SplatWork
Make, train, view and edit 3D Gaussian splats on your own machines. Capture on an iPhone, prepare and view on a
Mac, and train on whichever of your machines are free: your Mac's GPU, other Macs, and Linux or Windows PCs with NVIDIA cards.
Big scenes are split into blocks that train on several devices at once and are merged back into one splat.
Everything runs locally on your network. No cloud account, no uploads to anyone else.
Website · Download · Quick start · User guide ·
Documentation · Changelog · Contributing
The apps
| App | Runs on | What it does |
|---|
| SplatWork | Mac (Apple silicon, macOS 15+) | The main app. Turns photos, video, 360° video, stereo footage, iPhone captures and COLMAP / Nerfstudio datasets into splats. Native Metal renderer and trainer, editor, exporter, and the coordinator for every other device. |
| SplatWork for Linux | Linux PC with an NVIDIA GPU (Flatpak) | A CUDA worker (gsplat) that the Mac and iPhone send training and processing jobs to, and a full desktop studio of its own: viewer, new projects, editing and export on Linux. |
| SplatWork for Windows | Windows 10 / 11 PC (x64) with an NVIDIA GPU | The same worker and studio as SplatWork for Linux, packaged for Windows (MSI installer). |
| SplatWork Capture | iPhone (iOS 18+; LiDAR models record depth) | Records ARKit-posed frames with LiDAR depth and sends them to the Mac. Can also turn its own scans into splats on the phone, or hand the heavy steps to your Mac or PC. |
| splatprep | Mac or Linux (Python) | The processing toolkit inside the apps: learned feature matching and camera solving (COLMAP / GLOMAP), moving-object and sky masks (SAM 2), video splitting and benchmarks. Also usable on its own from the command line. |
Highlights
- Any source. Photos (blurry and duplicate shots rejected), video (sharpest frames picked), 360° video and photos
(cut into flat views), stereo footage (side-by-side, top/bottom, two files, Apple spatial video; metric scale from
the baseline), old home video (VHS and camcorder tapes: split into clips and cleaned up), iPhone LiDAR captures,
COLMAP and Nerfstudio datasets.
- Camera solving that fits the footage. Apple Object Capture built in, or COLMAP / GLOMAP with SIFT,
SuperPoint, DISK or ALIKED + LightGlue, LoFTR and more. iPhone poses are refined rather than re-solved, drone photos
are checked against their GPS, and cameras placed far off the camera path are dropped.
- People, cars and sky handled. Masks for people and pets (Apple Vision), open-vocabulary moving objects with
SAM 2 tracking, a mask editor with click-to-select, and a learned skybox so distant sky doesn't turn into haze.
- A real trainer on the Mac. A differentiable Metal rasterizer with analytic gradients, 3DGS-MCMC and adaptive
densification, anti-aliasing, depth and surface-normal supervision, per-photo colour correction and held-out
evaluation. On Linux the same settings drive gsplat.
- Your devices work as one. Pick the devices for each project; each stage runs where it is fastest, and large
scenes train as blocks on all of them. Pause, resume, retry a failed block on another device, and keep watching
the live preview from anywhere.
- Nothing is lost when something stops. Each processing stage is checkpointed, training saves checkpoints and
every finished block, and worker jobs carry on while the Mac sleeps. The Projects list shows where each project
stands and resumes it from any stage.
- View, clean up and export. Fast Metal viewer with orbit and fly navigation and depth, ellipse, opacity and
coverage modes. Auto clean removes splats no photo needs. Crop, brush selection and undo. Export PLY, SPZ or
.splat, 4K screenshots and turntable videos.
- Easy or Advanced. Easy shows presets and the few choices that matter; Advanced shows every parameter. Every
setting has a built-in explanation, collected in the settings reference.
Download
Ready-built apps are on the Releases page.
SplatWork for Mac (Apple silicon, macOS 15 or later): download the .dmg and drag SplatWork to Applications.
The app isn't notarized yet, so the first time, right-click it and choose Open. It updates itself from new
releases (SplatWork → Check for Updates…). A rolling developer build with the latest changes is published as a
pre-release.
SplatWork for Linux for Linux (NVIDIA GPU, any distribution with Flatpak): install it from the project's Flatpak
repository, so updates arrive through flatpak update, GNOME Software or KDE Discover:
flatpak install --user https://benevision.github.io/SplatWork/flatpak/splatworker.flatpakref
Open SplatWork for Linux; its setup assistant detects your driver and installs a matching PyTorch and gsplat
(about 2.7 GB, all inside the app's own folder). Then turn the server on. Details:
linux/README.md.
SplatWork for Windows (Windows 10 / 11, x64; NVIDIA GPU for training): download the .msi from the next
release and run it. It asks once for administrator rights to allow Macs on Private networks to find the PC. Its setup
assistant installs PyTorch, CUDA and gsplat like the Linux app does, and it updates itself from new releases. Details
and the current state: windows/README.md.
SplatWork Capture for iPhone isn't on the App Store yet. Build it from source with Xcode: see
ios/README.md.
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
