How to unblur an image (and when it is simply gone)

More people search for how to unblur an image than how to blur one. That tells you something about both jobs.

The honest answer has two halves. If a photo is soft because the camera missed focus or the file was shrunk, sharpening and upscaling can make it look better. If someone deliberately blurred a name or a code, you are usually not recovering it — and the app promising otherwise is inventing plausible pixels and calling it restoration.

What do unblur tools actually do?

Almost all of them do one of three things. They sharpen, which raises local contrast so edges read crisper. They upscale, which fills a bigger canvas with a model’s best guess at detail. Or they deconvolve, which tries to reverse a known blur kernel — real maths, and the only one of the three that is genuinely undoing anything.

Deconvolution needs the blur to be mild and fairly uniform. A gentle motion smear on a licence plate is its best case. A heavy Gaussian dumped over a phone number is not, because the information is not hiding in there faintly. It has been averaged away.

Why does a generative “AI unblur” look so convincing?

Because it is not reading your image. It is drawing a new one that would be consistent with the blur it sees. Feed the same blurred face to two of these tools and you often get two different people, both sharp, both confident. That is the tell.

It matters if you are trying to verify something. A generated digit is not evidence the original digit survived. Treat the output as an illustration, not a recovery. If you would act on it — call the number, accuse the person, file the report — you need the original file, not a model’s guess.

So when can blurred text really be read back?

When the text is short, the alphabet is small, and the blur is light. Researchers demonstrated this in a 2016 PETS paper on mosaicing and blurring as document redaction: the structure left behind is often enough to narrow the candidates dramatically. Six digits from a set of ten is a small search space.

A paragraph of body copy under a thick smear is the opposite case, and usually stays gone. So the risk is not evenly spread. It concentrates exactly where you care most — codes, PINs, card numbers, short surnames. We wrote that up in the PETS 2016 result.

Should you upload an image to an unblur site?

Think about what you are handing over. The picture you want unblurred is, by definition, a picture with something interesting in it. You are uploading the interesting part plus everything around it: the chat header, the timestamp, the account name in the corner.

For a holiday snap that came out soft, fine. For a document, an ID, a bank screen, or anything from work, no. Free unblur sites are free because the file is worth something to somebody. There is no reason to assume the copy is deleted just because the page said so.

How to make a blur nobody can undo

Stop relying on blur for the sensitive bits. A filled black bar replaces the pixels rather than smearing them, and once that is baked into a flat PNG there is no residual contrast for anyone to stretch. Nothing to deconvolve, nothing for a model to reconstruct.

Use blur where the goal is “not identifiable at a glance” — a bystander’s face, a background whiteboard. Use a bar for anything you would change or cancel if it leaked. The comparison in full is in blur versus bar.

A quick self-check before you send

• Zoom to 100% and look. Is the text gone, or just faint?
• Raise brightness and contrast. Does anything come back?
• Check you are attaching the exported PNG, not the original still sitting in Recents.
• Look at the edges of the frame — headers, notifications, the URL bar.

If a shape reappears when you push the brightness, that was a tint, not a redaction. Cover it again with a bar and export a fresh file.

You can do all of that here without uploading anything: open the blur tool, cover the part that matters, and download a PNG that has nothing left to recover.