How to Make a Gaussian Splat on a Mac Without a Graphics Card
The original Gaussian splat trainer needs an NVIDIA card, and no Mac has one. Here is why that matters, what your options are, and how to make a Gaussian splat on a Mac with a browser and a cloud queue.
You can make a Gaussian splat on a Mac without a graphics card, because the heavy part of the job no longer has to run on your Mac. A Gaussian splat is a 3D scene built from millions of tiny, soft-edged blobs of colour, trained from ordinary photos or video. Training is the slow step, and it is the step that historically wanted an NVIDIA card.
Macs do not have one. Apple Silicon ships with its own GPU, driven through Metal, and that is a different world from the CUDA software most splat tools were written for. The fix is to move training off the laptop and keep everything else in the browser.
Why can't a Mac run most Gaussian splat trainers?
The original research code for 3D Gaussian Splatting is built on PyTorch and custom CUDA extensions. Its README lists a CUDA-ready GPU with Compute Capability 7.0 or higher, and 24 GB of VRAM to train to paper evaluation quality. It states the code was tested on Windows 10 and Ubuntu Linux 22.04. macOS is not on that list.
CUDA belongs to NVIDIA. No Mac, Intel or Apple Silicon, ships with an NVIDIA GPU, so code that calls CUDA directly has nowhere to run. Porting it is a research project, not a setting.
The popular desktop apps inherit the same limit. Postshot, a widely used splat trainer, lists Windows 10 or later and an Nvidia GPU (GeForce RTX 2060, Quadro T400, or RTX 4000 and above) as its requirements on the Jawset site. A MacBook Pro is a fine machine for editing photos. It simply isn't the machine those programs ask for.
What are your options for gaussian splatting on Mac?
Three routes exist, and they trade effort for control.
| Route | What you need | Where training runs | Catch |
|---|---|---|---|
| Captures Studio in the browser | A Mac, a browser, photos or video | Cloud, typically 10 to 120 minutes | You depend on the queue, not your hardware |
| Native Metal or WebGPU trainer such as Brush | A Mac and some patience | Your own machine | Fewer features than the CUDA research code, and you manage the files yourself |
| Rent a Windows or Linux machine with an NVIDIA card | A cloud account and a terminal | A remote machine you configure | Setup, installs, and a remote desktop to babysit |
Brush deserves a fair mention. It is an open-source engine built on WebGPU rather than CUDA, and its README says it runs on macOS, Windows and Linux, on AMD, Nvidia and Intel cards, on Android and in a browser. Its browser demo is limited to Chromium browsers, and the WebGPU build lists Chrome 134 or newer on Windows and macOS. If you like tinkering and want training to stay on your own Mac, it is a real option.
For most photographers and designers, the browser route is shorter. Captures Studio processes in the cloud, so the Mac does not train anything. It takes photos, ordinary video (MP4, MOV), 360 footage from cameras such as the Insta360 or Ricoh Theta, and drone footage, and returns an interactive scene. No local GPU. No installs.
How do you capture a Gaussian splat on a Mac?
Capture decides quality more than any hardware does. Use whatever you already own: a phone, a mirrorless camera, a drone.
Photos or video
Walk slowly around the subject and keep it in frame. Overlap is the rule: each frame should share most of its content with the one before it. Video is the quicker path, and the video to 3D page covers how a clip becomes a scene. If you are shooting stills, photo to 3D covers the same ground for photographs.
Lock your settings
Fix exposure and white balance before you start. A scene whose brightness drifts between frames trains into a scene with cloudy patches. Shoot at a shutter speed fast enough to avoid motion blur, because blur reads as missing detail.
Move the files to your Mac
AirDrop or a cable works. Nothing is installed at this stage and nothing needs converting. Upload the files in the browser and the cloud does the rest.
Because training runs in the cloud, you can upload a capture and close the lid. Come back when the scene is ready rather than watching a progress bar on a machine that is not doing the work.
How does cloud processing replace the GPU?
Training a splat starts from a sparse point cloud recovered from your photos, then repeatedly adjusts millions of blobs until renders match the originals. That loop is GPU-bound, which is why NVIDIA hardware dominates it. In a cloud workflow, the same loop runs on a server GPU, and your Mac only uploads inputs and receives the result.
On Captures Studio this takes 10 to 120 minutes, depending on the capture. When it finishes, the scene opens in the browser. The viewer reads .spz, .ply, .splat and .ksplat, so you are not locked to one format.
What can a Mac do once the splat exists?
Quite a lot, because everything after training is light work. Viewing and editing run in the browser, with no render farm on your desk.
- Trim stray floaters and tidy the edges in the Splat Editor.
- Set guided camera paths and add hotspots carrying text, images, video or links.
- Scale the scene to real-world size from a reference object of known length, then measure distances in 3D.
- Compare two captures of the same place with Multi Splat, which aligns them and can render a 4D timelapse to MP4.
Sharing is one link, an embed or a QR code. It opens on a phone, tablet, desktop or VR headset without an app.
Where does this workflow fall down?
Cloud processing has real limits, and a Mac does not remove them.
- You need a connection. Uploading a 4K video over weak wifi is slow. Compress sensibly, or upload from a wired connection.
- Capture mistakes follow you. Mirrors, glass and chrome confuse reconstruction. Moving people and cars leave ghosts. A bad capture trains into a bad splat on any hardware.
- You get less low-level control. If you want to change training parameters yourself, a local tool such as Brush or a rented NVIDIA machine gives you the dials. A managed pipeline does not.
- Processing is a queue, not instant. Ten minutes and two hours are both normal. Plan around it.
FAQ
Can you do Gaussian splatting on a Mac?
Yes. Native tools such as Brush train on macOS using WebGPU, and cloud services process on remote GPUs while you work in a browser. What a Mac cannot do is run trainers that need CUDA, the NVIDIA-only toolkit used by the original reference code.
Does Apple Silicon work for Gaussian splatting?
Apple Silicon works for capture, viewing, editing and sharing. For training, it works with tools built on Metal or WebGPU. It does not run CUDA-based trainers, so most popular desktop apps are off the table unless training happens somewhere else.
Do I need an NVIDIA GPU to make a Gaussian splat?
Only if you train locally with CUDA software. The reference 3DGS code lists a CUDA-ready GPU, and Postshot lists an Nvidia card. If processing happens in the cloud, the NVIDIA hardware is someone else's and your computer needs only a browser.
Can I make a Gaussian splat from iPhone video on a Mac?
Yes. Record a slow, overlapping walk around the subject, AirDrop the clip to your Mac, and upload it in the browser. Captures Studio accepts MP4 and MOV video, so the file needs no conversion first.
How long does cloud splat processing take?
On Captures Studio, processing runs from 10 to 120 minutes. Short, well-lit captures finish near the low end. Long walkthroughs and large photo sets sit toward the high end.
Closing
A Gaussian splat on a Mac stopped being a hardware question once training moved off the laptop. The Mac was never the problem. The training step was tied to a specific vendor's toolkit, and moving that step to a server turns the laptop back into what it is good at: handling files, shaping a scene, and sending a link.
Capture slowly, keep your exposure fixed, and let the cloud do the arithmetic. What lands in your browser afterwards is a scene you can walk through, edit and send, from a machine with no graphics card to speak of.
All trademarks belong to their respective owners. No affiliation, partnership or endorsement is implied.