Face restoration for blurry photos, in your browser
OpenPixels (openpixels.app) is a free, open source photo enhancer that runs on your own device. Its face restoration finds every face in a photo, rebuilds each one with an AI model and pastes it back, so a soft or blurry face comes out clearer. The photo is never uploaded.
No account to try, no watermark. Saving the result takes a free sign-in.
How face restoration works.
Found
A small face detector, YuNet (about 230 KB), finds each face and where its eyes, nose and mouth are.
Rebuilt
Each face is cropped and rebuilt at 512 × 512 by GFPGAN 1.4, running on your own device.
Pasted back
The rebuilt face goes back into the photo through a feathered mask, so the edge does not show. How strongly sets how much of the rebuilt face is used.
Face restoration works on any photo with a face in it: a phone picture, a frame from a video call, a group photo, a scan. To unblur a face, the photo does not need to be old. For a whole old print, with colour and grain as well, see old photo restoration.
Face restoration in five steps.
- Open the app and press Choose a photo, or drop the photo anywhere on the page.
- Turn on Rebuild faces. The first time, the face model downloads once: about 170 MB.
- Leave How strongly where it starts, or lower it to keep more of the original face.
- Under Size, keep the photo’s size with Same size, just cleaner or enlarge it 2× or 4×, then press Enhance.
- Drag the slider to compare Original with Result, then press Save the photo. Saving takes a free sign-in.
While it works, the app says how many faces it is restoring. If it finds none, it says No faces found, and the rest of the photo is still enhanced.
Face restoration runs in the browser.
Face restoration is in the web app at openpixels.app and in the browser extension for Chrome, Edge and Firefox. The Windows, Mac and Android apps enlarge photos and clear noise, but do not rebuild faces yet.
The face model is a large download, and slow without a graphics card. It is kept on your device after the first time, so face restoration then works offline; the Downloads page lists it and lets you delete it.
A rebuilt face, not a recovered one.
Face restoration is a reconstruction. The model draws a likely face from what is left, so it can differ a little from the person, most of all when the face is tiny or badly blurred. Lower How strongly to keep more of the original. Do not use a rebuilt face for an ID or passport photo.
It does not repair scratches, creases or tears, and it does not change anything outside the faces. For the rest of a soft photo, see the image sharpener.
Try face restoration on your photo.
Small as well as blurry? Convert it to HD in the same run.
Open OpenPixels →Before you open it.
Is face restoration free?
Yes. There is no watermark, no credits and no limit on how many photos you restore. Trying it needs no account; saving a result takes a free sign-in. The photo is processed on your own device and never uploaded.
Which model does the face restoration use?
GFPGAN 1.4, whose weights are published under Apache-2.0, run on your own device. Faces are found first with YuNet from OpenCV Zoo. The repository’s MODELS.md lists every model file with its checksum.
Will the restored face look exactly like the person?
Not always. The face is rebuilt by a model, so it can differ a little from the person, most of all when the face in the photo is tiny. Turn How strongly down to keep more of the original face.
Can it unblur every face in a group photo?
Yes. Every face it finds is rebuilt, one after another, and the app says how many it is restoring.
Does face restoration work offline?
Yes, once the face model (about 170 MB) has downloaded. It is kept on your device, and the Downloads page lets you delete it.