Skip to main content

Unblur

Unblur a photo: remove the blur, keep the detail

Unblurring removes a blur the file still holds rather than turning up the sharpness. Movement during the exposure displaced the detail along a direction, so it can be pushed back: three to eight pixels of smear returns. Forty pixels is a different face, and no tool returns the original.

Try it on one smeared photo

First photo free — no card required

The four numbers that matter

Working range
3–8 px of smear
Handheld floor
1/focal length — 1/150 s on a phone
Children and pets
1/250 s and 1/500 s
Past 15 px
Plausible invention, not recovery
Cost
First free, then $1.50 a photo

Unblur first, then work out which blur you had

Unblurring means the same thing in every case: put back detail the file still holds but is no longer showing. What changes is why the detail left. Movement during the exposure displaces it along a direction — that is the case this page is built around, and the only one where recovery can be close to complete. Defocus averages the detail away instead of moving it, and compression or downscaling deletes it outright. Those are the repair-and-enlarge pass. If you do not yet know which one you are looking at, thirty seconds at 400% zoom settles it.

Three failures, three pages: movement is here, defocus and compression are on the repair page, and the diagnosis that separates them lives in the guide.

A moving subject is not a soft lens, and the fix is not sharpening

While the shutter is open, a sharp point of light does not stay a point. It travels, and the sensor records the whole path as a line. The photograph is the scene convolved with that path — the path is the point spread function, and every pixel in the frame has been smeared by the same one if the camera moved and the world stood still.

Recovery is therefore a deconvolution problem, not a sharpening problem. Sharpening raises contrast across an edge that is already in the right place; deconvolution takes energy that has been displaced along a known direction and puts it back where it came from. That is why it can produce something an unsharp mask cannot, and why it fails in a way an unsharp mask does not: it needs the path to be short and uniform.

In a real photograph nobody hands you the path, so it has to be estimated from the image. That is blind deconvolution, and the estimation is the fragile part. A straight smear against a static background is the easy case. A camera that rotated during the exposure, or a subject walking toward the lens, gives every region a different kernel — and a single inverse filter cannot undo a family of them.

Three different failures that all get called "blurry", and what recovery means for each

What happenedWhat the file containsWhat recovery can do
The subject moved, camera steadyThe subject smeared along its own path; the background is sharpDeconvolve the smeared region. Needs a mask, because the kernel differs from the background
The camera moved, world steadyEverything smeared by one kernel; no sharp reference anywhereDeconvolve the whole frame — the standard case, and the one that works best
Focus missed the subjectA disc of confusion, not a line. Nothing displaced, only averagedNot this problem. That is defocus, which has its own ceiling

The distinction matters because the three cases have different verdicts. A sharp background behind a smeared subject is good news: the kernel is estimable and the smear is isolated. A frame where everything has moved the same way is also tractable. A frame where the subject is soft for two reasons at once is usually a write-off.

How far things actually move: the arithmetic that decides everything

Displacement in pixels is focal length in pixels × subject speed × shutter time ÷ distance to subject. Focal length in pixels is not the lens figure — it is the actual count on the sensor, roughly the frame width in pixels times the crop factor. On a phone main camera at 26mm equivalent with 4,032 pixels across, that is about 2,650 pixels of focal length.

The classic rule of thumb — one over the focal length in seconds — falls out of the same equation. A 50mm lens on full frame wants 1/50 s or faster for a still subject, because that keeps hand tremor under about a pixel. A phone at 26mm equivalent is not safer just because the number is smaller: its sensor is small, so the same angular wobble covers more of the frame. 1/150 s is the realistic handheld floor.

The table below is that equation evaluated for the subjects people actually photograph. Read the columns as verdicts, not as advice: at 1/30 s a walking adult is past recovery before any software is involved, and no amount of processing changes that.

