Give it scattered keypoints and it matches a full splat reconstruction

Two underwater Gaussian splatting papers appeared on Monday, each declaring state of the art on an evaluation set it had assembled itself, neither citing the other. Olaya Álvarez-Tuñón and Stella Graßhof have now done the thing that was missing: taken five underwater splatting systems with public code, run them under a single protocol with shared poses, initialisation, compute budget and evaluator, and pushed them across four water regimes. The conclusion is that the architecture matters less than the setup, and that the numbers the field reports do not show it.

Why it matters: Underwater splatting has a benchmark culture problem that is easy to state and hard to see: the standard photometric metrics reward a model for fitting the water rather than the scene behind it. Haze is smooth, low-frequency and highly predictable, so a system that commits opacity to the veil scores well on held-out views while placing its geometry in the wrong place.

Arrival.Space sells hosted reconstruction: you upload video, their servers turn it into a splat, you get it back. A week ago the company published Splat.js under an MIT licence — a library that does the whole job in a browser tab, on the viewer's own GPU, without the photographs ever leaving the device. As of this afternoon it scores better on the standard benchmark than the 2024 papers it was presumably meant to be a toy beside.

Why it matters: The division of labour in splatting has been stable for two years: capture on a phone, solve and train on a server or a workstation with CUDA, view in a browser. The browser was the last mile, and everything upstream of it was somebody's compute bill.

The three rear cameras on an iPhone see the same scene from viewpoints about five degrees apart, simultaneously, every time the shutter fires. An Apple Vision Pro's stereo pair sits roughly fifteen degrees apart. A Lytro Illum records a thirteen-by-thirteen grid of them at once. Shamus Li and colleagues point out that essentially every reconstruction pipeline takes one of those streams and discards the rest, then spends considerable effort inventing the parallax it just threw out.

Why it matters: The received wisdom is that consumer camera baselines are too small to matter — five degrees of separation between phone lenses is nothing next to walking around an object. So the field standardised on a moving monocular camera, and when the camera cannot move enough, on learned priors that hallucinate the missing views.

vis.gl has published a Gaussian splat module for luma.gl, the WebGPU and WebGL framework that deck.gl is built on. The package went up as a beta on Friday and a third build followed on Monday; it is MIT-licensed, it is about four hundred kilobytes of TypeScript, and the part that matters is that it reads a file format someone else designed.

Why it matters: Spark, World Labs' three.js splat renderer, introduced the .RAD format with its 2.0 release: a precomputed level-of-detail tree stored so that a viewer can pull arbitrary chunks over HTTP range requests instead of downloading a scene whole. It is a good design, and until now it had one implementation — the one that shipped with it.

On Monday, a group led from Shanghai AI Laboratory posted AquaFlow, a streaming Gaussian splatting SLAM system for underwater video. The same day, a group at HKUST posted NemoSplat, a feed-forward 4D splatting model for the same setting. The two papers share a diagnosis, disagree completely about the cure, and independently arrive at the same uncomfortable result buried in their comparison tables.

Why it matters: Almost everything interesting in reconstruction over the past two years has been downstream of 3D vision foundation models — VGGT and its descendants, which estimate camera poses and dense geometry from uncalibrated images and made pose-free capture practical. They were trained on terrestrial imagery.