How to blur every bystander but keep your subject sharp

Last updated 2026-09-04 · 8 min read

What is the short version?

Blurring faces in public footage has two failure modes that pull in opposite directions. Miss a stranger and you have published someone who did not agree to be in your video; blur your own subject and you have published nothing worth watching.

1. Go to Vizard Agent with the footage.
2. Say which person must stay sharp, and give reference faces.
3. Ask it to verify in both directions before delivering.

What do you need before you start?

Footage and a few clear looks at the person who stays sharp. Reference faces from different angles are worth far more than one perfect portrait, because the recognition has to work on the profile shots and the half-turned ones, not the ones you would choose for a headshot.

• The footage. However long.
• Reference faces. Several, different angles.
• Who else is exempt. A co-host, a named guest.
• The blur style. Soft, pixelated, or a shape.
• Any local rules. They vary; the obligation is yours.

What do you type into Vizard Agent?

Name the exception rather than the rule. "Blur the faces" is ambiguous about your own subject, and the version of this instruction that actually works tells Vizard Agent who stays sharp and lets it treat every other face in the footage as the default case.

Prompt

Variants worth knowing:

• "Except mine." Names the exception explicitly.
• "Including in profile." The hardest case, said up front.
• "Check both directions." Missed strangers and false blurs.

What does Vizard Agent actually do?

Here is the order Vizard Agent worked in on a real nineteen-minute vlog filmed in a busy street. Step four is the one that stops the whole approach from failing quietly: it checks the signature against strangers before it blurs anything.

1. Sets up the face detection and recognition models.
2. Builds the subject's face signature from the references.
3. Crops the chosen reference faces and confirms they are all the same person.
4. Tests recognition against strangers' faces to see who wrongly matches.
5. Tunes detection to catch profile faces, which the default misses.
6. Checks the subject is reliably recognised across the whole video.
7. Writes the blur pass and tests it on two tricky segments first.
8. Inspects the blur results on those segments.
9. Expands the references and re-checks that strangers stay unmatched.
10. Runs the blur across all fifty-eight segments.
11. Spot-checks the blur quality at key moments.
12. Checks the face-track scores in the segments that look wrong.
13. Debugs why a bystander stayed sharp in one segment.
14. Measures track-level similarity for better separation.
15. Visualises the ambiguous tracks and identifies them.
16. Re-runs with refined identity rules and profile protection.
17. Debugs why the subject was still being blurred in another segment.
18. Fixes that segment specifically and re-checks it.
19. Checks whether the source already has blur baked in.
20. Extends the blur to a person who appears again later.

Steps thirteen and seventeen are the same work in opposite directions, and both have to be done. A pass that only checks for missed strangers will happily blur your presenter's face for four seconds in the middle of the video and nobody will catch it until the comments do.

Step nineteen is a detail worth stealing. Some source footage already has blur applied — from a camera app or a previous edit — and knowing which parts are already covered stops you double-treating them and tells you what the source considered sensitive.

What does the result look like?

Footage where every face that is not your subject is covered, including the ones in profile, the ones at the edge of frame and the ones who reappear later in the video. Your subject is sharp throughout, including when they turn away from the camera.

The verification is the deliverable as much as the video is. Knowing which segments were checked, and which two needed a second pass, is what makes it safe to publish.

When does this not work well?

Recognition needs faces it can actually resolve in the frame. Crowds at a distance, heavy motion blur and low light all reduce Vizard Agent to guessing, and a blur pass that guesses is not a compliance measure no matter how thorough the rest of it looks.

• Distant crowds. Faces too small to detect reliably.
• Very low light. Detection degrades sharply.
• People wearing masks or helmets. Detection may skip them entirely.
• Twins or close relatives. The signature may match both.
• Legal certainty. This is a tool, not advice; the obligation stays yours.

How do you fix a result that came back wrong?

Name the moment and say which of the two directions failed there. Vizard Agent keeps the face signature, the per-track scores and the per-segment results from the first pass, so a correction re-runs that one segment under an adjusted rule rather than reprocessing the entire video.

• "You missed someone at 4:20." That track re-scored and blurred.
• "You blurred me here." Profile protection applied to that segment.
• "This person should be sharp too." Added as a second exception.
• "They come back later." The blur extended into the later scenes.

How does Vizard Agent compare to doing it yourself?

By hand this is per-face rotoscoping across every segment, which is why most people either skip it or blur the whole background. The tooling exists; the reason it does not get done is that a nineteen-minute vlog has a great many faces in it.

Common questions

How many reference faces do I need? A handful from different angles. Vizard Agent will ask for more if the matching is shaky.

Will it blur me by mistake? It can, which is why Vizard Agent tests for that specifically before delivering.

What about profile shots? The default detection misses them. Vizard Agent tunes for them deliberately.

Can I exempt two people? Yes. Give references for each and Vizard Agent keeps both sharp.

What kind of blur can it use? Soft, pixelated or a shape. Tell Vizard Agent which; the choice is yours.

Does it handle someone appearing later? Yes, once identified. Vizard Agent extends the blur across the video.

Can it tell me which segments it checked? Yes. Ask Vizard Agent for the verification list; it is worth having.

What if the source is already blurred? Vizard Agent checks for baked-in blur and works around it.

Does it slow the render down? Yes. A long video with many faces is a genuinely long pass.

Can it blur bodies or just faces? Faces by default. Vizard Agent can cover more if you ask.

What about faces on screens in the shot? Say so. They are detected but you may want them treated differently.

Is this enough for GDPR? It is a tool, not legal advice. Vizard Agent cannot take that obligation on.

Can I see it before the full run? Yes. Vizard Agent tests on the trickiest segments first.

Why test against strangers first? Because a signature that matches half the street looks like it is working right up until you watch the whole video back.