2026-09-10
Does blurring or pixelating a photo actually hide what's underneath?
A face dragged over with a blur brush. A license plate smudged into a mosaic before a car-sale photo goes up. A name pixelated in a screenshot before it's shared. All three feel finished the moment the sensitive part turns to mush — but "looks unreadable to a person" and "the information is gone" are two different claims, and only one of them is actually true of a blur or a mosaic. The other is true of a solid black box. Here's the real research behind why NearIMG's own redact tool defaults to the one that looks less clever.
The two things people mean by "hide it"
NearIMG's blur tool is really a redaction tool: drag a box over a face, a plate, an ID number, and choose how to cover it. It offers exactly two modes, and they are not interchangeable:
| Mode | What happens to the pixels | Can it come back? |
|---|---|---|
| Solid block (the default) | The covered pixels are deleted and repainted as a flat color. Nothing about the original region survives in the file. | No — there is nothing left to reconstruct. |
| Mosaic / pixelate | The region is divided into blocks and each block is replaced with its own average color. The original detail is blended, not erased. | Sometimes — see below. |
Why "sometimes" isn't a small caveat
Averaging a block of pixels together is lossy, but it isn't random, and for two decades security researchers have been demonstrating how much of that loss can be clawed back. The most direct case is text: Bishop Fox researcher Dan Petro built a tool called Unredacter that takes a pixelated word and recovers the original letters, by exploiting the fact that a known pixelation algorithm applied to a small, guessable alphabet (digits, standard fonts) doesn't destroy enough information to make the problem hard. Bishop Fox's own writeup on the project is blunt about the takeaway: "never use text pixelation to redact sensitive information" — their recommendation is a solid bar that actually deletes the covered content, not a filter that scrambles it.
Faces hold up even worse. A 2016 paper out of Cornell and the University of Texas at Austin, "Defeating Image Obfuscation with Deep Learning" (McPherson, Shokri & Shmatikov), trained a neural network to re-identify faces in a standard face dataset after they'd been run through mosaicing. The network picked the correct person over 95% of the time. Blurring — the soft, gradient kind of obfuscation, not blocky pixelation — held up better in their tests but still fell to 57% correct identification, on a task where random guessing would succeed only rarely. The paper's point wasn't that mosaicing or blurring is useless for every purpose; it's that both were being treated as "this information is now gone," when the honest description is "this information is now harder to read," which is a much weaker guarantee once the person trying to read it isn't a human squinting at a thumbnail.
So is mosaic ever the right choice?
Yes — the two modes solve different problems. If the goal is "it should stay visible that someone's face was there, and no reasonable viewer needs to recover exactly who," a mosaic reads correctly and looks less like a hole was punched in the photo. The failure case is specifically short, guessable, structured content: a license plate is 6–8 characters from a small alphabet, an ID number follows a known format, a phone number is ten digits — all of these have far fewer possibilities than a stranger's face, which is exactly what made them easy targets for both Unredacter and the 2016 face-recognition result. NearIMG's redaction panel puts this choice in front of you directly instead of picking silently: solid block is preselected, and choosing mosaic surfaces a warning on the spot — "Pixelation only averages the pixels, it doesn't remove them — short, guessable details like an ID number or a licence plate can sometimes be recovered from it." That's the same finding above, written as a decision you make before you save, not a footnote you find out about later.
Doing the redaction itself, without uploading the photo to do it
The obvious tension with "blur this because it's sensitive" is that most blur tools are a website you upload the sensitive photo to first. NearIMG's redact tool runs the whole thing — drawing the boxes, applying either mode, re-encoding the file — on your own device:
- Add a photo and open the redaction view; drag a box over each face, plate or ID detail (arrow keys plus Enter work too).
- Pick solid block or mosaic per region — you're not locked into one mode for the whole photo.
- Apply, then Save. The regions are painted onto the full-resolution image on your device, the same way as every other NearIMG tool: one photo downloads directly, a whole batch downloads together as a ZIP.
One side effect worth knowing about: because the download is a freshly re-encoded file built from the edited pixels, the camera metadata baked into the original — including GPS coordinates, if your phone recorded any — isn't carried over either. Redacting a photo and stripping its location data happen in the same step, whether or not that's the reason you opened the tool.
What this tool doesn't do
Being straightforward about the edges of it:
- No automatic face or plate detection. You draw every box by hand; nothing scans the photo for you. That's slower for a busy scene, but it also means nothing about the image is being analyzed anywhere except the region you actually cover.
- It can't undo a mosaic you already posted. If a pixelated photo is already public somewhere, redoing it locally with a solid block doesn't reach the copy that's already out there. This only protects what you're about to publish.
- Solid black is permanent by design, so double-check the box before you save. There's an undo while you're working, but once you've saved and closed the file, the covered pixels are gone the same way they'd be gone from a printed photo with a piece of tape over it.
The short version
A blur or a mosaic can look identical to a solid black box in a quick glance at a photo, but only the solid block actually deletes what was there. Real research backs this up in both directions: Bishop Fox's Unredacter tool reversed pixelated text back into real words, and a 2016 Cornell/UT Austin study re-identified mosaiced faces over 95% of the time with a neural network, against 57% for blurring on the same faces. Use mosaic when the point is "obviously covered, exact identity doesn't matter." Use a solid block — NearIMG's default — for anything short, guessable, or actually sensitive: a plate, an ID number, an address on an envelope. And do it on the device the photo is already on, since uploading a sensitive photo somewhere else to redact it defeats the point before you've even started.