Drop in a landscape video and get a vertical crop back that follows the person on screen — face tracking, shot-cut detection, and a smoothed virtual-camera pan, computed frame by frame.
runs entirely in your browser — the file never leaves your machine
How it works
- Face tracking. The video is sampled several times a second and faces are detected with MediaPipe (WASM, on your CPU/GPU). The subject is chosen by size and centrality, with continuity gating so a passer-by doesn't steal the crop.
- Shot cuts. Frame-difference detection splits the video into shots, so the virtual camera snaps at real cuts instead of gliding across them.
- Camera path. Per shot, gaps are bridged and the path is smoothed with a zero-phase filter plus a pan-speed cap — the camera holds still when the subject does and commits to moves without jitter or lag.
- Render. A full-height 9:16 crop is drawn along the path and re-encoded in the browser with the original audio.
This page is the browser edition. The full offline pipeline swaps MediaPipe for ArcFace identity re-identification (the crop locks onto a specific person, not just any face), TransNetV2 neural shot detection, and an L1-optimal camera path solved as a convex program — measured on stage footage at 30% true tripod-static time with 70 px/f max in-scene velocity.