Gaussian splatting turns ordinary photos into a 3D scene you can walk through. Until now it meant uploading to someone's cloud or buying a tower with an NVIDIA card. SplatCast trains it on the Apple silicon you already own, natively in Metal, and your photos never leave the device.
SplatCast render
Photograph
The research is three years old and the results are stunning. The tools to make them have not caught up.
01 · Mobile
Most capture apps send your photos to a server, train there, and queue you behind everyone else. Apps that train on the phone itself keep scenes small.
SplatCast trains on iPad and Mac, offline, at full quality. No upload, no queue, nothing stored on someone else's servers.
02 · Scale
Desktop training lives in graphics-card memory: 24 GB on an RTX 4090. Large sites run out partway through, and the usual fix is to throw detail away.
SplatCast uses Apple silicon's unified memory, up to 512 GB on a Mac Studio. It finished a scene the 4090 could not.
03 · Mac
The reference trainer is written in CUDA, and the leading desktop app needs an NVIDIA card. On a Mac you were left with ports and command lines.
SplatCast is written for Metal from the first line: a drop-in app, about twice as fast as the 4090 reference on a Mac Studio.
A folder of photos, a video, or a COLMAP reconstruction.
Training starts by itself. Watch the scene sharpen and stop when it looks right.
Move through the scene in a native Metal viewer, then export PLY, SOG or SPZ.
SplatCast is a native app for Apple silicon. Train on a Mac Studio at your desk or a MacBook Pro on the move, and open the result on iPad.
Studio Display with Mac Studio, MacBook Pro and iPad, drawn around real captures of the app: the Riverview Room in the viewer, and the Clock Tower training. The iPad shows a frame rendered from the Clock Tower scene.
The reference for this kind of training is the original CUDA code on an NVIDIA RTX 4090. We trained the same photos with both, to the same 30,000 steps.
On a Mac Studio, SplatCast finished both benchmark scenes in a little over half the time the RTX 4090 took, at close to the same quality.
Bicycle, 194 photos
Rubble, 1,678 drone photos
Ziegeleipark: 512 drone photos at 1600 × 1200 and the same camera solve on both machines. The CUDA trainer ran out of graphics memory three times. SplatCast on the Mac Studio finished.
Steps completed, out of 30,000
Try 3 gave the 4090 expandable memory segments. The finished scene has 15.2 million splats.
The MacBook Pro beat the RTX 4090 on Bicycle as well. Training a scene no longer means sitting at a tower.
Mac Studio with M2 Ultra and 64 GB; MacBook Pro 14-inch with M4 Pro; RTX 4090 with 24 GB in a Threadripper workstation running the reference GraphDECO trainer. Times run from launch through the saved scene and leave out camera solving. RTX 4090 runs from September 2026; SplatCast runs from October 2026, the Mac Studio on the same prepared photos, the MacBook Pro on the original Bicycle photos resized to the same 1600 × 1063. Quality, scored with each project's own renderer: Bicycle 25.09 against 25.16 dB, Rubble 26.24 against 26.46 dB. Other CUDA trainers and memory offloading were not tested. Weights from Apple and NVIDIA specifications; the tower is drawn at a typical full-size case. Bicycle and Rubble are the Mip-NeRF 360 and Mill 19 research scenes; Ziegeleipark photos are CC0.
Each bar is one full training run of 30,000 steps. Quality is measured on photos held back from training, in decibels: higher is closer to the photograph.
Bonsai and Bicycle are the Mip-NeRF 360 research benchmarks. Clock Tower photos are 2.8 megapixels and Riverview Room photos 3.3, two to three times the size of the benchmarks.
Open one and look around before you train your own, here in the browser or in the app. Each was trained in SplatCast from openly licensed photos.
219 drone photos from Heritage3DMtl by Rucha Shende, Sudhir Mudur, Tiberiu Popa and Kaustubha Mendhurwar, Concordia University, CC BY 4.0. Undistorted and cropped.
16.6 MB in the browser
243 drone photos from Heritage3DMtl by Rucha Shende, Sudhir Mudur, Tiberiu Popa and Kaustubha Mendhurwar, Concordia University, CC BY 4.0. Camera positions rebuilt; undistorted.
18.8 MB in the browser
1,056 photos. From Eyeful Tower, © Meta Platforms, Inc. and affiliates, MIT licence.
17.9 MB in the browser
247 drone photos from Heritage3DMtl by Rucha Shende, Sudhir Mudur, Tiberiu Popa and Kaustubha Mendhurwar, Concordia University, CC BY 4.0. Camera positions rebuilt; undistorted; reduced to 1.2 million splats.
18.5 MB in the browser
3,804 rig photos. From Eyeful Tower, © Meta Platforms, Inc. and affiliates, MIT licence. Reduced to 1.2 million splats.
17.8 MB in the browser
In the browser, Riverview Room, City Hall and the apartment are reduced to 1.2 million splats so they load on a phone. The app ships every scene in full. The viewer is PlayCanvas's open-source SuperSplat viewer.
Scenes are processed and stored on this Mac. Nothing is uploaded.
Train as much as you like on every plan. Pay when you want to take your scenes elsewhere.
For trying it on your own photos.
$0
Recommended
For creators who share and sell what they capture.
$5.99 a month
or $59.99 a year, about $5.00 a month (17% less)
For studios and companies.
$14.99 a month
or $149.99 a year, about $12.50 a month (17% less)
| Feature | Free | Pro | Studio |
|---|---|---|---|
| Training and viewing | Unlimited | Unlimited | Unlimited |
| Viewer | Watermarked | Clean | Clean |
| Export as PLY, SOG or SPZ | – | ||
| Training resolution | Up to 1,600 px | Up to 1,600 px | Full resolution |
| Commercial use | – | Individuals | Companies |
| New training features first | – | – |
Prices in US dollars. Subscriptions renew until cancelled and can be cancelled any time from your account page or, for App Store purchases, your Apple Account.
Needs a Mac with Apple silicon and macOS 14 or later.
Download for Mac[Download link to be added at release]