Can you unblur a photo when the dog crossed the frame

A dog that crossed the frame smeared more than one pixel. Unblur the photo and look at the smear before you trust the fur.

My aunt sent a park photo with five words in the text: can you unblur a photo. Her black mixed breed, white chest flashing, was running left to right on a gravel path. The shutter stayed open long enough that his face landed in three places. I opened the file before I answered, because the question sounds like a yes or no and the pixels are a measurement.

The JPEG is 4032 by 3024 pixels and 2.1 MB. EXIF still reports one fortieth of a second, f/1.8, and ISO 80. Noon light, a phone, a dog who does not pose. I marked the leftmost nose highlight and the rightmost one. The travel is about 110 pixels. Between those ends I can see two more bright peaks, so the head is not a smooth smear. It is three exposures stacked in one frame.

TL;DR. The dog's nose travels about 110 pixels and the head is recorded three times, so there is no single face to recover. I can steady the gravel, the leash, and a little of the torso. I will not collapse three exposures into one invented snout. The file is a 4032 by 3024 JPEG. A mild pass is printable at 5 by 7 inches. Frame three of the burst has one head and is the file I would frame. I measured the streak before I promised anything.

What the streak is made of

A running dog is not a rigid box sliding sideways. The skull moves, the ears swing on a different arc, and the legs cycle faster than the chest. In this frame the chest smear is about 40 pixels. The front paws are closer to 150, and they leave a fan of toes rather than one paw. The tail is a gray ribbon about 70 pixels long with no readable tip.

That split matters. A deblur pass wants one description of how the camera and the subject moved. I can write one description for the torso and get a slightly tighter shoulder. I cannot write one description that is true for the paws and the skull at the same time. When I tried, the shoulder sharpened and the paws turned into extra claws.

The background tells me the camera was fairly still. The park bench about 40 feet back has slats I can count: five horizontal boards, edges soft by maybe 2 pixels, which is normal for a phone at that distance, not a pan. The gravel under the dog is almost sharp. If the whole frame had been jerked, the tool and the gravel would streak with the dog. They do not. The blur belongs to the animal.

I also checked vertical edges on the torso. Horizontal motion should spare a vertical line more than a horizontal one. The torso's vertical edges are still soft by about 6 pixels. So there is a second problem under the streak: the phone focused on the path a foot in front of the dog. Motion and missed focus are both in the file. Fixing only one of them leaves the other visible at a large print.

Three peaks are not one face

I opened the file in a viewer that reports the cursor position and I walked the nose highlights. The peaks sit near x of 1860, 1945, and 2010 on the 4032-wide frame. They are not evenly spaced. The last gap is shorter, which fits a head that slowed as the body stretched into the next stride. Even spacing would have looked more like camera shake. Uneven spacing looks like a subject.

Each peak has a hint of an eye, none of them round. The middle one is the least awful. If I crop tightly and squint, I can pretend it is the face. At 100 percent it is a bright wedge with a dark notch, and the notch is split. Relatives recognize a dog from the whole shape, the chest flash, the collar. They do not recognize him from a reconstructed iris. I told my aunt that directly.

The collar tag is a bright blob about 14 pixels across. No letters. A model that "reads" a tag here is guessing a phone number. I will not ship a guess that looks like contact information. The collar itself is a dark band about 9 pixels tall, and it survives as a band. That is enough to show he was wearing it.

Ears are the trap. The near ear flaps upward and leaves a thin dark flag above the skull. A deblur that assumes the head moved only sideways pulls that flag down into the skull and grows a lump. I ran that mistake once on a copy, compared it to the original, and deleted the copy. The lump was sharper than the real ear and therefore more convincing, which is the failure mode I watch for.

What I will and will not hand back

I cropped a 900 by 700 region around the dog so the pass was judging the animal and not the sky. A full-frame pass spends its capacity on leaves. Leaves were already fine. The unblur pass on that crop tightened the shoulder and the leash. The leash is a pale diagonal about 2 pixels wide, and it stayed a line instead of breaking into dashes. That was the success.

