Blur Images

What a blur actually destroys, and when it does not

A blur replaces each pixel with a weighted average of its neighbours. Done strongly enough it genuinely destroys information — there is no hidden layer underneath and no filter that reverses it, because many different originals average to the same result. Done weakly, it does not, and that is the part worth being careful about. A light blur over text is reversible in practice: the shapes are still statistically present, and researchers have reconstructed pixelated and lightly blurred characters by generating candidate text, blurring it the same way, and matching. Published work has recovered account numbers and names from screenshots people considered redacted.

So the rule is about strength rather than method. For anything made of characters — a bank number, an address, a licence plate, a name in a screenshot — the safe setting is coarse enough that the region carries no visible structure at all: not a soft smear where you can still see where the letters were, but a flat patch of a few blocks. A good check is to look at the result at normal size and ask whether you could guess the number of characters. If you can, it is not strong enough.

Pixelation and blur destroy information in different shapes. Pixelation averages each block down to a single colour, which at a coarse block size removes more than a blur of similar visual weight and also *reads* as deliberate redaction to whoever sees it — useful when you want it obvious that something was removed. A blur looks softer and more natural, which is right for a face in the background of a photograph and wrong for a document.

Everything outside the regions you mark is copied through untouched, and the file is written back in its original format at full quality — so blurring one corner of a photograph does not re-compress the rest of it. The result is flattened into the pixels, with no record of what was there, which is the point: keep your original somewhere if you may need it.

When you want something else

For a picture destined somewhere public, the blur is only half the job. A photograph carries its capture location, time and camera in its metadata, so a blurred house number tells nobody anything while the GPS coordinates in the same file tell them everything. Re-saving here strips that, and exiftool -all= photo.jpg is the way to be certain on a file you have edited elsewhere.

For faces at volume — a street photograph, a conference crowd, CCTV footage — marking boxes by hand does not scale. Tools that detect faces and blur them automatically exist and are free: deface does exactly this for images and video in one command, and ffmpeg can blur a tracked region across a clip. Both are the right answer past about a dozen faces.