Restore blurry faces in old photos

PicReviver finds every face in your picture and rebuilds sharp, natural detail (eyes, skin and features), free, no sign-up or watermark.

One photo at a time here. Need a batch, a different mode, or the before/after slider? Open the full tool.

What this mode does, and does not do

This page runs a dedicated face model on our server; you use it from the browser, with nothing to install, and it works the same whether the photo arrived as a scan, a phone shot or a screenshot. The button starts off, so nothing about a face changes unless you turn it on yourself, and then only the detected faces are rebuilt while the rest of the picture stays as uploaded. For the rest of the frame, the general restore old photos pass and the enhancer run as separate passes. Because the model is trained on faces, it knows what an eyelid, an iris and a lip edge usually look like, which is why it does better on a face than general sharpening. It runs at a deliberately conservative setting so the result stays recognisable rather than merely pretty. Each photo restore takes a few seconds, and the free tier covers five photos per run and ten runs a day.

How automatic face restoration works

Faces are the first thing to go in an ageing photo: a soft scan or a low-resolution snapshot loses exactly the fine detail our eyes care about most. PicReviver uses a dedicated face-restoration model that detects each face, aligns it, and rebuilds realistic detail in the eyes, skin and features, then blends it back into the original picture. Because it's trained specifically on faces, it produces a believable, natural look rather than the smeared or plastic result of ordinary sharpening. If you're comparing apps to restore photos before picking one, this runs as a website rather than an app: open it, restore a photo (or just the face), and download, with nothing to install first. To restore the whole image rather than just the face, the general restore pass runs as a separate pass. Face recovery here means rebuilding the face that is there, not swapping in a generic one, but on a small face the rebuilt detail can drift from the real person, so see the identity-check list below before trusting a restored face for something that matters.

How to restore a face

Perfect for

• Blurry or soft faces in old family portraits
• Low-resolution snapshots from early phones and cameras
• Faded scans where features have lost detail
• Group photos: every face is restored individually
• Photos you'd also like colourised or upscaled

What the face pass does, and what "fidelity" means

Face restoration is a separate job from general sharpening, which is why it has its own mode. The pass detects faces in the photo and rebuilds each one using a model trained specifically on faces, so it knows what an eyelid, an iris, a nostril and a lip edge are supposed to look like. That is why a face can come back sharp from a photo where general sharpening only made it look gritty.

Face models have a dial between two goals: stay close to the input pixels, or produce the most convincing-looking face. Push it towards "convincing" and you get a beautiful, confident face that may not be the person's. This tool runs at a deliberately conservative setting (fidelity 0.7, leaning towards the input) precisely to keep the result recognisable rather than merely pretty.

It still has a floor. If a face occupies only a few dozen pixels (someone in the third row of a group shot, a figure across a square) there simply isn't enough there to say who it is, and the model will fill in a plausible person. The smaller the face, the more of what you see is invention. Restoring photos where only the face has gone soft is what this mode is for; for the rest of the frame, use the general pass.

Check the identity, not the sharpness

The instinct is to judge the result by how crisp it looks. Judge it by whether it is still the same person. Open the before and after at 100% side by side and compare, in this order:

• Eye shape and spacing, including the corners. These carry most of the sense of "that's them".
• Distinctive marks: moles, scars, a crooked tooth, a chipped front tooth, freckles. A model that has smoothed these away has changed the face.
• Hairline and eyebrow shape. Easy to spot, easy for a model to normalise.
• Glasses. Frames get reconstructed as symmetrical, generic frames surprisingly often, and lens distortion of the eye behind them can be silently corrected.
• Teeth and mouth line in a smile. This is where a rebuilt face most often gains a set of even teeth nobody in the family had.

If any of those changed, the sharp version is a different face and the original is the true one. Keep both, label which is which, and use the restored version for showing rather than for record.

Two hard nos, while we're being honest. Don't use a restored face for identification — a passport or ID photo, a missing-person record, an insurance or legal claim — because it is reconstruction, and those uses need an untouched photo taken for the purpose; if you need a compliant ID photo, take a fresh one and use PassportNinja. And don't run the pass twice on its own output: the second run rebuilds the first run's invention, and identity drifts further with every pass.

What people actually upload

Between 3 August and 2 October 2026 the face tool on this page ran 82 times. Across the 5,436 photos uploaded to the site in that period, the most common sizes were small: 1,002 were between 128 and 256 KB and 694 between 64 and 128 KB, typical of pictures saved from chat apps and social media rather than files from a flatbed scanner. That matters for faces specifically, because file size and face detail are not the same thing: a small file of a group at a wedding can hold faces only a few dozen pixels across, which is exactly the case the fidelity guidance above warns about. When an upload is refused, it is more often an unreadable format or a scan with too many pixels than anything wrong with the photo itself.

Before you ask

By Sam Ridder — I build and run PicReviver on my own hardware, on my own. Who I am.