How to Remove Floaters From a Gaussian Splat Scene
Floaters, ghosting and a blown-out sky are the three problems a first splat almost always has. How to remove floaters from a gaussian splat, and deal with the other two, before writing off the capture.
A floater is a stray Gaussian sitting in empty space where there should be nothing: a fleck hanging over a couch, a faint smear above the roofline, a haze where the sky should be. Almost every first capture has some. Knowing how to remove floaters from a gaussian splat is mostly a matter of understanding why the optimizer put them there in the first place, because the fix is different depending on the cause.
The three problems tend to travel together. Floaters cluster where the reconstruction had the least to work with. Ghosting shows up where something moved. A blown-out sky causes both at once, and it is usually the reason the other two are worse than they need to be.
Why does a gaussian splat come out with floaters, ghosting or a blown-out sky?
A splat is built by an optimizer that compares a rendered guess against every photo of the scene and adjusts millions of Gaussians until the guess matches. It needs two things to do that well: texture to match between frames, and enough camera angles on the same surface to triangulate where it sits in space. Take either one away and the optimizer still has to put something there, and what it puts there is usually wrong.
A blown-out sky is the clearest case. An overexposed sky is close to a flat white field with no texture at all, so there is nothing for the optimizer to match frame to frame, and no parallax to tell it how far away that white actually is. Rather than leave the volume empty, it scatters low-confidence Gaussians through the space where the sky should be, and those scattered Gaussians are exactly what shows up as floaters near the top of the scene.
Ghosting has a different cause: something changed between the frames that were supposed to show the same static surface. A door that swung open partway through a walk-around, a shadow that moved as the sun shifted, a person crossing behind the camera. The optimizer can't tell which frame was "correct," so it averages both states into a soft double-image at that spot, which reads as a smear rather than a hard edge.
Floaters elsewhere in the scene are usually the same failure at a smaller scale: an edge of the capture volume where only one or two cameras saw a surface, a reflective countertop with no stable texture, a gap in coverage where nobody walked. Different cause, same result: the optimizer places a Gaussian it has almost no evidence for.
How to remove floaters from a gaussian splat, step by step
Fixing this in order matters. Each step removes a category of floater the next step can't touch on its own.
Crop the working volume first
Most of the floaters in a typical room capture live outside the space anyone actually cares about: the sky dome above a roofline, the floor beyond the last few feet anyone walked. Defining a bounding box around the real subject and dropping everything outside it removes that entire category in one pass, before touching a single individual Gaussian.
Trim the stray Gaussians that remain
What's left inside the crop is the harder set: floaters mixed in with the geometry you want to keep. These are usually the ones with the least evidence behind them, low opacity, unusually large, or seen from only a frame or two, and trimming them is the specific job a splat editor exists for. In Captures Studio's Splat Editor, that happens in the browser once a capture is back from processing, with no export into a separate desktop tool and back.
Deal with ghosting by hand, not by filter
A filter that removes low-confidence Gaussians won't reliably catch a ghost, because a ghost is often made of Gaussians the optimizer was fairly confident about, just confident about two conflicting things in the same place. This one gets selected and removed directly, or avoided at the source by re-shooting the frames where the movement happened.
Recapture the sky properly rather than patching it
Cropping out a blown-out sky removes the floaters it caused, but it does not put a sky back. There is no photographic information left to reconstruct once a region was captured at pure white. If a real sky matters for the final scene, the fix happens before reprocessing: expose for the brightest part of frame on the next pass, even if that means the ground reads a little dark, and let the cloud pipeline rebuild from a corrected set. That same pipeline that produced the first draft, typically 10 to 120 minutes depending on the capture, will do it again for the corrected one.
Where cleanup holds up
A crop and a trim pass will fix almost everything caused by missing coverage or a blown highlight, because the underlying geometry the optimizer needed was either recoverable from other frames or simply outside the region anyone needs. A living room shot with reasonable overlap, minus a scorched sky and a stray fleck over the sofa, comes back clean after ten minutes of editing.
Where cleanup falls down
Cleanup cannot recover a surface that was never photographed from a workable angle. If nobody captured the underside of a table or the far side of a narrow gap between furniture, no amount of trimming produces that geometry, because there was never evidence for it to begin with. Cleanup also can't fully separate a real repeating pattern, blinds, a patterned rug, a row of identical chairs, from a duplicate caused by ghosting, since both look like Gaussians repeated at a regular interval. That distinction still needs a person looking at the result, not an automated pass.
| Crop the volume | Trim by hand | Recapture | |
|---|---|---|---|
| Fixes | Sky floaters, out-of-bounds geometry | Isolated stray Gaussians | Missing or wrong source data |
| Speed | Seconds | Minutes | A new processing run |
| Undoes ghosting | No | Partially | Yes |
| Needs new photos | No | No | Yes |
If the sky matters for the final scene, expose for it on the capture pass rather than fixing it afterward. A slightly darker floor edits out in seconds; a missing sky does not come back.
Do you need special software to fix a gaussian splat scene?
The status quo for years has been a fragmented toolchain: install and configure a desktop pipeline such as Nerfstudio or a similar package, process the capture, then switch to a second tool to clean up what came out of it. All trademarks belong to their respective owners, and no affiliation, partnership or endorsement is implied. That workflow works, but it means a floater you noticed thirty seconds into a review costs a round trip through three separate pieces of software before you can even confirm it's gone.
A browser-based path skips that round trip. Captures Studio takes photo, video or 360 footage through the same cloud process either way, and the resulting scene opens directly in the Splat Editor for cropping and trimming, in formats like .spz and .ply, without a local GPU or an install. If you're still deciding how to shoot the capture in the first place, the interactive capture tutorial covers the overlap and exposure habits that mean less to trim afterward, and the plainer explanation of what a Gaussian splat actually stores is worth reading first if any of this is new.
FAQ
What causes floaters in a gaussian splat?
Floaters are stray Gaussians the optimizer placed with almost no supporting evidence, usually because a surface was seen from too few angles, had no matchable texture, or sat in an overexposed region like a blown-out sky. They cluster wherever the reconstruction had the least to work with.
Can a blown-out sky in a gaussian splat be fixed after capture?
Not by editing alone. An overexposed sky carries no texture or parallax for the optimizer to have used in the first place, so there is nothing left to recover. Cropping it out removes the floaters it causes; getting an actual sky back requires reshooting with correct exposure.
Why does my splat have ghosting or doubled geometry?
Ghosting happens when something changed between frames meant to show the same static surface, a moving door, a shifting shadow, a person walking through. The optimizer blends both states into a soft double image instead of picking one, since it has no way to know which frame was right.
Do I need special software to clean up a gaussian splat?
Some tool is required, since floaters live inside the scene data rather than on the surface. A browser-based splat editor removes the need for a separate desktop install; a traditional pipeline typically needs one tool to process the capture and another to clean it up.
Can floaters be removed automatically?
Partially. Cropping the working volume and filtering low-confidence Gaussians catches most of them without manual work. Ghosting and anything mixed closely into real geometry usually still needs a person to look at the result and decide what to cut.
How long does cleaning up a gaussian splat take?
Cropping and trimming a typical room capture takes minutes once the scene has finished processing. A full recapture and reprocess to fix a blown-out sky adds the same 10 to 120 minute cloud processing window the original capture went through.
Most of what gets written off as a bad capture is a crop and a trim pass away from being fine. The sky is the one part worth getting right before you leave the location, because it's the only failure on this list that a browser can't fix after the fact.