Tools

Unblurring a Photo: Five Methods, Measured on the Same Image

By Ayush Patnaik, founder of AccioPix

Published September 10, 2026 · Updated October 1, 2026 · 12 min read

You can unblur a photo when the blur is camera shake or subject motion, partly when it is mild soft focus, and not at all when the focus was badly missed. We tested that instead of asserting it: one sharp photo, blurred by six known amounts, then run through five unblur methods and compared with the original. On a 16-pixel camera shake the AI deblur model lifted the quality score by 2.0 dB (from 23.6 to 25.6); on a 3-pixel soft focus a deconvolver lifted it 1.5 dB (24.8 to 26.3); on a 6-pixel soft focus nothing managed more than 0.4 dB. The table, the crops at 100%, the files and the script are all below, and so is how to unblur a photo yourself, free.

Unblurring a photo: what comes back

You can unblur a photo to the extent the detail is still in the file. Blur is the camera spreading each point of the scene over a patch of pixels: a streak when the camera or subject moved, a soft disc when the lens focused somewhere else. Spreading is not deleting. A streak mixes neighbouring pixels in a known direction, and software can partly unmix them, which is why a shake-blurred licence plate can come back readable. A wide focus disc mixes so many pixels in every direction that the fine detail is averaged into nothing, and there is no unmixing it; a tool can only draw plausible edges where it guesses they were.

That is the entire answer. Every number below is an example of it. The rest of this page shows how much comes back, for each blur, by each method.

Fixing a blurry picture starts with knowing which blur you have

To fix a blurry picture, you must know which of three blurs it is. Each kind needs a different fix. Applying the wrong one makes it worse. Open the photograph at 100%. Look for a point of light, a catchlight in an eye, or a hard edge.

• Camera shake: every dot smears into the same short line, across subject and background. It is common in dark rooms with slow shutter speeds. This blur is the easiest to recover.
• Subject motion: The background stays sharp while the moving person blurs. It is recoverable when the streak is short. The AI model fixes it without being told which part moved.
• Soft focus: Points turn into soft, round discs. Something else, like an ear or the wall behind, may be sharp instead. You can fix it while the disc is small. Past that, you cannot.

Phone snaps in dim light are mostly camera shake. Wide-aperture portraits usually miss focus. A blurred child in a group shot is subject motion.

The results: one photo, six blurs, five methods

Each row uses the same 1600 × 1067 photo (a Pexels rooftop by Mat Reding, chosen for its brick and window detail). It was blurred by the amount in the first column, then repaired. The score gives PSNR against the sharp original in decibels, with SSIM in brackets. A higher number is closer to the original. Roughly 1 dB is a difference you can see at 100%. The two deconvolvers were given the exact blur they had to undo. They mark an upper bound for the classic approach, not a tool you can download. The AI model and our tool were told nothing.

Three things fall out of that table.

• Camera shake belongs to the AI model. NAFNet did not know the blur, yet it beat both deconvolvers that did on the 16- and 32-pixel shakes. It was also the only tool that helped the 32-pixel test. On soft focus it achieved nothing.
• Deconvolution handles focus, as long as the blur is slight. Set to the correct radius, a Wiener filter fixed most of a 1.5-pixel blur. It also pulled back a good share of a 3-pixel blur. At 6 pixels, no tool gained more than 0.4 dB, and the two deconvolvers made the photo worse.
• "Sharpen" is not "unblur". A sharpness slider applies an unsharp mask. That gave a gain of 1.9 dB on the slightly soft photo, but it lost ground on every shake.

At 100%, side by side

Numbers do not tell the full story. Here are 960-pixel crops of the same window at working size. Tap a version to swap them. Tap the picture to open the full file.

A camera shake of 16 pixels. Look at the window frame and the brick courses.

Sharp original Blurred: 16 px shake Wiener, true kernel NAFNet, blind (AI) AccioPix tool, Shake or motion

Sharp original · 1600 x 1067 · 491 KB

Download this file

The AI result uses the same engine as the tool, so they match down to the pixel. The two "true kernel" deconvolvers recover more brick texture than the AI. They also add more ringing. They had an unfair advantage, but they look worse where it matters, at the edges.

Soft focus, 3 pixels. Same crop.

Sharp original Blurred: 3 px soft focus Wiener, true kernel NAFNet, blind (AI) AccioPix tool, Soft focus Strong

Sharp original · 1600 x 1067 · 491 KB

Download this file

The AI changes very little here. That is fair: it was trained on motion, not focus. At the Strong setting, the deconvolver brings back the mortar lines. If you push it harder, the sky begins to ring. That tells you that you asked for detail that is not there.

Unblurring an image online, free, without uploading it

If you want to unblur an image online without sending it to a server, run the tool in your browser. The AccioPix unblur tool uses the two engines from this test. Your files never leave your device.

