How to blur every face so nobody can be identified
Tell Vizard Agent that nobody in the video may be recognisable and that nothing else may change. That is a pass-or-fail standard rather than a setting: a face missed for ten frames because it turned sideways defeats the whole purpose, so the tuning is done against the hardest face in the clip.
What is the short version?
You need to publish something with people in it and none of them can be identifiable. Automatic face blurring finds the easy faces. Vizard Agent works on the ones it misses, and on the things it wrongly blurs.
- Tell Vizard Agent every face, nothing else changed.
- Let it check the small and side-on faces specifically.
- Review the doubtful frames rather than trusting the count.
What do you need before you start?
The video, and a warning about who is in it. "There is a child in this, partly behind an adult" is worth saying, because small and partially hidden faces are where detection fails and where the consequence of failing is highest.
Say whether anything else may change. In the session behind this article the instruction was explicit that the video must be untouched apart from the faces, which rules out cropping, stabilising or re-framing as ways of solving the problem.
What do you type into Vizard Agent?
State the standard rather than the effect you want applied. "Blur the faces" describes an operation that can be performed badly and still count as done; "so it is impossible to tell who anyone is" describes the result you will actually be judged on.
Blur every face in this video completely, so it is impossible to tell who anyone is. Do not change anything else about the video.
Both halves matter. Without the second, a reasonable way to hide a face is to crop it out of frame, which is not what you asked for.
What does Vizard Agent actually do?
It gets the detection right before it blurs anything. Several detectors are tried on sample frames, the results are drawn as boxes and looked at, and the model that finds the difficult faces is the one that gets used — not the first one that returns something.
In that session the sequence ran:
- Found the video's rotation lived in metadata, which meant the detector was being shown sideways frames and finding nothing.
- Prepared a correctly oriented working copy without changing the visible frame.
- Compared detection models and switched to one that handles small and side-on faces.
- Built per-frame face coordinates rather than one moving box, so there are no gaps between detections.
- Tuned the threshold so real faces were masked but the water, the netting and a bucket were not.
- Kept stable face trajectories and discarded brief spurious detections.
Then it went through the doubtful frames one at a time, at full size with the boxes drawn, deciding in each case whether that was the child's face or an object — including the smallest, earliest appearance, specifically so it would not be missed.
What does the result look like?
Every face obscured for every single frame it appears in, with nothing else about the picture altered in any way. The blur follows the movement without jumping between frames, and objects that briefly resemble faces are left alone entirely.
Vizard Agent checks the delivered file in motion as well as on frames. A mask can be correct on every sampled frame and still flicker between them, and flicker is exactly what makes a face readable for an instant.
When does this not work well?
When someone is identifiable without their face. A distinctive uniform, a name on a shirt, a tattoo or a car registration all identify a person, and a face blur does not touch any of them. Say if the requirement is anonymity rather than face-hiding.
Very small or very blurred faces are ambiguous in both directions — hard to detect and arguably already unidentifiable. Vizard Agent will show you the borderline cases rather than deciding your threshold for you.
And this is obfuscation rather than redaction. A heavy blur is not reversible in any practical sense, but if the requirement is a legal one rather than a practical one, a solid opaque box is the safer instrument and Vizard Agent will use that instead if you ask.
How do you fix a result that came back wrong?
If a face got through anywhere, give the timestamp where you saw it. Vizard Agent inspects that frame at full size, works out why the detector missed it — the angle, the size, something in front of it — and adjusts for that specific case rather than lowering the threshold across the whole video.
If something that is not a face got blurred, name it. Lowering the threshold to catch a hard face is what causes this, and the fix is usually rejecting short unstable tracks rather than raising it again.
If the blur jumps around, that is stability rather than detection. Ask Vizard Agent to smooth the trajectories, which keeps the mask on the face between frames where detection is briefly uncertain.
How does Vizard Agent compare to doing it yourself?
One-click face blur is in every editor, and it works on a clip of adults looking at the camera. It fails quietly on the cases that matter — a face turned away, a child at the back, someone in shadow — and the failure is a few frames long, which is exactly long enough.
The difference is checking the failures rather than the successes. Looking at the frames where the detector was unsure, at full size, is the only way to know whether the standard has been met, and it is the step that a slider does not have.
Common questions
Will anything else change? No. Vizard Agent alters only the face regions.
What if a face is only visible for a second? It still gets blurred. Vizard Agent works from per-frame coordinates.
Can it blur a face turned away? Yes, with the right detector. Vizard Agent tests several on your footage.
Why did it blur a bucket? A low threshold catches false positives. Vizard Agent rejects unstable tracks instead.
Can I use a box instead of a blur? Yes, and it is the safer choice if the requirement is legal.
What about names on shirts? Not covered by a face blur. Tell Vizard Agent and it masks those too.
Does the audio change? No, unless you ask for voices to be altered as well.
How strong should the blur be? Strong enough that features are unreadable. Vizard Agent checks at full size.
Can it keep one person visible? Yes, that is a different job and Vizard Agent can do it.
Why did the detector find nothing at first? The video's rotation was in its metadata. Vizard Agent fixes that before detecting.
Does it work on a crowd? Yes, though the small faces at the back are the hard ones.
Can I see where it found faces? Yes. Ask Vizard Agent for the frames with the boxes drawn.
Will the file be re-encoded? Yes, the picture is. Vizard Agent keeps the original geometry.
What about reflections in glass? Say so. Vizard Agent will not find a reflected face on its own.
Can it handle a moving camera? Yes. Per-frame detection does not depend on the camera holding still.
How do I know it did not miss one? Review the frames with the boxes drawn. Vizard Agent produces them for exactly that check.