Automatic Face Blurring: How It Works and When to Review Manually

Automatic face blurring finds faces in a video and applies a moving effect to them. It saves you from placing a mask on every frame, but it can still miss a face, lose one behind an obstruction, or follow the wrong person when people cross.

Those are different problems. Understanding which one happened makes fixing the clip much easier.

Three jobs happen behind the button

Detection looks for face-shaped regions in an image. It may find a clear front-facing person but miss a tiny profile near the edge.

Tracking connects locations over time. If a person walks across the room, the mask needs to move with them. If they pass behind a door, the system has less visual information to work with until they reappear.

Rendering writes the chosen effect into the output. A well-positioned mask can still look too weak, while a very strong effect can sit on the wrong region.

Here is a useful diagnosis: if there is no mask, investigate detection; if the mask wanders, investigate tracking; if it stays attached but the face looks recognizable, investigate the treatment.

What the detection presets change in Unseen

The local import screen offers Relaxed, Balanced, Thorough, and Maximum. They control how frequently the local pipeline looks for faces, with different tracking behavior between detections.

Thorough is the recommended starting point. Maximum checks more frequently and is worth trying for fast movement, dense scenes, or appearances lasting only a few frames. Relaxed and Balanced can reduce work on calmer footage.

Maximum is not a guarantee. A face hidden behind a hand is still hidden from the detector, and more attempts cannot recover detail absent from the source. Cloud processing does not use this local preset control.

Why the face list is not a coverage report

In Select faces, the thumbnails help you choose who to hide. They do not prove that every appearance of that person has been found. A person can also produce multiple entries if their appearances are split.

For example, a presenter turns away, leaves the shot, and returns after a cut. The front-facing thumbnail may look perfect even though the brief return needs another look. Always inspect the video itself.

Review the moments most likely to fail

Start with every entrance and camera cut. Then look at head turns, people crossing, faces behind glasses or hands, and reflections. Finally, watch the whole result; a list of known hard cases cannot anticipate every miss.

Record a problem as a timestamp and a symptom: “00:18, profile appears without a mask” is more useful than “tracking is bad.” Rerun with a more thorough local preset when detection frequency seems relevant. If it still fails, open Review & fix faces, draw a box on the face at that moment and track it through the shot (step by step), or trim the shot.

Try the basic face-blur workflow first, then use the missed-frame guide when a specific problem appears.