiyasaki · 2026-11-11
• [source] SpaghettiSplats decouples geometry from convolution for direct physics simulation — taiyasaki · 2026-11-11
• SpaghettiSplats led by first-year PhD student George Shramko at SFU — taiyasaki · 2026-11-11