Can you reverse a blurred or pixelated image?

It depends what was blurred, and the honest answer splits three ways. An out-of-focus or motion-blurred photo can be sharpened by AI, but only approximated, because a tool cannot recover detail the lens never captured. Text that was deliberately pixelated to hide a secret can often be reconstructed exactly, because pixelation is a fixed, solvable mapping. And a content-independent scramble, where each block is sampled at random from across the region, reverses to nothing, because no relationship to the original pixels is left to solve for. So unblurring a fuzzy photo and reversing a redaction are two different problems. The first is restoration, and it is lossy. The second is an attack, and it works well enough on short pixelated text that you should never pixelate a secret in the first place.

Three different questions hiding in one

People search "unblur an image" meaning wildly different things, which is why the usual answers are so muddled. One person has a treasured photo that came out fuzzy and wants it sharp. Another has a screenshot where someone pixelated a code and wants to read it. A third wants to know whether their own redaction can be undone. These are not the same problem, and they do not have the same answer. Here is each case, what is actually recoverable, and by what.

An out-of-focus photo: sharpen, do not expect a miracle

A blurred photo is the true sharp image run through a blur, which maths treats as a convolution with a blur kernel. Undoing that is called deconvolution, and doing it without knowing the exact blur is an ill-posed problem: many different sharp images could have produced the same blur, so there is no single correct answer to solve back to. That is why exact recovery of fine detail, the digits on a distant plate or the features of a small face, is not really possible. The information was lost when the shot was taken.

What the tools do instead is estimate. An AI upscaler is trained on millions of sharp-and-blurred pairs and generates a plausible sharp version that fits the blurry input. The output often looks great, and for a family snapshot that is exactly what you want. But it is a confident guess, not a measurement. Do not trust an AI-sharpened image as evidence, because the detail it added is invented to look right, not read off the original.

Pixelated text: yes, this really does reverse

This is the case where the answer is a clear yes, and it is the reason you should never pixelate anything secret. If you are here to work out how to hide text safely instead, read how to redact a screenshot on a Mac. For the mechanism itself, two tools tell the story.

Depix, 2020. Sipke Mellema showed that text pixelated with a linear box filter is solvable. In that kind of pixelation, each block is simply the average of the pixels underneath it, and the same input always produces the same block, so it is deterministic. Depix renders a De Bruijn sequence of characters in the same font, pixelates it the same way, and matches the resulting blocks against the redacted image to read the characters back out.

Unredacter, 2022. Dan Petro at Bishop Fox pushed it further with a recursive search: it renders a candidate string in the known font, re-pixelates it with the same settings, and compares block by block, extending the guess one character at a time. It is not magic. It needs the font, and it is sensitive to exact whitespace and rendering, so it is not a general photo tool. But it works well on precisely the short, structured strings people pixelate, like an account number or a two-factor code. Petro's advice was blunt: use black bars covering the whole text, and never use anything else, no pixelization, no blurring, no fuzzing, no swirling.

A note on plain blur, since people reach for it too. A Gaussian blur is not a stronger version of pixelation, it is a weaker protection. A low-radius blur can be estimated back, and the strong, proven attacks above are on pixelation. Either way, obscuring a secret is not the same as removing it.

A content-independent scramble: nothing to solve for

Both attacks above rely on the same weakness. Ordinary pixelation is a fixed mapping from the original pixels to the blocks, so there is something to solve. Break that mapping and the attack has nothing to work with. That is what a content-independent scramble does, and it is how ScrubShot's Scrub tool works.

Instead of setting each mosaic block to the average of the pixels beneath it, ScrubShot fills each block from a few pixels sampled at random from across the whole selected region. The block pattern is decoupled from whatever sat underneath, so it keeps the rough colors of the area and still looks like a redaction, but there is no per-block relationship to the original text for a solver to walk. It reverses to nothing by design. The scrub is baked straight into the image too, so the original pixels are gone, not hidden on a layer you could peel off.

So can AI unblur an image or not?

Both, depending on what you mean, and that is the whole confusion. AI is good at making a fuzzy photo look sharper, by generating detail that plausibly fits, so if you just want a nicer-looking picture, yes, it helps. AI is not what breaks a redaction, though. Pixelated text is recovered by matching against the known font, a deterministic search, not by an AI dreaming up characters. And nothing, AI or otherwise, recovers a content-independent scramble, because the information is not merely hidden, it was never kept. The practical takeaway is simple: to read a fuzzy photo, reach for an AI sharpener and treat it as a guess; to hide a secret so no tool can read it, do not blur or pixelate it at all.

FAQ

Can you reverse a blurred image?

It depends what was blurred. An out-of-focus or motion-blurred photo can be sharpened by AI, but only approximated, because the detail the lens never captured is gone. Text that was deliberately pixelated to hide a secret can often be reconstructed exactly, because pixelation is a fixed, solvable mapping. A content-independent scramble, where each block is drawn at random from across the region, reverses to nothing.

Can AI unblur an image?

AI can make a blurry photo look sharper by generating plausible detail from what it learned on millions of images, so it is a good-looking guess, not a true recovery. That is fine for a family snapshot and misleading for anything you need to be exact, like a license plate or a face. And it is not what reverses a redaction: pixelated text is recovered by matching re-pixelated guesses in the known font, not by an AI hallucinating detail.

Can a blurry photo be corrected?

Partly. Deblurring a photo is an ill-posed problem, meaning many different sharp images could have produced the same blur, so no tool can pick out the one true original. Modern AI upscalers and sharpeners estimate a likely sharp version and often look impressive, but they are inventing detail that fits, not restoring detail that was recorded. Treat the result as an approximation, never as evidence.

Is pixelating text safe for hiding it?

No. Pixelated text is reconstructable. Depix in 2020 and Bishop Fox's Unredacter in 2022 both recover the original characters by rendering guesses in the same font, pixelating them with the same settings, and matching block by block. It works best on the short, structured text people tend to pixelate, like account numbers and codes. Dan Petro's advice was blunt: use black bars, no pixelization, no blurring.

Can ScrubShot's scrub be reversed?

No, and that is the point of how it works. A normal pixelate sets each block to the average of the pixels beneath it, which is the fixed mapping Depix and Unredacter solve. ScrubShot instead fills each block from a few pixels sampled at random from across the whole selected region, so the pattern holds no relationship to the text underneath. There is nothing to solve for, and the scrub is baked into the pixels, not sitting on a removable layer.