Blur Faces and Plates Before You Post, Without Uploading the Photo

A photo you are about to post has three people in the background who never agreed to be in it, and a parked car whose plate is perfectly legible. You want them gone before the image leaves your device, not after. That is the job Privacy Blur does, and the interesting part is not the blurring. It is everything that has to happen correctly before a single pixel is smudged.

Detection runs only when you ask for it

Two detectors are available and neither one is loaded when the page opens. Faces use a roughly 15 MB model. Licence plates are far heavier: about 81 MB, plus a shared runtime that may not be cached yet. Downloading either without asking would be a surprise on a metered connection and a bigger surprise on a phone.

So the download is a decision you make, once, per detector. After that the model is cached and the second use is instant. If you only ever blur faces, you never pay for the plate model at all.

This is also why the two detectors are separate buttons rather than one "find everything" sweep. A sweep would have to fetch both.

An empty result stays empty

Here is the behaviour that matters most, and it is a refusal rather than a feature.

When a detector finds nothing, the tool reports that it found nothing. It does not fall back to a guess, it does not draw a speculative box in the middle of the frame, and it does not quietly hand you an image with a blur applied somewhere plausible.

That sounds obvious until you consider the alternative. A privacy tool that invents regions is worse than no tool, because it produces an image that looks processed. You would ship it believing a face had been covered. A visible "no faces detected" tells you to place the regions yourself, which is the correct next step and takes about four seconds.

Detection also is not perfect on hard input. A face at a steep angle, mostly out of frame, heavily shadowed, or very small in a wide shot may be missed. That is exactly the case where an invented box would be most dangerous and an honest empty result is most useful.

Every region is reviewable before anything is applied

Detected regions arrive as editable rectangles, not as a finished effect. You can move them, resize them, delete a false positive, and add a region the detector missed. Nothing is baked until you apply.

Three failure modes this catches:

The box is too tight. Detectors frame the face, not the head. A tight box leaves the hairline, the ears and the jaw visible, which is often enough to recognise someone. Widening the region takes one drag.

The box is on the wrong thing. Patterned fabric, a printed face on a poster, a reflection. Deleting is faster than working around it.

Something was missed. The person in the background who is turned away, the second plate on a trailer. Add a region by hand.

Blur or pixelation, and they are not equivalent

Both are offered and the choice is not cosmetic.

Pixelation is easier to read as deliberate. It looks like an intervention, which is often what you want on a published image: nobody thinks the photo was accidentally out of focus.

Blur looks more natural but a weak blur is genuinely reversible in the sense that matters. If the region is large, the content is high contrast, and the radius is small, a person can often still tell who it is. If you use blur, be generous with it, and check the result at full size rather than at thumbnail size. A blur that hides a face in a 200 pixel preview may not hide it at 2000 pixels.

Video is tracked, not frame by frame

On video the regions are tracked across frames rather than re-detected on every one. This is what makes the tool usable on a clip of any length, and it has a consequence worth knowing: if the subject moves sharply, leaves the frame and returns, or the shot cuts, the tracked region may drift or lose its subject.

So the review step matters even more on video. Scrub through the clip, not just the first second. The frame where the tracking slips is the frame where the blur was not applied.

Nothing is uploaded

Detection and blurring both run on your device. The models are fetched once and then cached, the image or video never leaves the browser, and there is no server that holds a copy of the thing you were specifically trying to make less identifiable.

That property is the whole reason to do this locally. A privacy tool that uploads your unredacted image to a server in order to redact it has, at the moment of upload, already done the thing you were trying to prevent.

A short checklist

Before you post:

1. Run the detector you actually need, and accept the download once.
2. Read the count. If it says none, place regions yourself.
3. Widen every face box past the hairline and jaw.
4. Delete false positives.
5. Choose pixelation if the redaction should be obvious, blur if it should be discreet, and use a large radius either way.
6. On video, scrub the whole clip and check the fast-moving parts.
7. Check the exported file at full size, not in a preview.

The tool does the tedious part. The judgement about what needs covering, and how much, stays with you, which is the correct division of labour for anything that is going to be published.