Cover the faces before you share the photo

Why this has to run locally

Most online face blurrers upload your photograph to a server, find the faces there, and send back a redacted copy. Consider what that means. To hide the faces, you have first handed the unredacted original — the exact thing you were trying not to share — to a third party, along with its location metadata.

The face model used here is a few megabytes and runs in your browser. There is no reason to involve a server, and every reason not to.

Blur, pixelate, or block

They are not equally safe, and the difference matters more than it looks.

Solid block is absolute. The pixels are replaced. Nothing can be recovered because nothing remains. It looks blunt, which is sometimes exactly right, and you can set its colour — white blocks disappear into a pale background, while a red one announces that something was deliberately removed.

Pixelation genuinely destroys detail by averaging each block down to a single value, and shows that it has done so. At three to ten blocks across a face it is irreversible in any practical sense.

Blur is the most natural-looking, and here it is not a true Gaussian blur at all. A real blur only attenuates information and is invertible in principle, which is why published research has recovered identities from mildly blurred faces. What this does instead is resample the face down to a handful of samples and scale it back up smoothly. It reads as a heavy blur and discards the detail exactly as pixelation does.

Both settings have a floor, because a privacy control whose weakest position leaves a recognisable face is worse than no control — it hands you false confidence. The slider chooses how the result looks, never whether it works.

Check the result, every time

Automatic face detection misses things. It is good at reasonably front-facing faces at a useful size, and much weaker on profiles, faces turned away, small faces in the background, heavy shadow, motion blur and partial occlusion.

The failure mode of a tool like this is silent: you get a picture that looks redacted, you share it, and one face in the back row was never covered. So look over the output before it leaves your hands. The tool tells you how many faces it found precisely so you can compare that against how many you can count.

Metadata, which is the part people forget

A photograph carries more than an image. EXIF metadata routinely includes the exact GPS coordinates where it was taken, the timestamp, and the device that took it. A picture with every face blurred and the location intact can still tell a reader which school, which clinic, which house.

Because the output here is re-encoded from a canvas, none of that metadata survives into the saved file. That is a genuine benefit and worth knowing about even on photos with no faces in them at all.

What redaction does not cover

Faces are one identifier among several. A photograph can identify someone through a distinctive tattoo, a name badge, a licence plate, a house number, a uniform, a reflection in a window, or the simple fact of who else is in the frame. Screenshots of documents are worse still: covering the photograph on an identity card leaves the name, number and address perfectly readable.

This tool covers faces. Deciding what else in the frame identifies someone is a judgement no automatic tool should be trusted to make for you.

This tool processes images on your device and transmits nothing. It cannot guarantee that every face in an image is detected, and you remain responsible for checking the result before sharing. For material with serious privacy or legal consequences, use pixelation or solid blocking rather than blur, and review the whole frame rather than the faces alone.

Questions people ask

Is my photo uploaded anywhere?

No, and for this tool that is the whole point. The face model runs in your browser using WebAssembly, and the redacted image is produced on your device. A redaction tool that uploaded your unredacted photograph to find the faces would have defeated itself before it started, because the original would already have left your device.

Can the blur be reversed?

Pixelation and blocking genuinely discard the detail rather than hiding it, so there is nothing to recover from the saved file. Heavy blur is also effectively irreversible. Light blur is riskier: research has shown that faces obscured with mild blurring can sometimes be re-identified, so if the stakes are real, use pixelation or a solid block rather than a gentle blur.

Does it find every face?

No, and you should check. It works well on reasonably front-facing faces at a decent size. Profiles, faces turned far to one side, very small faces in the background, heavy shadow and motion blur are all frequently missed. Always look over the result before sharing, because a tool like this failing quietly is worse than not using one.

Why can I untick individual faces?

Because consent is usually uneven. In a group photo one or two people may be happy to appear while the rest are not, and blanket redaction is not always what you want. Every detected face starts covered and you uncover the ones that should stay.

Does this remove metadata from the photo?

Yes, as a side effect. The saved file is re-encoded from a canvas, so EXIF data including GPS coordinates, camera model and timestamp does not survive. That matters: a photo with faces blurred but location metadata intact can still identify where someone was.

What about faces on identity documents or screenshots?

The face will be covered, but everything else on the document will not. Names, numbers, addresses and barcodes all remain readable. Redacting a document properly means covering every identifying field, not only the photograph.