The head did not improve in a way I would sign. The three peaks moved closer by a few pixels and the middle notch got darker, which reads as an eye if you want it to. I dragged the original across the result and the "eye" had no match in any single peak. It was an average. Averages of faces look plausible and belong to nobody. I rejected the head region and kept the original pixels there, pasted back with a soft edge about 30 pixels wide so the seam would not show at arm's length.

The path needed nothing. At a 5 by 7 inch print the 4-pixel gravel smear is smaller than the dot of the printer. I checked that by dividing. 5 inches across 4032 pixels is tiny, but I was not printing the full frame edge to edge. The crop I printed put the dog at about 4 inches tall. His 110-pixel nose travel becomes a streak you can see, and you should see it, because it is what the shutter recorded. Hiding it would be a different picture.

When the softness is a missed focus with one subject position, I use the blurry photo repair rather than treating every soft file as motion. This park frame is the mixed case, so I used the motion-aware pass only on the torso and I stopped. Mixing both tools on the head, one after another, was worse. The second pass treated the first pass's ghosts as edges and outlined them.

Checks I make before I promise a print

I keep these on a card above the desk. They are the difference between "I will try" and "I will frame this."

• Read the shutter time and the ISO before judging the smear by eye.
• Count distinct highlights of the same feature, such as a nose, a collar tag, or a single eye.
• Measure the streak in pixels on the longest travel and on the torso separately.
• Look at a still background edge. If it streaked too, the camera moved, and the kernel may be shared.
• Find every burst frame before editing the worst one.
• Decide the print inches, then decide whether the streak is visible at that size.
• Reject any new eye, tooth, or tag letter that has no counterpart in the original.
• Keep the original JPEG beside the attempt, with the same filename plus a suffix, and do not overwrite.

The burst saved this job. Four frames, less than a second apart. Frame one is the triple head. Frame two clips the tail. Frame three shows one head, side-on, with the near ear cut by the frame edge. Frame four is a rump and a path. Frame three is the portrait. The cut ear is a crop problem, and a crop problem is honest. I can print frame three at 5 by 7 with the ear ending at the border, and my aunt knows the ear continued outside the picture. She cannot know which of the three noses was the "real" one inside frame one.

A table of the regions I measured

I wrote this down so I would not talk myself into a full-frame miracle after a promising shoulder.

Region	Travel in pixels	What I delivered
Park tool slats	about 2	Left untouched
Gravel under the dog	about 4	Left untouched
Leash	about 15 along its length	Kept the 2-pixel line
Torso and shoulder	about 40	Mild tighten, then stop
Head and nose	about 110, three peaks	Original pixels pasted back
Front paws	about 150	Original, streak visible
Collar tag	blob of 14, no letters	Shape only, no invented text

The table is the whole job. Most of the pixels I was asked to fix are in the bottom rows, and those rows say original. The rows I could help are the ones a relative mentions second, after the face. I say that out loud before I start, so the face is not a surprise disappointment at the end.

A drawing I use when I explain the limit

People hear "blur" and picture a slider. I keep a generated drawing on the tool for that conversation. It is not a dog and it is not a person. It is flat shapes: a round head, a teal rectangle, a panel with a yellow circle. When I blur that drawing, every edge gets a single soft disk, the same in every direction. That is focus blur, or a demonstration of it. The park dog is not that pattern.

The drawing is useful because the softness is even. I can show my aunt a round head that was one head, then a soft version of that same head. Her file does not look like the soft drawing. Her file looks like three heads sharing a neck. If a tool makes the drawing's yellow circle round again, that only proves the tool can undo an even disk on a simple shape. It does not prove it can pick among three noses.

I also keep the sharp version of that drawing in the same folder so I remember what "recovered" would mean. Recovered would mean the circle comes back as a circle because the circle was in the source and the blur was applied after. The dog's face was never a single sharp face inside this exposure. There is no sharp original hiding under the streak. The sharp originals are the other burst frames, and one of them is missing an ear at the border.

