Blur faces and anything else, on your device VEIL

It works at full resolution, which is the whole trick

The photograph you see on screen is a preview, scaled to fit. The boxes you draw are mapped back onto the original pixel grid, and the pixelation is applied there, at whatever resolution the camera actually produced. The export is the full-size image with the covered areas destroyed in it.

The common way to get this wrong is to redact the preview: draw a box on a 900-pixel-wide version, export, and hand over an image where the covered area was 900 pixels wide and the detail is gone from the copy but not from the source. Worse is the version where a viewer draws a black rectangle as an annotation, which is a separate object sitting on top of the picture and can be moved aside by anyone who opens the file properly. Neither happens here: the pixels are rewritten.

Nothing is uploaded, and there is a practical reason beyond principle. The photographs people need to redact are photographs of children, of injuries, of documents, of number plates outside a house. Those are exactly the files that should not become an HTTP request to somebody else's server, however good their intentions and however clear their retention policy.

Face detection finds most faces, and you check the rest

The one-tap detection runs a small model in this tab and puts a box on each face it finds. It is a convenience for a group photograph, not an authority. Detection misses faces in profile, faces partly behind something, faces at the edge of the frame, faces that are small, faces in poor light, and it has been consistently less reliable on darker skin across the whole history of the technology.

So the workflow is: let it find what it can, then look. Every face it missed is one you draw a box over yourself, and the tool is built for that because drawing is the primary action and detection is the shortcut.

Then widen your idea of what identifies somebody. A face is the obvious thing and frequently not the thing that matters: a name badge, a number plate, a house number, a screen with an email client open, a school uniform, a distinctive tattoo, a reflection in a window, a prescription label, a delivery note on a worktop. Anything you can draw a box over can be covered, and the box is the same box.

Pixelation, and what it leaves behind

The block size is yours to set, and larger is safer. Pixelation keeps a coarse average of what was underneath, and where the set of possible originals is small, an attacker can render each candidate, pixelate it the same way, and look for the match. That has been done to recover redacted number plates and short strings from published images. Against a face in a crowd it is not a realistic attack; against a six-character plate or a postcode it is a weekend.

If the information must be gone rather than obscured, cover it solidly: SCREENSHIELD does that for a screenshot and RECAST for a video.

Last, the export is a new image built from pixels, so the original's metadata does not carry into it. That is worth knowing in both directions: it means the GPS coordinates and the camera serial do not travel with the redacted copy, and it means the redacted copy is not the original if you needed the original preserved. EXIF shows what the source was carrying and CLOAK strips it without altering the picture.