CorbeauSplat

CorbeauSplat is a free desktop application for macOS that puts a single graphical interface over a Gaussian splatting pipeline, from raw video or images through to a trained and viewable scene. It is written by Frederick de Witt, who states in the README that he is not a professional developer and built the tool with AI assistance to automate the workflow for a documentary film called Le Corbeau, which is where the name comes from.

The app orchestrates other people's tools rather than implementing its own. COLMAP handles feature extraction, matching and mapping, with GLOMAP available as an alternative mapper. Brush trains the splats locally on the Mac. upscayl-ncnn optionally upscales input images first, with six curated models verified against pinned checksums. SuperSplat runs in a built-in tab for local viewing and editing, listening only on 127.0.0.1. The PlayCanvas splat-transform CLI powers format conversion between PLY, SPZ, GLB and CSV, with spherical harmonic band reduction, decimation, NaN filtering, floater removal and Morton reordering. Apple ML Sharp generates a model from a single image or a video.

Version 2.0.0 arrived as the first stable release of the v2 line, following a beta series and a release candidate. The headline change in v2 is that the steps stopped being separate islands. A single Launch button in the Project panel runs the whole chain under a mode selector covering Gsplat, Sharp and 4DGS, with optional steps ticked in an Automation block, and the same chain is exposed from the command line as a pipeline subcommand.

Two modules are marked experimental. The 4DGS path prepares a dataset from a folder of synchronized camera videos into Nerfstudio format, and stops there, because no Apple Silicon 4D trainer exists to hand it to. The 360 extractor converts equirectangular video into planar image sets using cube map, ring or Fibonacci layouts, with AI masking of the operator.

Requirements are macOS with Apple Silicon recommended, Python 3.13 or 3.11, Xcode command line tools, Homebrew and Git. There is no PyPI package and the v2.0.0 release carries no binary assets, so installation is a git clone followed by a launcher script that detects and installs missing dependencies itself. The licence is MIT. The interface is localized across nine languages, with French as the source language, so several panel and step names render in French including Nettoyage, Export and Visualiser.

Radiance Fields covered the beta.4 release, which fixed the 4DGS COLMAP step six days before the stable build.