toruhashimoto/3dgs-floater-cleaner

3DGS Floater Cleaner is a Windows desktop tool for cleaning up 3D Gaussian Splatting reconstructions. 3D Gaussian Splatting, often shortened to 3DGS, is a technique for turning a set of photos into a 3D scene made of many small blob shaped points. A common problem with these reconstructions is floaters, stray points that hang in midair and only become visible when you view the scene from certain angles. This tool takes photo data that has already been aligned with the RealityScan app and exported in COLMAP format, then trains a 3DGS scene with far fewer floaters, producing a cleaned up .ply output file with one click. The key finding behind the tool, described in its included FINDINGS document, is that cleaning up the point cloud before training does not actually reduce floaters, because the training process rebuilds background structure regardless of the starting points. What does work is a geometry regularizer called scale_reg, applied during training itself, which shrinks or removes floating points while leaving image quality mostly unchanged. The tool offers five detail level presets, from lower detail and faster processing up to a high detail default setting, and lets the user choose how strongly scale_reg is applied. After training finishes, the tool automatically counts remaining floaters and can open the result directly in SuperSplat, a separate viewer and editor, for further editing or export. An optional before and after comparison mode can show the floater reduction against an unprocessed baseline, though this roughly doubles processing time. Running the tool requires a Windows computer with an NVIDIA GPU, the LichtFeld Studio training engine which the tool can auto detect, and a standard Python 3 installation. The project bundles its own launcher script that sets up a virtual environment on first run, so no manual Python package installation is required to get started.