OOOSplat turns smartphone video into a 3D scene locally, in one click

Gaussian Splatting has long since moved beyond research demos: scenes look photorealistic and can be explored in a browser in real time. But getting from phone footage to a finished splat still feels like an obstacle course. You need to extract frames with FFmpeg, calculate camera positions in COLMAP, run training, then find something to view and crop the result. The OOOLab team has packaged that entire chain into a single application and released it as an open-source project called OOOSplat.

What's inside

OOOSplat is a cross-platform desktop application for Windows, macOS on Apple Silicon and Linux. It is built with Rust and Tauri, with a React interface. The central idea is self-sufficiency: the package already includes preconfigured binaries for FFmpeg, FFprobe, COLMAP and the Brush training engine. Nothing needs to be installed separately, and nothing is sent to anyone else's servers.

The pipeline starts with a single button. The application takes a video (MP4, MOV), extracts key frames, calculates camera positions through COLMAP and optimizes the Gaussians in Brush. The output is a final.ply file. If the machine has an NVIDIA graphics card, COLMAP uses it automatically; otherwise, it runs on the CPU, just more slowly.

What it looks like in practice

Imagine you are building a service for a real estate agency or a product catalog with 3D previews. Previously, an employee would film a room on their phone, and everything after that depended on an engineer who knew how to run COLMAP. Now the footage can be dropped into the application to produce a scene without touching the command line.

After training, the result opens in a built-in 3D viewport using WebGPU and WebGL. You can move and rotate the scene and use a bounding box to crop out the floating debris around the edges, a common problem with splats. Export options include PLY, SOG, SPZ and SPLAT, as well as a standalone HTML page with a lightweight web player. That offers a straightforward use case: upload a single file to a hosting service and send the client a link.

There are some useful extras, too. The app supports 360×180 panoramas in ERP format. GPS data from EXIF is used to calibrate the metric scale, giving the scene real-world dimensions. Transparent alpha masks are supported, while re-shoot mode lets you add new frames and update a scene without retraining from scratch.

Limitations and how to approach them

A few caveats up front. This is a new open-source project, and there is still little feedback on its stability or the quality of its results with complex scenes. Splat quality itself depends on the footage: sharp frames, good overlap and no moving objects. No wrapper can change those requirements. The CPU fallback is better suited to testing an idea than to regular use, and training speed on Apple Silicon and Linux should be measured with your own data. You should also check the licenses of the bundled components in advance if you plan to integrate the output into a commercial product.

The practical implications vary by role. For developers, the repository is a useful reference for bundling COLMAP, Brush and FFmpeg into a single Tauri package, as well as a ready-made foundation for a fork tailored to their needs. For product managers, testing ideas involving 3D content has become cheaper: a pilot needs neither a cloud budget nor a dedicated photogrammetry specialist. For founders working with private footage of properties, production facilities or interiors, local processing is valuable in itself: the data never leaves the machine.

My advice is simple: take two or three real videos from your field, run them through OOOSplat and inspect the results in the browser player. If the quality meets your needs, you have a working prototype of a 3D feature in an evening. If it does not, you will quickly discover where the pipeline runs into trouble: the footage, the hardware or the approach itself. The repository is on GitHub under the ooolabdev organization.