Smear in pixels by subject and shutter, phone main camera at 26mm equivalent

SubjectSpeed1/30 s1/125 s1/250 s1/500 s1/1000 s
Adult walking, 3 m1.4 m/s41 px10 px5 px2.5 px1 px
Toddler, 2 m1.8 m/s40 px10 px5 px2 px1 px
Child's hands, 2 m2.0 m/s44 px11 px5 px3 px1 px
Footballer, 6 m7 m/s51 px12 px6 px3 px1.5 px
Car at 50 km/h, 15 m14 m/s41 px10 px5 px2.5 px1 px
Bride walking, 4 m1.2 m/s26 px6 px3 px1.5 px1 px

One caveat that makes the numbers optimistic: they assume shutter speed is the only limit. At 1/500 s indoors, the sensor is underexposed enough that noise removal will destroy more detail than the smear ever would. In low light the real choice is between a grainy sharp frame and a clean smeared one, and the sharp frame almost always wins.

What comes back, what does not, and where inference becomes fabrication

Under about three pixels of smear the photograph is simply sharp enough; you are looking at noise and compression, not at movement. Between three and eight pixels deconvolution is genuinely effective: the neighbours have not yet overlapped, the kernel is estimable, and the recovered edge is the original edge.

Between eight and fifteen pixels it becomes reconstruction. The algorithm still produces something that reads as sharp, but it is resolving ambiguity with priors rather than recovering displaced energy. At this point the answer stops being photography and becomes illustration — acceptable where nobody compares the result to the truth, unacceptable anywhere the truth matters.

Past fifteen pixels and any fixed threshold you care to name, the question changes. A face that moved forty pixels no longer contains that face; what the model returns is a face shaped by its training data, positioned where the smear was. It will look like the person. It is not a photograph of them.

Recovery verdicts by smear length

SmearWhat is actually happeningVerdict
0–3 pxBelow the resolution of the detailSharp. Fix noise, not movement
3–8 pxEnergy displaced but recoverableDeconvolution returns the original edge — this is the working range
8–15 pxAmbiguity exceeds evidenceReads sharp, partly invented. Fine for a wall print
15–40 pxFeatures dragged into each otherIllustration. Do not present it as a photograph of anything
40 px and beyondThe subject is a smearNothing to recover. Reshoot, or accept a different image

The hard cases are text, number plates and identifying detail, because the reader can verify them. A drawn eyelash is invisible at print size; a drawn digit on a licence plate is a false statement. If the output will be used to prove something, smear past eight pixels disqualifies the file.

By subject: what each one actually needs

The same equation produces different advice because the subjects differ in speed, in distance, and in how much the viewer can tell. These are the five cases that account for most of the smeared photographs people upload.

Children

Fast, close, and usually indoors. A toddler covers about 1.8 metres a second and rarely holds still for a pose, so 1/250 s is the floor and 1/500 s is comfortable. Because they are close, the distance term works against you as well — halving the distance doubles the smear.

Recovery: often good, because a child photographed at 1/125 s indoors is typically 5–10 px — the top of the working range.

Pets

Faster than children and closer to the ground, where the light is worse. A dog at full speed is 8–10 m/s; at 1/500 s and three metres that is still 20 px. Catching them at rest is not a processing problem, it is the only reliable answer.

Recovery: poor for action, good for a head turned at 1/250 s.

Sport

The subject is far away, which helps, and fast, which does not. 1/1000 s at f/2.8 in daylight is the standard for a reason. Panning — tracking the subject so it stays still in the frame — smears the background instead, which is usually the better trade and makes the subject sharp in-camera.

Recovery: a panned shot is sharp on the subject; do not process it. A missed pan is unrecoverable.

Weddings in low light

The hardest case, because the light is deliberately dim and the moments are unposed. A bride walking up an aisle at 1.2 m/s in 1/60 s light is 13 px — past recovery, but only just, and it is the single most common upload in this category.

Recovery: marginal. Anything shot at 1/125 s or faster is comfortably recoverable; 1/60 s is a coin toss.