Focus blur, motion, and the JPEG on top

Why the vertical softness changed the pass

The 6 pixels of vertical softness change how hard I push. A motion model that only streaks sideways will leave a halo above the back, a light line where it tried to pack the blur into a sharper spine. I saw that halo on the test copy, about 2 pixels wide, brighter than the fur by a step I could see on a gray card crop. Halos are a reason to stop, same as they are on a product shot. I threw out any pass that drew a bright rim along the back.

JPEG adds a third texture. The file is 2.1 MB for 12 megapixels, which is a moderate squeeze, not a ruined one. Along the streak, the blocks are 8 pixels and they follow the smear, so the ghosts have square shoulders if you zoom in. A pass that "sharpens" those shoulders carves the compression into the fur. I judged the result at 100 percent and then at the print size. At print size the blocks vanish. At 100 percent they tell me not to add edge contrast.

Color stayed put, which I was glad of. The white chest did not pick up a green fringe. The grass did not bleed into the black back except where the motion itself mixed them, a band of about 8 pixels along the outline. That band is optical, recorded by the sensor, and I left it. Cleaning it to a razor edge would invent a contour the shutter never saw.

Noise was low. ISO 80 at noon does not speckle. That meant I did not need a denoise before the deblur, and I did not want one. Denoise on a streak softens the very peaks I was using as landmarks. I needed those peaks to stay so I could see whether the pass moved them.

Print size, viewing distance, and the note I sent

I print test strips rather than arguing in the abstract. On a 4 by 6 inch strip of the triple-head crop, the motion reads as motion. You see a running dog. You do not count noses unless you look for them. On an 8 by 10 the three peaks separate and guests ask if the printer glitched. I stopped at 5 by 7 on a matte sheet. Matte hides a little of the JPEG edge. Gloss would have outlined it.

Viewing distance for a 5 by 7 held in the hand is about 12 to 14 inches. At that distance my eye does not resolve the 6-pixel focus softness on the torso after the mild pass. It does resolve the 110-pixel head travel, because that travel is a large fraction of the head. The math and the strip agreed. I like it when they agree. When they do not, I trust the strip.

The note I sent with the files was three lines. Streak at the nose, about 110 pixels, three peaks. Pro I will stand behind, 5 by 7, from burst frame three. What I will not draw: eyes, tag text, a single merged muzzle. My aunt answered with a photo of frame three on the fridge. Frame one stayed in the shared album with no edit. She liked seeing how fast he was.

A walking dog, so this does not become every motion file

Last winter I had a beagle on a sidewalk, same aunt, different problem. The nose moved about 18 pixels. One peak, not three. Shutter was one one-hundredth of a second, and the dog was walking, not running. One kernel fit the head and the body well enough. I delivered an 8 by 10. The collar tag still had no letters, and I still did not invent any. The difference was the count of peaks and the length of the streak, not the species.

I keep both numbers in the same notebook so I do not tell the next person that motion blur is hopeless. Eighteen pixels with one peak is a repair. One hundred and ten pixels with three peaks is a choice of frame. If someone only has the bad frame, the honest output is a smaller print where the streak reads as speed, plus a sentence about what was not recovered.

Video is sometimes the better source, and it was not available here. A phone video at 30 frames a second would have given me exposures near one one-thousandth if the shutter was fast, or smeared frames if it was not. I ask for the video when it exists. I do not ask a still model to pretend it saw the frames the camera did not save. On this afternoon the burst was the video, four frames, and frame three won.

How I stored the attempts

The folder has the original JPEG, the four burst frames exported before any gallery "enhance" could touch them, the rejected full-frame pass, the torso-only pass, and the 5 by 7 print file. The print file is a TIFF at the print pixels, 1500 by 2100, which is 300 pixels per inch at 5 by 7. I did not upscale the triple head to get there. I cropped and placed frame three on a canvas of that size. Upscaling the ghosts would have made larger ghosts.

Names are dull on purpose. park-dog-burst-03-print.tif sits next to park-dog-burst-01-original.jpg. If I am hit by a bus, my aunt's next helper can see which file was the camera and which file was the print. I have opened too many family folders where the only survivor was a sharpened export and the phone original was gone.