1. Open the tool on a computer with an up-to-date Chrome, Edge or Safari. The AI engine needs your GPU. The focus engine works anywhere.
2. Choose your blur type. Use shake or motion for streaks. Choose soft focus for round blur, beginning at Medium.
3. Drop in the photos. You can use JPG, PNG, WebP or camera RAW, one image or a folder. The AI model downloads once, about 71 MB, and stays cached.
4. Hover to compare each image against the original, or click to view it at full size.
5. Download one photo or the full set in a ZIP, at the original resolution.

If Shake or motion helped little, the blur probably comes from focus. Start again in Soft focus mode. If Strong leaves rings, retreat to Medium. If nothing works, turn to the section after next.

Unblurring a photo on iPhone or Android

To unblur a photo on a Pixel, use Google's Photo Unblur: open the photo in Google Photos, tap Edit, then Tools or Actions, then Unblur. Google says it "improves out-of-focus blurry photos, both new and old" and that the feature "is available on Pixel 4a and later" with 4 GB of memory (Google Photos Help). On other Android phones, Google Photos shows the same editor without that button.

On an iPhone, the Photos app's Adjust panel has light and colour sliders (Apple's guide lists "Exposure, Brilliance, or Highlights" as examples) and the Apple Intelligence Clean Up tool removes objects, but there is no deblur (Apple Support). The practical route on either phone is a browser tool: the focus engine above runs on any phone, and the AI engine on phones whose browser has WebGPU. Otherwise send the photo to a laptop and do it there.

Deleting instead

A photographer with 2,000 frames from a reception doesn't unblur; they cull. The free blurry photo remover scores a whole folder for sharpness and lets one slider separate the soft frames, and how to batch remove blurry photos shows what that detector sees at each blur level in this same test. One finding worth knowing: after the AI repaired the 16-pixel shake, the sharpness detector still scored the result at 9% of the original. Unblurred photos look better than they measure, so cull first, then rescue the keepers you can't reshoot.

The test method

The sharp reference is a 6240-pixel Pexels photo resized to 1600 pixels on the long edge, which is also the working size our tool uses. Six blurred inputs were made from it with known kernels: Gaussian blur at sigma 1.5, 3 and 6 pixels for soft focus, and a straight-line motion kernel of 8, 16 and 32 pixels at 15 degrees for camera shake. Each input went through: an unsharp mask (radius 2, 150%); Wiener deconvolution and 30 iterations of Richardson-Lucy, both given the exact kernel; NAFNet, the MIT-licensed motion-deblurring network from Megvii, run blind in 512-pixel tiles; and the AccioPix unblur tool in a headless Chromium with WebGPU, in Shake or motion mode for the shakes and Soft focus mode (Medium or Strong) for the soft cases. PSNR and SSIM were computed against the reference at the same size, and the variance-of-the-Laplacian sharpness score was recorded for every image. The script is scripts/unblur-lab.py in the AccioPix repository; the files, including the reference, every input and every tool output, are in the lab ZIP (6.4 MB). Synthetic blur is cleaner than real blur, so treat these as the best case for each method; real shake has a wobbly path and real focus miss varies across the frame.

Unblurring photos: FAQs

Can you unblur a photo?

It works on camera shake and subject movement. It works in part on mild soft focus. It fails when focus is badly missed. In our test, an AI deblur model repaired most of a 16-pixel shake. A deconvolver restored a worthwhile amount of a 3-pixel soft focus. Nothing fixed a 6-pixel one.

What is the best way to unblur a picture?

Fit the fix to the blur. Use a motion-trained AI model like NAFNet for shake and movement. Run Wiener deconvolution with a sensible radius for mild soft focus. Save a sharpen slider for a photo that is barely soft. The AccioPix unblur tool provides the first two in the browser.

Why do some unblur apps make faces look fake?

They use face-restoration models that draw a believable face instead of recovering the real one. The output is sharp. Often, it looks like a slightly different person. Deblurring and deconvolution work on the pixels that are there. Restoration invents. For a photo of someone you know, that difference matters.

Does unblurring reduce image quality?

The process can add ringing around edges and amplify noise, especially if deconvolution is pushed too hard. Look for a halo along hard edges or a wavy texture across flat areas like sky. Pull the strength back until it clears.

Can I unblur a photo for free without uploading it?

Yes. The AccioPix unblur tool works directly in your browser. The AI model downloads to your device once, and your photos never leave it. You face no watermark, no account, and no limit.

References

• Google Photos Help, Edit a photo or video on your Pixel phone or tablet. Read 1 October 2026.
• Apple Support, Edit photos and videos on iPhone. Read 1 October 2026.
• Chen, L., Chu, X., Zhang, X. and Sun, J., 2022. Simple baselines for image restoration. Computer Vision, ECCV 2022, Lecture Notes in Computer Science, pp. 17-33.
• Richardson, W. H., 1972. Bayesian-based iterative method of image restoration. Journal of the Optical Society of America, 62(1), pp. 55-59.
• Lucy, L. B., 1974. An iterative technique for the rectification of observed distributions. The Astronomical Journal, 79, pp. 745-754.
• Wang, Z., Bovik, A. C., Sheikh, H. R. and Simoncelli, E. P., 2004. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4), pp. 600-612.