Blurring Faces and Number Plates Before You Post
Blur is reversible if you do it wrong, and most people do it wrong. Which methods actually destroy the information, what else in the frame identifies someone, and how to check before publishing.
Updated August 23, 20266 min read
Covering a face in a photograph seems like a solved problem: drag a blur over it and publish. The difficulty is that not every method of obscuring actually destroys the underlying information, that the face is often not the only thing identifying the person, and that the failure is silent — a photo published with reversible redaction looks exactly like one published with irreversible redaction until somebody demonstrates otherwise.
Why some blurs can be undone
A blur is a mathematical operation: each output pixel becomes a weighted average of its neighbours. Averaging discards information, but not all of it, and the operation is well understood enough to be partially inverted. Given a mild blur and knowledge of its radius, deconvolution recovers a surprising amount of the original — enough for a face to become recognisable again.
Pixelation is worse, and this is counter-intuitive because it looks more destructive. A pixelated block is the average colour of the pixels beneath it, so a face reduced to a grid of coloured squares is a low-resolution photograph of that face — a real image, just small. If the set of possible originals is limited, which it always is for text and number plates and often is for faces from a known group, a search over candidates matched against those averages recovers the original reliably. The same technique has repeatedly been used to reverse pixelated text.
| Method | Recoverable? | Use it when |
|---|---|---|
| Light blur | Yes — often trivially | Never, for privacy |
| Pixelation / mosaic | Yes — a known attack | Never, for privacy |
| Heavy blur, radius above ~15% of the face | Unlikely, but not impossible | Casual anonymity, low stakes |
| Solid fill or emoji sticker | No | Anything that matters |
| Cropping the person out | No | When composition allows it |
The face is not the only identifier
Anonymising a photograph properly means asking who could identify this person, not just whether their face is visible. Somebody who already knows them needs far less than a stranger does.
- Number plates and vehicles. A plate is a direct lookup for anyone with access, and an unusual car is nearly as good in a small area.
- Name badges, lanyards, ID cards. Frequently legible when you zoom in, and frequently missed because they are small in the frame.
- Screens in the background. Monitors and phone displays catch email addresses, client names, open documents and internal dashboards.
- Tattoos, distinctive clothing, hair. These identify reliably to anyone who knows the person, which is usually the audience that matters.
- House numbers, street signs, shopfronts. Together these place a photo precisely, which matters when the point of anonymising it was to hide a location.
- Documents on a desk. Letterheads, addresses on envelopes and printed schedules are all readable at modern camera resolutions.
The region blur tool handles all of these — you mark the areas rather than relying on face detection, which by definition only finds faces. Where the composition permits it, cropping the offending part out of the frame entirely is stronger than covering it, because there is nothing left to recover.
The metadata problem
You can blur every face in a photograph perfectly and still publish the exact address where it was taken, because that information is not in the picture. Phone cameras write EXIF metadata into the file: GPS coordinates to within a few metres, the timestamp to the second and the device model — and many dedicated cameras add a body serial number that links every photo that camera has ever taken.
Some platforms strip this on upload and some do not, and the ones that do strip it are not consistent about which fields. Do not rely on it. Check what your file actually carries with the metadata viewer, and remove it deliberately before publishing — stripping EXIF metadata covers the fields worth caring about and the ones that are harmless.
A workflow that holds up
Decide what you are protecting against
Hiding a bystander from casual viewers is a different job from protecting a source whose identity somebody has a motive to uncover. The first tolerates a heavy blur; the second needs solid blocks or a crop, and probably a second pair of eyes.
Let detection find the faces, then check its work
The face blur tool locates faces automatically, which is fast and catches the ones you would skim past. It will miss profiles, partial faces at the edge of frame, faces in reflections and faces on a poster in the background. Review every one and add what it missed.
Sweep the frame for everything else
Go corner to corner at full zoom looking for plates, badges, screens and documents. This is the step people skip and it is where most failures originate.
Use enough strength
Blur until nothing about the region reads as a face — no eye positions, no jawline, no expression. If you can still tell which way somebody is looking, it is not strong enough.
Export, then inspect the export
Open the finished file and zoom to 400% on each covered region. Check the metadata of the exported file separately. Only then publish.
Detect and cover faces on your device — the photo is never uploaded, so no server ever holds the unblurred original.
Blur faces in a photoWhere the file goes matters as much as the blur
There is a contradiction at the heart of most online blurring tools. You are anonymising a photograph because it contains something that should not be widely seen, and the tool's first action is to upload the unmodified original to a server you do not control. Whatever the privacy policy says, the un-redacted image now exists somewhere else, in logs, in backups, in a processing queue.
Every tool here runs entirely on your device, not our servers: the photograph is read, processed and written back out as a new file by your own hardware. Nothing is uploaded, which means the only copy of the original that ever existed outside your device is the one you choose to share. For photographs of children, medical documents, protest footage or anything under a confidentiality obligation, that distinction is the whole point rather than a marketing line.
One last thing worth knowing about publishing: many platforms re-encode images on upload, which can slightly soften a heavy blur but never restores what was destroyed. Redaction that is genuinely irreversible stays irreversible after any amount of resizing, recompression or screenshotting. If a technique only holds up at the original resolution, it was not strong enough to begin with.
Frequently asked questions
Can a blurred face be unblurred?
A light blur often can be, because blurring averages pixels rather than deleting them, and the operation can be partially inverted when its strength is known. Pixelation is worse still, since each block is the average colour of a real region and can be matched against candidate images. Only a very heavy blur or a solid fill genuinely destroys the information.
Is pixelating a face safe?
No. A mosaic is a low-resolution photograph of the thing it covers, and reconstructing the original from the block averages is a well-established technique, especially where the range of possibilities is limited. Use a solid block or a very heavy blur instead when it matters.
Do I need permission to post photos of people?
It depends on where you are and the context, and this is not legal advice. Many jurisdictions treat photographs of identifiable people as personal data, with stricter rules for children, medical settings and anything suggesting political or religious affiliation. Blurring is one way to reduce the question; asking is the other.
Does blurring a photo remove its location data?
No. GPS coordinates, timestamps and device details live in the file's EXIF metadata, entirely separate from the pixels, and many editors copy them through to the exported file. Check and strip metadata as a distinct step after blurring, on the exported file rather than the original.