Blur a license plate

There is no automatic plate detection here

This site finds faces for you. It does not find plates for you, and that deserves an explanation rather than a quietly missing feature.

Faces are handled by the face tool, which uses Google's MediaPipe BlazeFace model. That model is published under the Apache-2.0 licence and is small enough to download to a phone once you have picked a file. Nothing equivalent exists for number plates. The plate detectors that genuinely work are either licensed for research or non-commercial use, which rules them out on a site funded by advertising, or they are large enough that pulling them over mobile data would take longer than doing the job by hand.

Variety is the deeper problem. Plate proportions, mounting height, background colour, character set and typeface differ enormously from one country to the next, and a single photograph can hold a plate at an angle, in shadow, or half hidden behind a tow bar. A detector that finds three plates out of five is worse than no detector at all: it teaches you to trust it, then misses the fourth one on the day it matters.

So this page asks you for a box. Drag it over the plate, then drag another over any second plate in the frame. It takes about four seconds and it is correct every time, which is more than an unreliable model can offer. The same hand-drawn box is how the black out text tool works, for the same reason.

Why a solid bar, not a blur

The default here is a solid black bar, and on a license plate that is not a matter of taste. It is the only mode that actually removes the information.

Consider what a plate is. It is a short string of characters, drawn from a small and highly structured alphabet, set in a known typeface at known spacing inside a rectangle of known proportions. Almost everything about it is fixed before the shutter opens. The only unknown is which characters.

Blur and pixelation do not delete pixels. They average them. A blurred plate still carries the shadow of every dark stroke against its pale background: where the ink was heavy, the average stays dark. That residue is useless when the thing underneath could have been anything at all. It is extremely useful when the thing underneath could only have been one of a limited set of characters, because then nobody has to read the plate. They only have to rank guesses. Render a candidate string in the right font, blur it the same way, compare it against the picture, repeat. The closest match is the answer, and the work is arithmetic rather than insight.

That is the difference between a plate and a face. A face is a large, irregular subject with no finite alphabet behind it, so averaging its pixels throws a great deal away — though not always as much as people assume, which is the subject of whether AI can unblur a face. A plate is the opposite: small, structured, and drawn from a short list. Structure is precisely what survives averaging.

A solid bar replaces the pixels with a flat colour and cannot be undone. Pixelation and blur average the pixels, and averaged pixels can sometimes be partly recovered. If you want the three modes set out side by side, blur, pixelate or black box compares them. For a plate, take the bar. Save pixelation for cases where you want a viewer to see that something was covered and the exact content does not matter.

Where plates hide in a photo you did not think about

The plate you meant to cover is the easy one. These are the ones people miss.

• Your own plate reflected in the paintwork, usually in a door or across the bonnet on a bright day.
• A plate in a wing mirror — small, sharp, and sitting in the middle of the frame in any three-quarter shot.
• The neighbour's car at the kerb behind you. It is not your plate, but it is somebody's, and it places your car on a particular street.
• A reflection in a shop window or a glass door behind the car, which can show the plate of a vehicle that is not even in the shot.
• The small plate on a trailer, a caravan or a bike rack, easy to overlook because it is not where a plate normally sits.

One practical point on top of that: across a set of photos you almost always publish both the front and the rear plate, because you photograph the car from both ends. Covering one covers nothing. Work through every image in the set, not only the one you thought was the risky one.

The rest of the car photo

The plate is the obvious identifier and rarely the only one. A VIN plate at the base of the windscreen, a service book, a registration document photographed to prove the mileage, an insurance certificate, a delivery label on a box in the boot — all of that is text, and text is covered with the black out text tool. Faces caught in the background of a driveway shot belong to the face tool. The file itself may also carry the GPS coordinates of the spot where you took the picture, which the metadata stripper will show you. Saving a photo from this page re-encodes it, and that drops the EXIF block along the way.

For the full sequence — which photos to take, what to cover in each, and what to leave in so the advert still looks honest — read selling your car online.

What this tool does

Where it runs

Entirely in this browser tab. No image is uploaded, because there is no server to upload it to.

Modes

Solid bar (the default, irreversible), pixelate, blur.

Accepts

JPG, PNG, WebP, GIF, BMP, AVIF, SVG, HEIC/HEIF and TIFF. Camera RAW is refused.

Exports

PNG at the resolution of the file you opened, or JPG. No watermark, no cap on how many photos you do.

Metadata

Re-encoding through the canvas strips EXIF, including GPS coordinates.

Not supported

Video. Still images only.