How to blur faces, plates and screens out of a photo, without uploading it
A photograph of a protest, a classroom, an accident, a shop floor. You want to post it and you do not want to put strangers, children or a number plate on the internet. The mechanism matters here more than the button does.
Written · Updated
Open the photo in VEIL. Press Find and blur faces and it looks for faces itself, using a small detection model that runs in the tab; drag a box over anything it missed, or over a plate, a badge, a screen or a street sign. Each box is pixelated into the picture at its full resolution, and the file you save is a fresh PNG with the pixels changed for good. The photograph is never uploaded, and once the page has loaded it works with the network off.
Pixelation, not blur
These are not the same operation and the difference has consequences.
A Gaussian blur averages each pixel with its neighbours. It is a smooth, reversible-looking transformation, and in some conditions it is genuinely reversible: with a known blur radius, a known face and enough compute, deblurring is a solved problem, and there have been real cases of blurred faces and blurred text being recovered. A light blur over a number plate is decoration.
Pixelation replaces each block with one average colour. Information inside the block is not attenuated, it is gone. There is nothing left to undo, provided the blocks are large enough. That is why VEIL pixelates rather than blurs, whatever the button says, and why the block size is scaled to the box rather than fixed.
One caveat that applies to both, honestly: if somebody has the original photograph, no amount of anything on the copy helps. This protects the copy you publish.
What to cover, beyond the obvious
Faces are what people remember to do. The things that identify somebody in a photograph and get forgotten:
- Number plates, including reflections of them in windows and paintwork.
- Screens. A laptop or phone in shot, legible or not: a modern camera resolves more than your eye did at the time.
- Lanyards, badges and name tags, and the printed rota on the wall behind everybody.
- House numbers, street signs and shop fronts, which place the photograph precisely.
- Tattoos, distinctive clothing and mobility aids, which identify people in a small community as reliably as a face.
- Children's school uniforms. A crest is an address.
What the saved file carries
The copy VEIL writes is drawn from pixels into a new PNG, so the camera model, the timestamps and the GPS coordinates in the original do not come with it; the metadata is gone as a side effect of how the file is made, not as a promise. That is worth knowing both ways round: it is why the exported copy is safe to post, and why the original on your phone still has all of it. If you need the original cleaned as well, CLOAK strips the metadata from it, and EXIF shows you what was in there first.
It is a PNG, which is lossless and larger than the JPEG you started with. If the size matters, put it through SHRINK afterwards; the pixelation is already burned in, so nothing comes back.
Video, screenshots and documents
Different jobs, different tools, and using the wrong one is the usual way this goes wrong.
Why it matters that the photo stays here
The photographs people want faces taken out of are, by definition, photographs of identifiable people who did not consent to being posted. Uploading one to a free site to anonymise it puts the un-anonymised version on somebody else's machine, which is the exact thing the edit was for. In several jurisdictions a photograph of an identifiable person is personal data, and handing it to a third-party processor is a decision with a name and a paper trail. Doing it in the tab avoids the question entirely rather than answering it.
The short version
| The job | Here | Notes |
|---|---|---|
| Find the faces in a photograph automatically | Yes | A detection model that runs in the tab |
| Cover anything else by hand | Yes | Plates, screens, badges, signs |
| Make it irreversible | Yes | Pixelation, not a blur that can be undone |
| Keep the photograph's full resolution | Yes | The edit is at full size, not on a preview |
| Drop the camera and GPS data | Yes | The export is a fresh PNG |
| Keep the photo off somebody else's server | Yes | Read in memory; works offline |
| Blur a face in a video | No | RECAST does moving pictures |
| Redact text in a PDF | No | SEAL; a box over PDF text hides nothing |
| Protect a photograph somebody else already has | No | Nothing can; this is about the copy you publish |
Questions people ask about blurring faces in a photo
Is pixelation better than blurring for hiding a face?
Yes, and not by a small margin. A blur averages pixels together and can in some conditions be computationally undone: there are documented recoveries of blurred faces and blurred text. Pixelation with a large enough block replaces the detail with a single colour, and there is nothing left to recover. Use a blur for a look; use pixelation to hide something.
Can a pixelated face be un-pixelated by AI?
An AI can invent a plausible face from a pixelated one. It cannot recover the real one, because the information is not in the file any more. What comes out is a guess that looks convincing and is not evidence of anything. The risk to watch is the block size: too fine a grid leaves enough structure to narrow the field, which is why the blocks here scale with the box.
Does this remove the location data from my photo?
The copy it saves, yes: it is a new PNG drawn from pixels, so no camera, timestamp or GPS data comes across. The original on your device keeps all of it. If you need that cleaned too, CLOAK does it, and EXIF will show you what was in there.
Will it find every face?
No automatic detector does. It is good at faces looking roughly at the camera and worse at profiles, partly hidden faces, small faces in a crowd and faces in poor light. Check the result yourself and drag boxes over what it missed; that is what the manual boxes are for.
Is it safe to use an online face blurring site?
You would be uploading the photograph of identifiable people, un-blurred, to somebody else's machine, in order to blur it. In much of the world that photograph is personal data and the upload is a disclosure with a name attached. Doing it in the tab does not raise the question.
Can I blur faces in a video?
Not here; RECAST handles video. It is a different problem: faces move between frames, so a single box does nothing and the covering has to follow.
Related tools
- Can your writing style identify you?Yes, on a few hundred words, from the words you never choose: function words, punctuation and sentence rhythm.
- Share a run or a ride without giving away where you liveA GPS track starts and ends at your door, with the time on every point, and a privacy zone hides that on one site only.
- Can you make a voice unrecognisable by changing the pitch?Partly, and less than you think. It defeats a stranger, not somebody who knows you, and one number undoes it.
- How to anonymise names in a spreadsheet, and keep the same person the sameFind and Replace is not consistent across a second file and a lookup table is the re-identification key written down.
- Check a dataset before you publish itPostcode, date of birth and sex identify most of a population on their own. What k-anonymity measures, what generalising costs, and why a comfortable k can still give the whole answer away.
All 9 guides under “Keeping who you are out of it”. Part of the reading for newsrooms.
A letter, now and then
An occasional letter when there is something worth reading: the new guides and what changed in the tools. Never more than once a month and nothing else, and you can leave at any time on the unsubscribe page.