Daily digest — 2026-04-25

Gaussian Splatting day. The plan was: read the original paper, then read its three most-cited follow-ons, then figure out where the surface story is.

The anchor is 3D Gaussian Splatting (Kerbl et al., 2023). Compared to Instant-NGP, splatting trades implicit-volume ray marching for explicit anisotropic primitives rasterised front-to-back. The headline number — real-time at training-time quality on Mip-NeRF 360 — is what reshaped the field. The graphdeco-inria/gaussian-splatting reference is non-commercial, which has been awkward for downstream products; the nerfstudio-project/gsplat Apache-2.0 reimplementation is what almost everyone is actually running. The gsplat math supplement is worth reading alongside the original paper — the projection math is more subtle than the rasteriser code makes it look.

SuGaR (Guédon et al., 2023) is the surface story for splatting: regularise Gaussians toward a manifold, then extract a mesh via Poisson reconstruction on the resulting near-surface distribution. It's the closest splatting analogue to NeuS, and the comparison is worth doing — SuGaR is faster but more capture-sensitive.

2D Gaussian Splatting (Huang et al., 2024) collapses each primitive to a view-consistent oriented disk. The geometry is cleaner because the depth ambiguity inherent in volumetric 3D Gaussians goes away — at a small cost in view-synthesis fidelity. The right move when you actually want a mesh.

Scaffold-GS (Lu et al., 2023) replaces the "free Gaussian soup" with anchor points that emit view-conditioned offsets, which prunes redundant primitives in low-frequency regions. Worth contrasting with the densification revisit later this week.

Five papers, ~50 min. Tomorrow I want to see what happens when you push splatting into SLAM territory.