Hexels replace the Gaussian kernel in 2D Gaussian Splatting with a soft hexagonal support function and a linear per-primitive color field, leaving the rasterization and compositing pipeline unchanged

computer-visiongaussian-splattingnovel-view-synthesisrenderingcomputer-graphics3d-reconstructionoutside

Hexels, introduced by Kaushik, Paul & Yuille (Johns Hopkins CCVL lab, arXiv:2610.37668, 2026), are planar surface primitives pitched as a drop-in replacement for the elliptical Gaussian kernel used in 2D Gaussian Splatting (2DGS) and the broader Gaussian-splatting family of scene representations. Each Hexel is centered at a 3D point with two orthogonal tangent directions and two scales — the same oriented-planar-patch parameterization 2DGS uses for its surfels. The substitution is specifically at the level of the per-primitive weighting function: instead of a Gaussian falloff, a Hexel's spatial support is a soft hexagonal boundary — a signed-distance-like term passed through a sigmoid, with a sharpness parameter controlling edge crispness. Color is not constant across a primitive, as in standard Gaussian or triangle splats, but varies linearly from a base RGB value via two RGB gradient vectors, letting one primitive represent smooth shading across its footprint.

The paper is explicit that everything else carries over unchanged: perspective-correct ray–plane intersection and front-to-back depth-ordered alpha blending are the same machinery 2DGS already uses. Only the weighting function and the color model change. This narrow, surgical framing — new primitive shape and color model, same rasterizer — is what lets the paper attribute its benchmark gains (see the matched-budget comparison notes below) specifically to the hexagon-plus-gradient design rather than to an unrelated pipeline change.

The hexagon choice is not arbitrary: the paper cites Thomas C. Hales's 2001 proof of the honeycomb conjecture (hexagons tile a plane with less perimeter than squares or triangles for equal area) as its justification, and reports its own appendix derivation that an optimally-oriented hexagon covers roughly 30% more area than a same-sized square for equal bounded color error — a figure this vault has not yet independently read in the appendix itself.