Weekly synthesis — 2026-W17

The thread running through this week's reading: 3D Gaussian Splatting is no longer a standalone view-synthesis method. It's swallowing the two adjacent sub-fields — implicit-surface reconstruction and visual SLAM — that the NeRF era had treated as separate problems. Three years after 3DGS, the explicit-primitive representation is doing most of what the implicit-field representation used to.

The surface absorption

The 2021 surface trilogy (UNISURF, NeuS, VolSDF) settled on a particular design: represent geometry as a signed distance function, render it via volume rendering, train per-scene from photometric loss. MonoSDF added monocular cues to make the training tractable on fewer views.

Two splatting papers from this week argue that you don't need any of that machinery to get a mesh. SuGaR (Guédon et al., 2023) regularises 3D Gaussians toward a manifold and then runs Poisson surface reconstruction on the densely-sampled near-surface points. 2D Gaussian Splatting (Huang et al., 2024) goes further and collapses each primitive to a view-consistent oriented disk, sidestepping the depth ambiguity that 3D Gaussians have when you ask for geometry instead of radiance. The "surface story" for splatting is no longer "bolt an SDF onto it" — it's "constrain the primitives to behave like a surface in the first place".

This isn't strictly better than NeuS-line methods. The implicit-surface reconstructions are still sharper on objects with thin structures and self-occlusion. But splatting trains in minutes where NeuS takes hours, and that ratio is what is winning over downstream users.

The SLAM absorption

The neural-SLAM line — starting with NICE-SLAM (Zhu et al., 2021) and continuing through Point-SLAM and Co-SLAM — was built around implicit feature grids or point-cloud feature stores, with NeRF-style volume rendering as the photometric supervision signal. Tracking ran as joint optimisation of pose and the implicit field. Quality was good; speed never quite was.

GS-SLAM and Gaussian Splatting SLAM / MonoGS, both from late 2023, do the obvious substitution: same online keyframe management, same pose-and-map joint optimisation, but the map is now an ever-growing pile of 3D Gaussians rasterised at frame rate. MonoGS reports the highest-fidelity photometric reconstructions of any monocular SLAM system in this corpus. Notably the classical-geometry baseline — DROID-SLAM — is still the tracking-accuracy reference, because deep BA on learned features remains the most robust pose engine. The Gaussian Splatting SLAM systems are beating DROID-SLAM on the map but not on the trajectory.

This is the open question that runs across both absorptions: how do these systems hold up under capture conditions that the splatting paradigm implicitly assumes are benign? See multi-view-consistency for the detailed version.

Density, generalisation, and the sparse-view frontier

Two more arrivals worth flagging. Revising Densification (Bulò et al., 2024) is a corrective on the original 3DGS recipe — the clone / split / prune schedule was undertuned, and a more principled treatment of the split criterion produces meaningfully better fidelity at the same primitive count. This is the kind of paper the corpus has too few of: nothing fancy, just careful re-examination. It complements Scaffold-GS's argument that the "free Gaussian soup" wastes capacity in low-frequency regions.

On the generalisation front, MVSplat (Chen et al., 2024) and latentSplat (Wewer et al., 2024) both feed-forward predict Gaussian parameters from sparse-view inputs, eliminating per-scene optimisation entirely. Splatting plus learned priors — the next phase that next week's reading will pick up in earnest.

What to read next

Two threads to chase into W18: (1) the diffusion-into-3D line that DreamGaussian bridged into the splatting representation, and (2) the large reconstruction model line (LRM and successors) that asks whether per-scene optimisation is necessary at all.