4DGS Pipeline

Detect Sync

Cameras move between shots, so each sample time gets its own COLMAP solve (one frame snapshot per camera, at trim_start + sample time). Preview first — COLMAP+train per timestep is expensive.

Auto uses MASt3R instead of COLMAP for small camera counts (2 raw cams, or 2 cams x stereo-eye-split) — COLMAP can't run below 3 images and struggles on a moving subject even at 4. "Fill" replaces Sample times with an evenly-spaced list across the range at the target FPS; "All frames" uses each video's native frame rate (capped at 500 samples — re-run per timestep is the whole chain, not incremental, so more samples means proportionally longer runs).

Drag a box around the subject on each camera you want masked (unmarked cameras stay full-frame). One box per camera is enough — SAM2 tracks it across every timestep. Runs a separate joint solve + masked train + compress, written alongside the regular pipeline (doesn't touch it).

Review trained PLYs, apply per-frame transforms and crop regions.