A black bar across the eyes, and it stays there.
2 files freeNo cardNothing uploaded
Blurred
Recoverable
Covered
Nothing left to recover
Nothing survives under it
Every pixel is replaced with the same value. There is no kernel to invert and no surviving fingerprint to match against a list of candidates.
It is measured, not guessed
The rectangle comes from the eye-ring landmarks and is then dilated. Where those are not known confidently, the whole head is covered instead.
It stays on through the clip
Tracked rather than placed. The bar moves inside the same motion the head does, and gaps in detection are carried across rather than dropped.
What the bar is good at, and what it is not
Measured per pixel covered, a solid bar is the strongest treatment available. It is not an average of anything, so there is no arithmetic to run backwards, and it leaves no residue to compare against a suspect list the way a mosaic does. In that narrow sense it is the end of the argument.
The weakness is not the bar. It is how little of the face a bar covers. A strip across the eyes removes the region automated matching depends on most, and leaves everything a human being recognises a friend by.
So the useful question is never “is a censor bar strong enough”. It is who is looking. A stranger scrolling past, or a face-matching system, is substantially defeated by an eye bar. Somebody who has met the person is not defeated by it at all — they will know from the jaw, the hairline, the shoulders and the way the mouth sits, and no amount of blackness across the eyes changes that.
When an eye bar is the right choice
- A bystander in a photograph who is not the subject and does not need to be identifiable to anyone.
- A document, a report or a filing, where the mark needs to read as a deliberate removal rather than as a bad photograph.
- Publication, where the face must not be matchable by a system even though the people in the room already know who it is.
When to cover the whole head instead
- Anyone who could be recognised by someone who knows them — which, in a small community or a workplace, is most people in most footage.
- Anywhere a tattoo, a scar or unusual hair is visible, because those identify as reliably as a face and an eye bar does not touch them.
- Anything where being identified has a consequence serious enough that you would not want to be arguing about margins afterwards.
How this one is drawn
- Where it comes from
- The eye-ring landmarks, with a floor on the half-height so a squint or a downward glance does not produce a thin bar. It is then dilated along and across the eye line.
- Its shape
- A flat rectangle with softly rounded corners, in pure black, running past the outer corner of each eye. A bar laid over a photograph, not a mask fitted to a face.
- Through a video
- Tracked per face. Gaps between detections are filled, short gaps past the end of a track are carried, and a track lost for too long escalates to covering the whole head.
- Afterwards
- The written file is decoded and read back frame by frame. Anything the first pass missed is covered again, wider, and anything still unresolved is shown to you with its timestamp.
- Other covers
- A frost and a streak, if a bar is too blunt for the picture. Both destroy the pixels rather than averaging them — though only the bar leaves literally nothing.
The way a bar actually fails
Not by being reversed. By not being there.
One uncovered frame in a thirty-second clip is one clean photograph of the person.
A tracker that loses a face during a turn leaves three or four frames uncovered, and nobody notices because nobody steps through an export frame by frame — which anybody wanting to identify the person absolutely will. That is the failure this pipeline is built around: the whole finished file is decoded again and re-checked after it is written, and what cannot be cleared is reported rather than quietly shipped.
The second failure is worse and is not a video problem at all. A bar drawn as an overlay — a shape on a layer above the picture, in an editor or a PDF — is not a redaction if the export keeps the layers, or if the thing underneath is live text rather than pixels. Published court filings have been un-redacted by selecting the text under the black rectangle and pasting it elsewhere. If you are removing text rather than a face, use a tool that deletes the text.
Questions
Is a black bar over the eyes enough to hide someone?
It depends entirely on who is looking. Against a stranger and against automated face matching, a bar across the eyes is strong — the eye region carries a great deal of what those systems use. Against somebody who knows the person it is close to useless: a jaw, a hairline, a mouth and a posture are more than enough for recognition. If the risk is a person rather than a system, cover the whole head.
Can a censor bar be removed or reversed?
Not if it is in the pixels. A solid bar replaces everything underneath with one colour, so there is no information left in the covered region to recover and nothing to compare against a list of candidates — unlike a blur, which keeps the face in smeared form. The way a bar actually fails is different: an overlay that was never flattened into the exported file, or a frame somewhere in the clip where it was not drawn.
Does the bar follow the face in a video?
Yes. Each face is tracked, so the bar sits inside the same movement the head does rather than being pinned to a point on the screen. Where the detector loses a face — a steep turn, deep shadow — the track is carried across the gap, and where it has been lost too long to carry honestly, the whole head is covered instead of the bar being placed on a guess.
What size should the bar be?
Wide enough to run past the outside corner of each eye and deep enough that no part of either eye shows at any point in the clip. The bar here is built from the eye landmarks and then dilated, so it is proportioned to the face rather than to a number somebody typed. Where the eye positions are not known confidently, the covered area is escalated rather than estimated.
Is a bar better than pixelation?
Per pixel covered, yes — a mosaic leaves the surviving block colours, which can be matched against a short list of suspects, and a bar leaves nothing at all. In practice the choice is usually about how the result should read: a bar announces that something was removed, which is what you want in a document, and a mosaic is less visually aggressive, which is what people often want in a photograph.
Can I put a bar on a scanned document or a screenshot?
If it is an image with a face in it, yes — it is handled the same way as any other photograph. But do not use any image tool to redact text in a document. Text redaction needs the text itself removed from the file rather than a shape drawn over it, and drawing a black rectangle on top of live text is the single most famous way redactions fail.
Put the bar on.
A photograph, a batch of them, or a video where it has to stay on the face for every frame. It runs in this browser and nothing is uploaded.
2 files freeNo cardNothing uploaded