3D Gaussian Splatting on a Mac
The easiest way to do 3D Gaussian Splatting on a Mac is a native, all-in-one app such as RadianceKit: you import photos or a video, and it aligns, trains, edits, and exports the scene on your Apple Silicon GPU — no command line, no Python, no cloud. If you prefer a free, open-source route, you can also assemble a command-line pipeline. This guide compares every option and walks through the workflow.
What you need on a Mac
3D Gaussian Splatting is GPU-heavy, so you want an Apple Silicon Mac (M1 or later). Training runs on the GPU through Metal, which makes an Intel Mac impractical. 16 GB of unified memory is comfortable for typical scenes, and more helps with larger captures. You also need macOS recent enough for your tool — RadianceKit, for example, requires macOS 26 Tahoe or later.
Your input is ordinary photos or a video of an object or space, shot with enough overlap and steady lighting. Everything else — camera alignment, training, and export — is handled by the software.
The fastest route: a native all-in-one app
A native Mac app removes the hardest part of Gaussian Splatting: the setup. RadianceKit runs the whole pipeline locally on the Apple Silicon GPU. You drop in photos or a video, Apple Photogrammetry computes the camera positions, Gaussian Splatting training builds the scene, and you explore it in real time and export it — without installing COLMAP, Python, or any command-line tools, and without uploading anything to the cloud.
It offers a Simple Mode (import, press Start, get a scene) and an Expert Mode with a 3D viewport, training inspector, live loss curves, and an interactive editor to brush away floating artifacts. Export covers PLY, Compressed PLY, SPZ, glTF, .splat, and SOG, plus orbit videos and self-contained web viewers. This is the best fit if you want results quickly and want everything to stay on your machine.
Free and open-source pipelines
If you are comfortable with the command line and want a zero-cost route, you can build a pipeline from open-source parts. COLMAP (or PyCOLMAP) reconstructs camera poses from your images, and a Metal-capable trainer such as OpenSplat or Brush performs the Gaussian Splatting on Apple Silicon. This is flexible and free, but it means installing dependencies, managing Python or build toolchains, and stitching the steps together yourself.
Choose this path for research, custom workflows, or integration with your own code — not for the quickest result from a handful of holiday photos.
Viewers for existing splats
If you already have a .ply or .splat file and only need to view it, a dedicated viewer such as MetalSplatter renders Gaussian Splats with Metal on macOS, iOS, and visionOS. Viewers do not train scenes — they display ones you produced elsewhere.
The workflow, step by step
- 1 Capture — Shoot 30–200 overlapping photos, or a slow video, of your subject from many angles with consistent lighting.
- 2 Import — Load the photos or video into the app (or into COLMAP for the open-source route).
- 3 Align — Let the software compute camera positions — Apple Photogrammetry in RadianceKit, COLMAP in the open-source pipeline.
- 4 Train — Run Gaussian Splatting training on the Apple Silicon GPU to build millions of 3D splats.
- 5 Edit & export — Clean up stray splats, then export to PLY, SPZ, glTF, .splat, SOG, an orbit video, or a shareable web viewer.
Mac Gaussian Splatting tools at a glance
| Tool | Type | Best for |
|---|---|---|
| RadianceKit | Native Mac app | Fastest local, no-setup workflow |
| OpenSplat | Open-source CLI | Free, cross-platform, scriptable |
| Brush + COLMAP | Open-source CLI | Free DIY pipeline on Apple Silicon |
| MetalSplatter | Viewer | Viewing existing .ply / .splat scenes |
Bottom line
For most people on a Mac, a native app like RadianceKit is the fastest way to go from photos to a finished 3D Gaussian Splat — locally, privately, and without touching a terminal. If you need a free, open-ended setup and do not mind the command line, an OpenSplat or Brush + COLMAP pipeline is the alternative.
Frequently asked questions
What's the difference between Gaussian Splatting and NeRF?
Both turn photos into a 3D scene, but Gaussian Splatting represents it as millions of small 3D “splats” that render in real time, while a NeRF stores the scene in a neural network that is slower to display. In practice Gaussian Splatting trains faster and views in real time, which is why RadianceKit uses it, and it tends to look sharper on everyday captures.
Is Gaussian Splatting better than traditional photogrammetry?
They solve different problems. Traditional photogrammetry builds a textured mesh that is good for measurements and editing in 3D tools. Gaussian Splatting reconstructs the look of a scene — reflections, fine detail, soft edges — for photorealistic real-time viewing. For a lifelike 3D capture to share, splatting usually looks better; for a measurable, editable model, a mesh is still the right tool.
How many photos do I need for a good result?
For a single object, 30–100 sharp, overlapping photos from many angles usually give a good result; rooms and larger scenes benefit from more. Consistent lighting and enough overlap matter more than sheer count. You can also shoot a slow video and let RadianceKit sample the frames.