Bring faces in old, blurry or damaged photos back into focus. Mark each face, and a face restoration network (GFPGAN or GPEN) rebuilds eyes, skin and hair detail, while a classic clean-up removes dust, scratches, grain and yellowing from the rest of the picture. Everything runs on your own device, with a before/after slider to compare.
Last updated:
BeforeAfter
Working…
Faces to restore
Drag a box over each face, from forehead to chin, and resize it with its corner handle.
originalfull
Use your own ONNX model
Load a face restoration model (.onnx) that takes a 1×3×N×N RGB face in −1…1 and returns the same, such as GFPGAN, GPEN or RestoreFormer exports.
Advertisement
Advertisement
How to use the Old Photo Face Restorer
Add a photo by dropping it on the box, choosing it, or pasting it. Scan prints at 600 DPI or more for the best result. A sample old portrait is shown first.
Mark the faces: drag the box over a face so it covers forehead to chin with a little margin, and resize it with the corner handle. Click Add face box for more people.
Click Restore faces. The first time, the AI model downloads and is cached; each face then takes a few seconds.
Adjust Restoration strength until the face looks natural, and use the Before / after slider to compare.
Download the restored photo as PNG or JPG.
What happens to each face
Crop. The square inside each face box is cut out and resized to the model's input size: 512 × 512 pixels for GFPGAN and GPEN-BFR 512, 256 × 256 for GPEN-BFR 256.
Normalize. Pixel values are scaled from 0-255 to −1…1 and arranged as a 1 × 3 × N × N tensor (red, green and blue planes).
Restore. The network encodes the degraded face, looks up matching facial structure in its generative prior and decodes a sharp face of the same size.
Blend. The result is resized back to the box and mixed into the photo with a soft edge, weighted by the Restoration strength, so there is no visible seam.
Tips for restoring old photos
Problem
What to do
Small faces in a group photo
Scan at 600-1200 DPI, then box each face separately. The tool works on up to 1600 px of the photo, so crop group shots into parts if faces are tiny.
Strong yellow or brown tint
Keep Remove color cast on. For real color, run the result through the Colorize Black and White Photo tool afterwards.
Face looks too smooth or "plastic"
Lower Restoration strength to 50-70% or try GPEN-BFR 512.
Dust and scratches everywhere
Keep the whole-photo clean-up on; it removes thin, light or dark marks while keeping edges.
Phone photo of a print
Shoot straight on in soft daylight without flash to avoid glare, and fill the frame with the print.
Always keep your original scan. Restoration is an interpretation, so save the result as a new file and label it as restored when sharing family history.
Frequently asked questions
How does AI face restoration work?
Face restoration networks such as GFPGAN and GPEN were trained on millions of pairs of sharp faces and artificially degraded copies (blurred, noisy, compressed). They learned what real eyes, skin, teeth and hair look like, so given a blurry face they generate a sharp, plausible version with the same pose and features. Both use a generative face model (StyleGAN-style) as a prior, which is why they produce realistic detail instead of just sharpening.
Will the restored face look exactly like the person?
Not always. The network invents detail that is consistent with the blurry original, so on very small or very damaged faces the result can look slightly different from the real person, for example in eye shape or teeth. Lower the Restoration strength to keep more of the original, and compare carefully with the before view, especially for family history work.
Which model should I choose?
GFPGAN 1.4 usually gives the most detail and is a good default. GPEN-BFR 512 often looks a bit more natural on skin. GPEN-BFR 256 works at a lower resolution but downloads and runs faster, which helps on phones or slow connections. Model files are large, especially GFPGAN, but they download only once and are then cached.
How big should the face box be?
Cover the face from forehead to chin and from ear to ear with a little margin, roughly like a passport photo crop. The network expects a face that is centered and upright and fills most of the box. Tilted or side-on faces work less well.
What does the quick clean-up do without AI?
It finds dust spots and thin scratches with a median filter and replaces them, reduces film grain with an edge-preserving bilateral filter, stretches faded contrast, removes yellow or sepia casts with a gray-world white balance and adds gentle sharpening. It works instantly and offline but cannot recreate missing facial detail.
Can it fix torn or missing parts of a photo?
Not large ones. Face restoration rebuilds detail inside faces, and the clean-up handles small dust and scratches. Big tears, missing corners or water stains need inpainting; try the Remove Object from Photo tool for those areas first.
Are my photos uploaded?
No. The model files are downloaded to your browser, and your photo is processed on your own device. Nothing is sent to a server, which keeps family photos private.
Are all the tools on FreeToolCabin tested for quality and accuracy?
Yes! Every tool on our site undergoes automated Q/A testing in real browser environments to ensure high precision, correct output, and zero errors. If you happen to spot a bug or have a suggestion, please contact us at [email protected] and our team will review it right away.