2D Gaussian Splatting for Geometrically Accurate Radiance Fields

Huang et al. observe that 3D GaussianSplatting (3DGS), while excellent for NovelViewSynthesis, is a poor representation when accurate geometry is required. Each 3D Gaussian is a volumetric primitive: its silhouette changes shape from one camera angle to another, so the surface inferred from a 3DGS field is not view-consistent. The authors propose 2D Gaussian Splatting (2DGS), which collapses each 3D volumetric primitive to a flat, oriented Gaussian disk in 3D space, intrinsically representing a tangent plane to the underlying surface.

A 2D disk has a single normal and a well-defined location in 3D, which makes the geometry it represents view-consistent: every camera sees the same plane. Recovering radiance from a set of such disks requires a perspective-correct splatting process, derived through explicit ray-splat intersection rather than the screen-space affine approximation that 3DGS uses. The authors implement this perspective-correct rasterisation efficiently on the GPU so that the runtime cost remains close to that of 3DGS.

Two regularisers strengthen the geometric output further: a depth distortion term encourages the depth distribution along each ray to be sharply peaked on a single disk, and a normal consistency term encourages the disks' normals to agree with the gradient of the rendered depth map. Together these produce noise-free meshes and clean surface normals that are competitive with NeuralImplicitSurface methods, while retaining 3DGS- style training speed and real-time rendering. 2DGS sits alongside SuGaR as one of the two main lines of work bridging the gap between fast Gaussian-based radiance fields and accurate geometry, and the released code has become a common baseline for surface-aware Gaussian methods.