Tiny Online Tools
Three different kinds of rubbish
A Gaussian splat capture almost always carries three separate problems, and lumping them together makes it impossible to tell what a cleaner actually did. This tool runs them as three passes and reports each one on its own.
Near-transparent splats
Optimisation leaves behind Gaussians whose opacity has collapsed towards zero. They draw nothing but still occupy memory, still take part in the depth sort every frame, and still go into the file. On many captures they are a tenth of the total. The threshold is yours to set; 0.02 is a safe default that removes what is genuinely invisible.
Blown-up splats
A splat whose largest axis is a serious fraction of the whole scene is not describing a surface — it is a smear of translucent colour laid across everything behind it. The size limit is expressed as a fraction of the scene diagonal, so it means the same thing whether you are cleaning a coffee cup or a cathedral.
Floaters
These are the wisps of nothing hanging in mid-air, usually where the capture had too few views. They are found statistically: each splat's mean distance to its nearest neighbours is compared with the scene average, and anything more than a few standard deviations out is treated as isolated.
This test cannot tell a floater from a genuinely isolated object. A lamp hanging in space looks exactly like noise to it. That is why the pass can be switched off, and why the strictness is adjustable.
Check before you keep it
The two viewports share a camera, so you can orbit either side and see the same angle on both. Look at the cleaned side before you export — the point of a cleaner is not to remove as much as possible, it is to remove only what you did not want.
The cleaned capture can be written as PLY, SPZ, KSPLAT or SPLAT. The whole thing runs in a background worker inside this page; the capture is never uploaded.