What the unblur image tool fixes
It handles the everyday kinds of blur: a phone that moved while the shutter was open, a hand that shook in low light, a pet or a child that turned at the wrong moment, and a lens that focused slightly in front of or behind the subject. Edges pull back together, hair and eyelashes separate again, and surfaces regain their texture.
It also cleans up the problems that often come with blur, such as noise from a dark scene and blocky compression from a messaging app. One run usually takes under ten seconds, and the result has the same width and height as the photo you uploaded.
How AI unblurring works
The tool uses SeedVR2, a restoration model from ByteDance Seed and the S-Lab at Nanyang Technological University, presented at the ICLR 2026 conference and released under the Apache 2.0 license. It runs on fal.ai GPUs at a scale of 1x, so it removes blur without enlarging the picture.
SeedVR2 is a diffusion model trained to turn degraded footage back into clean frames. It predicts what a sharp version of your photo looks like, using what remains in the blurred pixels as a guide. Mild blur therefore comes back faithfully, while detail that the blur wiped out is filled in with something plausible rather than recovered.
Our test: camera shake and missed focus
We blurred two sharp photos and ran them through the tool. On a portrait smeared by horizontal camera shake, strands of hair, eyebrows and eyelashes came back crisp, the skin regained its texture, and the face stayed recognizably the same person, very close to the original photo.
On a landscape blurred evenly, like an out-of-focus lens, the result looked sharp and convincing, but the trees along the ridge were not the original trees: the model drew its own branches where the blur had erased them. For scenery that is usually fine; for anything where exact detail matters, compare the result with the original.
Faces, text and documents
Faces keep their overall features, expression and identity, because those survive moderate blur, but the finest details, such as individual lashes or skin pores, are reconstructed. The tool is not designed to change age, expression or appearance, and it is not meant for identity documents or photos that must stay unaltered.
Text needs the most care. A slightly shaky photo of a receipt or a page usually becomes readable, but where letters were smeared together the model can produce clean characters that are not the right ones. Check every number and name against the original before relying on it, and use the image to text tool only on results you have verified.
What it cannot unblur
Blur or pixelation that someone added on purpose to hide a face, a license plate or a document cannot be undone. The original pixels were replaced, so any sharp result would be invented, not revealed; please do not use the tool to try to identify people or read hidden information.
Very heavy blur, where a subject is smeared across dozens of pixels, produces a sharp but imaginary image. Backgrounds blurred deliberately by a portrait mode or a fast lens are part of the photo too, and sharpening them rarely looks natural.
Getting the best result
Upload the original photo from your camera roll, not a screenshot or a copy that a messaging app already compressed. Crop to the subject if it is small in the frame, so its details get more of the model’s attention; photos larger than about 4 megapixels are reduced to that size before processing.
Skip photos that were already sharpened in an editing app, since the halos around their edges confuse the model. If the blur is so strong that the result looks invented, retaking the photo with a faster shutter or a steadier hand is the better fix whenever that is still possible.
Unblur first, then upscale
Blurry photos are often small too, especially pictures saved from chats or old phones. Fix the blur first at the original size, then enlarge the sharpened result 2x or 4x with the AI image upscaler, so the enlargement starts from crisp edges instead of magnifying the blur.
Together the two tools turn a small, shaky picture into one you can post or print. Keep the untouched original as well, so you can always return to it.
Credits, privacy and file formats
Each run costs 1 credit, taken from the 2 free credits everyone gets each day or from a credit pack, and a run that fails does not use one. The tool accepts JPG, PNG and WebP files up to 10 MB and returns a high-quality JPG.
Your photo is sent to fal.ai, which runs the model, and the result is stored so you can download it. Every request passes an automated content check before processing. Avoid uploading documents you would not share with a cloud service.