What Is SuperSplat, and How Do You Clean Up a Gaussian Splat?
Part 11 of our Gaussian Splatting Series the search-led one: "supersplat" is a breakout query on YouTube, and "how do I clean up a Gaussian splat" is the question behind it.
1. Why clean up a splat at all?
When we use the Lixel K2 or any scanner, for that matter to record a Gaussian splat, we record the environment around us as best we can. That is a wonderful opportunity to capture reality. But the design process within that reality often requires us to remove things from the scanned environment.
Take a renovation. The client asks us to oversee the redesign of an interior, and we scan a room or a large interior in which everything currently placed in the space is about to be removed. The only things that will remain are the walls, the ceiling, and the floor. Everything else in the scan is history the moment the design process starts.
So the splat needs cleaning. In some software that's an AI process click-to-remove-furniture, as you can see in tools like Matterport. In the Gaussian splat workflow it's a manual one: in Lixel CyberColor (LCC) Studio you select a 3D eraser box a cube you stretch, pull, and move to cover the items you don't want to see in the scan and cut them out.
2. What do you remove, and why?
Three categories, in practice:
The contents of the space: furniture, loose objects, anything the renovation will remove. We clear it to make room for the new design.
Light artifacts and blur: reflections, ceiling blur, and any smeared data the capture left behind.
Out-of-range data, anything the scanner picked up at the perimeter, far beyond the area of interest.
That last one deserves explanation. Anything we scan at its most exterior edge still carries some Gaussian splat information there is still light-sensing data appearing at the perimeter of the scan. The K2 can scan up to 40–50 metres, and it will happily return scannable data 50 metres away from you: the street outside the window, the neighbour's roof. That data has no context value for the design. It just costs file size and visual noise.
One important warning before you start cutting: the LCC software is not using AI to fill in the space behind what you remove. If we splat a hallway with a beautiful tiling pattern, and there are objects standing on that floor, deleting those objects does not reveal a repaired floor underneath. It reveals a gaping hole in the scan the scanner never saw that patch of tile, so there is nothing there. And because the eraser box is a 3D volume, it's easy to accidentally cut through the floor itself and suddenly see through to the void beyond the scannable area. Deciding what to remove, and why, is of the utmost importance it's the first act of the design process, not housekeeping.
3. Capture well and you'll clean less
Most cleanup problems are capture problems. Two settings matter.
First, processing quality. LCC Studio processes a scan at slow, normal, or fast. By giving the computer more time, we get more density, more accuracy, more reliable depth.
Second and more important, scanning speed. Move the scanner through the space at below one metre per second. The camera on these devices is a digital shutter rather than a mechanical one, and that reduced walking speed gives it enough time to collect clean frames without motion blur. Remember what you're actually doing: moving through a space while trying to capture static data. Any blur you introduce while walking becomes blur baked into the splat which becomes cleanup work later, or worse, data you can't recover.
4. Destructive versus non-destructive edits: our rule
When we scan, we deliberately gather data three to five metres past the area we need to design. That margin gives us the view down the corridors and enough context to create renders that sit inside a believable environment rather than floating in a void.
Everything beyond that margin gets destructively removed in LCC Studio, before export. This is the edit you never need to undo, and it greatly reduces the file size headed for SketchUp.
The numbers make the case. The scanner captures at roughly a gigabyte a minute, so raw project data runs into tens of gigabytes. By the time the LiDAR and photogrammetry data are assembled into a Gaussian splat, that 20 GB might become 200 MB. The file we import into SketchUp usually lands between 100 and 300 megabytes occasionally 800 MB to a gigabyte and a 200–300 MB splat is a relatively light file to carry as a reference inside a SketchUp model. Remember, at that point it's not a mesh. It's not thousands of edges combined into shapes. It is simply a Gaussian splat: oval light shapes, reflecting light from different angles and viewpoints.
Everything inside the margin the actual design context we keep, and we edit it non-destructively once it's in SketchUp. Cut, hide, unhide: decisions we're allowed to change our minds about, all the way through the design process.
2. What is SuperSplat?
SuperSplat is a free, open-source Gaussian splat editor from the PlayCanvas team that runs entirely in the browser. Type superspl.at into your address bar and the site opens on the community gallery and the ability to explore how other people are using the software is fantastic. Apartments, churches, a car left in the woods, highly zoomed sections of cloth, high-detail garages, Mont-Saint-Michel in France, a fountain, a lady seated on the ground. It's the fastest way to calibrate your expectations of what a good splat looks like.
On the left-hand side, under the logo, the sidebar gives you two buttons: Editor and Convert. Click Editor and you're presented with a 3D environment not so different from Unreal Editor or Blender. Drag and drop a PLY file onto it or go File → Import — and your splat from LCC, or an SPZ file, loads into the environment. (You may get a note suggesting you upload the original data first; we haven't hit issues importing directly.)
Once the data is in, you're given tools similar to LCC Studio selection brushes, selection area tools, sphere selects with which you can protect the areas you want to keep, and cut, hide, or delete the rest. Crucially, SuperSplat lets you do genuinely destructive editing here, which is not immediately possible in the CyberColor environment once you've exported for a pure live viewer. You can also transform and scale the splat itself, set up a scene with specific cameras, and render out an image or a video walkthrough along a camera path.
What does SuperSplat cost?
The editor itself: nothing. SuperSplat is free and open source under the MIT license you can support it, file feedback, and follow the roadmap on GitHub and Discord, and read the documentation and blog for detail.
A free account lets you publish splats and keeps a small library of your content at no cost. Beyond that there are paid plans personal at around $15 a month and organisational at around $50 which add more cloud storage and hosting, unlimited private projects, archive and restore, REST API access, a customisable loading screen for your own clients, a downloadable app for self-hosting, and team management so your whole organisation works from the same environment.
Why we still prefer cut, hide, and unhide inside SketchUp
So we've arrived in SketchUp. The .lcc2 or PLY file is imported, and we have an environment we can draw over. From here, our preference is firm: bring the splat in largely intact, and use the non-destructive tools, the scissor icon in the Gaussian Splat toolbar to cut and hide sections, plus SketchUp's native machinery of tags, hidden objects, groups and components, and the famous section cut with its corresponding styles.
The reason is the whole point of scanning in the first place: we want everything we captured available as context throughout the design process. At presentation time we might cut and hide elements of the splat so attention goes to the design itself rather than to a distracting corner of the captured environment. We're not trying to obscure the truth of the site, we want the freedom to cut, hide, and unhide any section of the splat at any point, and then bring it back.
Destructive edits belong upstream, where they reduce file size. Non-destructive edits belong in SketchUp, where the design decisions happen. Get that order right and you end up with a composition of scanned reality with your new design placed inside it the foreseeable future, standing in the captured present. That is what makes Gaussian splatting so beautiful as a way of relating ideas to clients.
Together, SuperSplat and LCC form a nice bridge between formats: a web-based editor and a desktop-based one, both feeding the SketchUp environment where the drawing happens.
Questions about our Gaussian splatting series, the blogs, or the YouTube videos, or want to work with us on a design for your retail space? Leave a comment, or reach us at info@studiojdb.com. Thanks for taking the time to read, and have a wonderful week.
