Skip to main content
Back

OOOSplat Turns Raw Footage Into 3D Scenes Without Uploading Anything

OOOSplat is a one-click desktop app that converts phone videos and photos into 3D Gaussian Splats entirely on your own machine.

Repo, Full Access
  • OOOSplat converts videos or image folders into 3D Gaussian Splats locally in one click
  • Bundles FFmpeg, COLMAP, and Brush inside a Tauri desktop app written in Rust and React
  • Runs on Windows, Apple Silicon macOS Alpha, and Ubuntu 24.04 Alpha under Apache-2.0
  • Three quality presets (Fast, Balanced, Fine) control frame rate, resolution, and training iterations
  • Exports to PLY, compressed PLY, SPZ v4, SOG, MP4, and single file offline HTML
  • Preview workspace supports WebGPU editing, cropping, point deletion, WASD walk, and panoramas

OOOSplat 0.7.1 packages video processing, camera reconstruction, Gaussian Splatting training, editing, and export into one local desktop application. It turns overlapping footage or photographs into a navigable 3D scene represented by colored Gaussian particles, replacing a workflow that typically requires several tools and custom command-line scripts. The repository has 2.2k GitHub stars.

The Apache-2.0 licensed application runs on Windows, Apple silicon macOS, and Ubuntu 24.04 LTS x86_64. Its React interface invokes a native Rust backend through Tauri's inter-process communication layer, so generation and preview require no local HTTP service. Users select a video or image folder, project directory, and quality preset; the application handles the remaining stages without sending source media to a hosted reconstruction service.

One app handles five pipeline stages

OOOSplat bundles FFmpeg, FFprobe, COLMAP, and Brush, then coordinates their inputs and outputs:

  1. FFprobe inspects the source video, and FFmpeg extracts frames using a target frame rate or manual sampling ratio.
  2. COLMAP detects matching image features, estimates camera positions, and reconstructs sparse scene geometry. It selects CPU or CUDA processing automatically.
  3. For panoramic input, the application preserves the full equirectangular projection during spherical reconstruction. It converts registered views into six cube faces for training.