Street photography

Usually intentional. A blurred passer-by is a compositional choice, and deconvolving it removes the reason the frame exists. If the blur is not deliberate, the cause is normally a slow shutter chosen for a dark alley rather than a missed focus.

Recovery: possible but often wrong to do. Decide whether you want the photograph you took.

Preventing it next time

Shutter speed is the only lever that fully solves this, and the table below is what to set. Everything else is a workaround.

Shutter speed by subject, for a handheld camera

SubjectMinimum shutterWhy
Landscape on a tripodAnythingNothing moves, and the tripod removes the camera from the equation
Landscape handheld, wide lens1/125 sHand tremor, not subject motion
Adults posed1/125 sSmall movements, comfortable margin
Children1/250 sFaster, closer, less predictable
Pets1/500 sFast and low to the ground, where the light is poor
Sport1/1000 s7 m/s and beyond
Anything indoors in the evening1/125 s and raise ISOA grainy sharp frame beats a clean smeared one

On a phone, four things actually work: burst mode and keep the sharpest frame, tap-to-focus before the shot rather than after, optical zoom instead of digital (a 2× crop keeps every pixel; an 8× crop is upscaling with extra steps), and holding the shutter to lock exposure so the camera does not slow the shutter to brighten a dark scene. Turning on the flash in a dark room is the fifth, and it is the one people refuse to use.

Where this page ends

This page owns one question: a subject that moved during the exposure. If the photograph is soft because focus missed, the physics are different and the ceiling is lower — that is deblurring defocus and compression damage. If you do not yet know which one you are looking at, the thirty-second diagnosis settles it with a single edge at 400% zoom.

If the file is short of pixels rather than smeared — a small print you want at A3 — the question is print size and dpi, not recovery, and upscaling without inventing detail is the page that answers it with numbers.

Motion blur, answered

Can a photo be unblurred if the subject moved?

Partly, and it depends on distance rather than on how bad it looks. A smear of three to eight pixels is recoverable because neighbouring features have not yet been dragged into each other. Past about fifteen pixels the face is a smear and anything sharp afterwards was drawn by the model.

What shutter speed stops motion blur?

Children need about 1/250 s, sport 1/500 to 1/1000 s, and a handheld phone about 1/150 s because its small sensor sweeps a wider angle per hand movement. The rule is one over the focal length in seconds; the caveat is that it assumes a still subject.

Can motion blur be removed completely?

No. Deconvolution reverses a known path, and the path is only estimable while it is short and uniform. A curved path, a rotating camera, or a subject moving toward the lens gives every pixel a different kernel, and no single inverse filter recovers that.

Does the tool know which way the subject moved?

It estimates the point spread function from the image itself, which is why the technique is called blind deconvolution. A straight smear across a static background is the easy case. Two subjects moving in different directions in one frame is not solvable by any single-kernel method.

Why are photos of children always blurred?

A toddler moves at roughly 2 metres a second, and indoor light rarely allows faster than 1/60 s. At three metres that is about 30 pixels of travel on a phone sensor — far past recovery. The fix is shutter speed and light, not post-processing.

Is this better than a deblur filter in an editor?

They are different operations. An unsharp mask raises local contrast at edges, which fakes acuity and does not move any detail. Deconvolution redistributes smeared energy back toward a point, which is real recovery while the path is short. Both are worse than a second exposure.

What happens if the subject moved more than the tool can fix?

You get a plausible reconstruction and should treat it as one: fine for a wall print, not fine for evidence, insurance claims, identification, or an archival record of what the negative contains. The tool cannot tell you where that line is; the pixel count can.

Does it cost more to unblur a photo?

No — one credit per render, the same as every other mode: $1.50 singly, from $0.70 in the 100-credit pack, or from $12 a month annually. The first photo is free, which is the cheapest way to see what your smear length actually allows.

Measure the smear before you spend a credit on it.

Find a high-contrast edge, zoom to 400%, and count the pixels the tail runs. Under eight, upload it. Over fifteen, the honest answer is a different photograph.

Try it on one smeared photo