ooolabdev/ooosplat

OOOSplat – One‑click 3D Gaussian Splatting on your desktop

What it does

• Takes a regular 360° video or a folder of sequential photos and turns it into a 3D Gaussian Splatting model (a point‑cloud‑like representation that can be rendered from any viewpoint).
• The whole pipeline runs locally – no cloud upload, no remote API. It bundles the required tools (FFmpeg, COLMAP, Brush) and orchestrates them from a graphical UI built with React + Tauri.

Why it matters

• Gaussian splatting is a recent neural‑graphics technique that produces high‑quality, real‑time renderable 3‑D scenes from images. It’s being adopted for VR/AR, game assets, and visual effects.
• OOOSplat makes this technology accessible to non‑programmers: you just pick a video/folder, choose a quality preset (Fast / Balanced / Fine), and the app does the heavy lifting – frame extraction, camera pose reconstruction, model training, and PLY export.

Key features

Feature	What you get
One‑click generation	Supports MP4, MOV, JPG/PNG sequences. The app extracts frames, runs COLMAP for structure‑from‑motion, trains a Gaussian splat model with Brush, and writes final.ply automatically.
Three quality presets	Fast (low resolution, ~8 k iterations), Balanced (mid‑resolution, ~15 k iterations), Fine (high‑resolution, up to 30 k iterations). The app auto‑adjusts frame rate, resolution and training settings based on your GPU memory.
Transparent‑material support	Handles alpha‑channel MOV/PNG, generating black‑and‑white masks for COLMAP so transparent objects are preserved.
Multi‑task manager	Run several projects side‑by‑side, see live progress, logs, hardware usage, pause/cancel, and resume from checkpoints.
Local MCP (Model Control Protocol)	An optional local AI‑agent interface (Codex‑style) that can create, start, query, and cancel tasks via a loop‑back HTTP endpoint.
Edit & preview	View the generated PLY, rotate/scale/translate, select and delete individual Gaussians, undo/redo, and export edited models.
Export options	Export vertical or horizontal MP4 videos, a single‑file offline HTML viewer (WebGPU/WebGL2), or the raw PLY files.
Cross‑platform	Pre‑built binaries for Windows 11 (x64), Apple‑Silicon macOS 15+ (Alpha), and Ubuntu 24.04 LTS (x86_64 Alpha).
Privacy‑first	All source media, intermediate data and models stay on your machine; only optional anonymous telemetry is sent (no media, no paths).

How it works – the pipeline

1. Input handling – Video is probed with FFprobe; FFmpeg extracts frames at the chosen FPS and resolution. Image folders are sorted automatically.
2. COLMAP reconstruction – Feature extraction & matching (GPU‑accelerated CUDA when an NVIDIA card is present, otherwise CPU). Incremental SfM builds camera poses and a sparse point cloud.
3. Brush training – The OOOBrush CLI consumes the COLMAP output and trains a Gaussian splat model on the GPU (Vulkan on Windows/Linux, Metal on Apple Silicon).
4. Validation & export – The resulting final.ply is verified and written to the project folder. The UI can then preview or further edit the model.

System requirements

• Windows: Windows 11 x64, WebView2 runtime, NVIDIA GPU with CUDA 13.2+ (or CPU fallback).
• macOS: Apple‑Silicon (M1/M2…) only; uses built‑in Metal backend, no CUDA.
• Linux: Ubuntu 24.04 LTS x86_64; requires Vulkan driver for GPU training, otherwise CPU only.
• GPU with at least 8 GB VRAM is recommended for the Fine preset; the app will automatically downgrade if memory is insufficient.

Typical workflow

Select video / image folder → Choose project folder → Pick quality preset → Start →
   [FFmpeg extracts frames] → [COLMAP builds camera poses] → [Brush trains Gaussian splats] →
   final.ply appears → Preview / edit → Export (MP4, HTML, or PLY)

Developer notes

• The UI is a React 19 front‑end rendered inside a Tauri 2 window; the backend is Rust (Tokio async).
• Engines (FFmpeg, COLMAP, Brush) are bundled and verified via npm run setup:engines; source code for the custom COLMAP fork and Brush lives in the same organization (ooolabdev).
• A diagnostic CLI (splatstudio) is provided for health checks, frame planning, and running the full pipeline from the command line.
• Build scripts handle Windows NSIS installers, macOS unsigned DMG, and Debian packages.

Who should use it

• Creators who want to generate 3‑D assets from handheld 360° footage without learning command‑line SfM tools.
• Researchers prototyping Gaussian‑splatting pipelines who need a reproducible, GUI‑driven workflow.
• Hobbyists interested in VR/AR content creation who value privacy (all data stays local).

---

Bottom line: OOOSplat is a polished, cross‑platform desktop application that packages the entire Gaussian‑splatting workflow—frame extraction, camera reconstruction, neural training, and export—into a single click, making cutting‑edge neural rendering accessible to anyone with a modern PC.