Open Tools for Gaussian Splats
Reconstruction, cleanup, conversion, research, and publishing
Gaussian splats, 3DGS, SplatEdit, Brush, COLMAP, SuperSplat, Nerfstudio, gsplat, LingBot-Map, Apple Silicon

A toolchain rather than one application
Producing a Gaussian splat normally involves several distinct jobs. One tool prepares camera positions, another trains the 3DGS scene, and another cleans or publishes the result.
This directory concentrates on open-source, free, or locally usable tools. Commercial capture services are intentionally not included.
| Job | Recommended starting point | Result |
|---|---|---|
| Prepare photographs and camera poses on macOS | COLMAP Gradio | COLMAP dataset |
| Train 3DGS on Apple Silicon | Brush | Trained Gaussian splat |
| Edit and extend a splat | SplatEdit | Cleaned or transformed splat |
| Clean and publish in a browser | SuperSplat | Web scene or optimized file |
| Research and custom training | Nerfstudio/Splatfacto and gsplat | Experimental 3DGS pipeline |
| Streaming geometric reconstruction | LingBot-Map | Poses, depth, and point-cloud reconstruction |
For complete installation and processing instructions, continue with Making 3DGS on Apple Silicon.
Reconstruction on macOS
COLMAP Gradio Interface
COLMAP Gradio places a simple browser interface in front of COLMAP. It was made specifically to prepare data for 3DGS on macOS without requiring the user to construct every COLMAP command manually.
The interface can:
- resize an image set while preserving the originals;
- extract image features;
- run exhaustive, sequential, or spatial feature matching;
- solve camera positions and a sparse point cloud;
- produce the COLMAP files needed by a compatible trainer.
The expected workspace contains an images directory. Successful processing creates database.db and a sparse/0 directory containing cameras.bin, images.bin, and points3D.bin.
Role in the pipeline: camera solving and dataset preparation. It does not train the Gaussian splat itself.
Brush
Brush is an open-source Gaussian splat reconstruction engine built with WebGPU-compatible technology and the Burn machine-learning framework. It supports macOS, Windows, Linux, several GPU families, Android, and browser execution instead of depending exclusively on NVIDIA CUDA.
Brush accepts COLMAP data or a Nerfstudio-format dataset. During training, the user can navigate the scene, observe the reconstruction developing, and compare the current rendering with its input views. Brush also provides a CLI, a viewer for PLY and compressed PLY splats, and support for image masks.
Role in the Mac pipeline: train and inspect the 3DGS scene using Apple Silicon GPU resources.
Use COLMAP Gradio to solve the photographs and Brush to train the Gaussian splat. This separates the well-established structure-from-motion stage from the WebGPU-compatible 3DGS stage.
Editing and cleanup
SplatEdit
SplatEdit is a free browser-based Gaussian-splat editor built upon SuperSplat and extended with additional creation, selection, conversion, and spatial tools.
Alongside familiar transform, lasso, polygon, brush, flood, and attribute-based selection, SplatEdit provides operations such as:
- densifying selected areas;
- dividing a scene into a grid;
- creating proxy geometry from a splat;
- projecting and color-correcting imagery;
- grounding and georeferencing a scene;
- selecting by transparency, density, size, brightness, color, or spherical harmonic data;
- exporting PLY, SPLAT, KSPLAT, SOG, Streamed SOG, SPZ, GLB proxies, viewer packages, images, and video.
Because it runs in the browser, it is immediately usable on a Mac without a CUDA installation. Large scenes still depend on browser and GPU memory.
Best for: detailed cleanup, segmentation, scene division, proxy creation, conversion, and preparing a splat for another application.
SuperSplat
SuperSplat Editor is an open-source, browser-based editor from PlayCanvas. It can crop scenes, remove floaters, transform and combine splats, adjust appearance, create camera animation, and export several delivery formats.
The editor works locally until the user chooses to publish. Its standalone HTML and viewer-package exports are useful for self-hosting, while the hosted workflow can compress large scenes into streamed levels of detail.
Best for: fast cleanup, optimization, camera animation, browser publishing, and self-hosted viewing.
Research and programmable pipelines
Nerfstudio Splatfacto
Nerfstudio’s Splatfacto is an open 3DGS implementation intended for experimentation and reproducible technical pipelines. It works with COLMAP-derived data and uses gsplat as its rasterization backend.
The standard ecosystem is most straightforward on CUDA-capable hardware. Therefore, it belongs in this general tools directory but is not the primary recommendation for a native Apple Silicon workflow.
gsplat
gsplat is a CUDA-accelerated library for Gaussian rasterization and optimization. It is a development component rather than a finished capture application.
Best for: researchers and developers implementing or modifying 3DGS training methods on suitable hardware.
LingBot-Map
LingBot-Map is a feed-forward model for streaming 3D reconstruction. It combines coordinate grounding, dense geometric cues, trajectory memory, and long-sequence processing. Its demo can take an image folder or video and present the reconstructed geometry and camera path in a browser viewer.
LingBot-Map is relevant because it addresses long, streaming captures and camera drift differently from traditional iterative reconstruction. However, it should not be confused with a conventional 3D Gaussian Splatting trainer: its documented output and visualization pipeline center on depth, poses, and point-cloud reconstruction.
Its documented installation uses PyTorch with CUDA 12.8, FlashInfer, and—when using the offline renderer—NVIDIA Kaolin and custom CUDA extensions. It is therefore not part of the Apple Silicon workflow described on this site.
Best for: research into long-sequence, feed-forward, and streaming 3D reconstruction on an NVIDIA system.
Viewing and delivery
Browser
Brush can view PLY and compressed PLY splats, while SuperSplat and SplatEdit can export or publish browser-oriented packages. Browser delivery is convenient but should use an optimized format rather than an unnecessarily large archival PLY.
Apple Vision Pro
RealityKit on recent visionOS versions provides native Gaussian-splat APIs. Apple’s Gaussian splats on visionOS documentation demonstrates the spatial presentation stage. See also the USDZ concept page.
Recommended open workflow
- Capture a static, overlapping image sequence.
- Create the COLMAP dataset with COLMAP Gradio.
- Train and inspect the scene in Brush on Apple Silicon.
- Export a full-quality master.
- Clean and orient the scene in SplatEdit or SuperSplat.
- Produce an optimized copy for browser or spatial delivery.
- Preserve the source photographs, COLMAP solve, and full-quality splat.
This modular workflow keeps the irreplaceable capture and reconstruction data separate from smaller files made for a particular viewer.