I also wrote the measurements into a text file in that folder: shutter, pixel travels, the x positions of the three peaks, the print size, and the sentence I refused to cross. The next running-dog job will not match these numbers. The habit of writing them down is what transfers.

What the search phrase is asking for

People type can you unblur a photo because a button in another app said the picture could be fixed, or because a friend did fix a different picture. The phrase hides the measurement. "A photo" might be a portrait focused an inch too far, a license photo in shade, or a dog in three places. Those are three jobs. I can often help the first. I can sometimes help the second if the photo was large enough and I am allowed to say when the characters are not really there. I cannot promise the third.

If you only have frame one, say so, and ask for a print size where the motion still looks like motion. If you have the burst, send all of it, including the frames you think are worse. Frame three looked worse to my aunt because the ear was cut off. It was the only frame with one head. I would rather crop an ear than invent a face.

The longer note on whether a soft file can come back at all is is it possible to unblur a photo. When the blur is a single smear and you want a sequence, I wrote how can I remove blur from a picture. When the kernel really is one motion and the subject held its shape, the steps live in how to deblur a picture. Read those against your own streak length. If you count three noses, you are still in this essay, not in those.

Ears, paws, and the chest flash

I spent an extra half hour on the chest flash because it is how the family recognizes him in small prints. The white patch is about 80 pixels wide in frame three and it is clean. In frame one it is stretched to about 130 pixels and it overlaps a shadow from the leg. A pass that "corrects" frame one's patch to the frame three width is copying a different moment. I did not do that. Recognition at 5 by 7 comes from frame three's patch, which was recorded once.

Paws were a request. She wanted to see the nails. The nails are a row of specks inside the 150-pixel fan. I can see four specks in one cluster and three in another. I cannot assign them to a single paw position without dropping two of the clusters. Dropping them would make a tidy paw that is a lie. The print shows the fan. At 5 by 7 it reads as speed, which is what a running dog is.

The far ear in frame three is fully inside the frame and it is sharp enough to show the notch where he has a small tear from an old scrape. That tear is about 6 pixels. It survived because I did not denoise and I did not sharpen hard. I checked it at 100 percent against the original before I made the print TIFF. Small scars are how you know the pass behaved. If the tear had healed itself, I would have rejected the file.

What I would do on a reshoot

I told her this for the next park day, not as a scolding. Shutter at one five-hundredth or faster if the phone lets her lock it. Burst on. Stand so the dog runs across the light, not into a bright sky that fools the focus. Fill more of the frame with the dog if the face matters, but not so close that one ear is doomed at the border. Frame three was a good picture that arrived by luck. A locked shutter would have made frame one good too.

If the phone will not lock the shutter, a burst still helps, because one of the frames often catches the feet at a turnaround when the head is slower. I have seen that in other jobs: the streak collapses for a single frame when the subject changes direction. You only get that frame if the burst was already running before the turn.

None of that repairs the JPEG we already have. It prevents the next one. I keep the advice in the same message as the limit so the limit does not sound like a shrug. I measured this file. I printed the frame that could hold a face. I left the triple exposure alone.

The line I will not cross

I can unblur a photo when the subject occupied one place and the blur is a smear around that place. I can sometimes unblur a photo when the smear is short, under a couple dozen pixels, and the feature I care about has one peak. I cannot unblur a photo of a head that visited three places during the exposure and hand you one of those places as if the sensor had chosen it.

The park file is on the far side of that line at the face, and on the near side of it at the leash and the shoulder. Splitting the file by region is the whole craft here. A single slider has to be right for every region at once, and this frame will not allow that. I would rather mail a 5 by 7 of burst frame three than a sharp fiction of frame one.

She still has the running picture. It shows a dog who was fast, a path that stayed put, and a collar with no readable tag. That is what the camera kept. My work was to notice which frame had one face, to keep my hands off that face, and to write down the 110 pixels so the next request does not start from hope.