The short version of the pipeline — from a clip to a finished splat in four steps. Each step links to a detailed page under Advanced when you want to dig in.

1. Extract

Add your video clip to the queue. Raw fisheye footage is recommended for interiors, stitched 360 for exteriors. The app detects the input type and extracts frames from the video, automatically keeping the sharpest frame in each interval. You can adjust how often frames are extracted.

For raw fisheye, choose 1, 5 or 9 splits per frame — 5 splits is recommended, so start there. Stitched 360 material always uses a 6-face cubemap.

→ More detail: Extraction.

2. Masking

Use masking to remove things you don't want in the reconstruction — the person holding the camera, moving objects, the sky. Type a keyword like "person" to mask it automatically, or paint areas by hand.

With raw fisheye material, the area outside the image circle needs to be masked. There are preset masks for .insv and .osv files — you can adjust them or add your own.

You can also paint out unwanted elements with the brush, either frame by frame or across all frames at once.

The masking step downloads an AI model (~3 GB) the first time you use it. This is a one-time download.

→ More detail: Masking.

3. Alignment

Alignment runs on a custom build of COLMAP.

Matcher — Sequential is recommended. If your clip has GPS data, use Spatial.
GLOMAP — much faster, so leave it on. If you scanned multiple rooms and the alignment comes out wrong, try turning it off.
Geometry timeline check — experimental; fixes similar-looking places being merged into one.
Feature quality preset — start with Default, or use Low on a low-spec computer.
Detail photos — add extra photos from a smartphone, DSLR or drone; they are registered into the main alignment.

→ Full settings reference: Alignment.

4. Train

Connect the path to your chosen splat trainer — LichtFeld, Postshot, Brush — or select no train. If you select no train, all the training data you need is in the 0_train_data folder (images, masks, sparse).

→ More detail: Train.