Tiny Online Tools
Fewer splats, same scene
A capture straight out of training is not optimised for anything. It has as many Gaussians as the optimiser happened to end up with, and a large share of them contribute almost nothing to what you see. Removing those is the single most effective way to make a capture usable on a phone or in a web page — and unlike quantisation, the splats that survive are untouched.
Three ways to choose what goes
By contribution is the default. Each splat is scored by its opacity multiplied by the area of its largest cross-section — a stand-in for how much of the screen it covers — and the lowest scores go first. Detail survives; thin background haze goes early.
Evenly keeps every nth splat. It is blind to importance, but it preserves the spatial distribution exactly, which is the right choice for an already-uniform capture where importance scoring would strip one region and leave another dense.
By opacity drops everything below a threshold. It is the only one of the three whose behaviour you can predict without running it, which makes it a good match for a capture full of near-invisible splats.
The contribution score has no camera in it. It approximates screen coverage rather than measuring it, so a splat that matters from one viewpoint and not another gets a single score. Check the result from the angles you care about.
See the cost before you pay it
The target count and the file size it implies are shown while you drag the slider — no work is done until you ask for it, because the size of a splat file follows directly from the count and the spherical-harmonic degree.
Compare, then export
The two viewports share a camera. Orbit either one and the other follows, which is the only honest way to judge whether 40 % of the splats were expendable. Export as PLY, SPZ, KSPLAT or SPLAT when you are satisfied.
Everything runs in a background worker in this page. The capture is never uploaded.