Face-tracking crop

On every frame of the source video, a face-detection model locates the dominant face and emits a bounding box. The crop window follows the box with smoothing applied so it doesn't jump frame-to-frame. When the face leaves the frame entirely (cutaway, B-roll), the crop falls back to a center crop or to the last known position depending on configuration.

Quality face-tracking handles two common edge cases: multiple faces (the algorithm picks the most confident detection or splits the screen) and fast motion (Kalman-filter-style smoothing prevents the crop from chasing every twitch).

• Heavy graphics overlays — the algorithm can confuse a face on a thumbnail in the corner with the actual speaker.
• Profile shots — frontal faces detect more reliably than side profiles.
• Dim or backlit footage — low-contrast frames produce low-confidence detections.

When a clip's tracking is wrong, manual override is faster than re-running the model with different parameters. Clipperz exposes a per-clip crop override in the editor for the few cases where automation needs help.