MiniPx
Add to ChromePricingBlogAccount
No. 05 — AI Tools · Cloud HDPro · on our GPU

Old photo faces, recovered.

Everything else in an old photograph can be read around — a soft doorway, a grey sky. The face is the part you actually wanted. Our server finds each one, rebuilds it, and proves it still belongs to the person who was standing there.

Pro · Cloud HDDeleted on completionUp to 25 MB · 24 MP
Restore the faces in a photoCloud HD · our GPU

Pro to restore your photo →

How it works

  1. Pick the photograph — JPEG, PNG or WebP, up to 25 MB and 24 megapixels. Crop tightly first if you want one person only.
  2. Our GPU restores every face it finds and checks each one against your original. If a rebuilt face drifts too far from the person in your photo, the run is refused rather than handing back an invented face — and you are not charged.
  3. Download the restored PNG at the same resolution you uploaded, with the rest of the frame unchanged.

This page is the photo restorer, tuned to what the model does with a face and where it stops.

All Image Tools

AI Tools

Pro Tools

Compress

Convert

Resize

Compress to Size

Compress for Platform

Tools & Utilities

By Device

Compare

Solutions by use case

For Developers & Web

For Sellers & Ecommerce

For Job Applicants & Govt Forms

For Creators & Social

For Work & Documents

Restore Faces in Old Photos — Pro, on our server

The run begins with a detection pass, and everything else follows from it. Before any restoration happens, the tool looks through your photograph for faces. Each one it finds is cut out, aligned, rebuilt at high resolution by RestoreFormer++, and blended back into the frame it came from. The rest of the image passes through untouched. Three consequences fall straight out of that design, and it is worth knowing all three before you spend a run: a photograph with no findable face is declined rather than processed into something barely different; a photograph with nine faces gets nine reconstructions; and nothing outside a face is altered, which is why a tear or a stain across the sky comes back exactly as it went in.

What makes a face restorable is its pixel budget, not its age. An 1890s cabinet card where the sitter’s head fills a third of the frame has far more to work with than a 2011 phone snapshot of six people in a garden. So the useful lever is almost always digitisation, not the model: scan the print at a higher resolution, or crop closer to the person with the on-device crop tool before you upload. Pose and occlusion matter next. Frontal and three-quarter faces do best; a strict profile has fewer of the landmarks the reconstruction and the identity check both lean on; hats, spectacles and a hand at the chin usually survive intact, because the model only rebuilds what it can see. When a frame holds a whole family at mixed distances, the family archive page covers how to work through it.

Reconstruction is not recovery, and it is worth being precise about the difference. At the coarse scale — the shape of a jaw, the position of an eye, the fall of hair — the model is working from what your photograph actually contains. At the finest scale, individual eyelashes and skin pores, it is producing detail that is plausible for a human face rather than detail recovered from your print, because that detail was never in the file. This is why every restored face is then compared against the original with a face-recognition model, and why a restoration that has drifted is refused instead of handed back. Treat a restored photograph as an enhancement for looking at and printing, not as evidence. Nothing is charged for a refused run, so the honest answer costs you nothing. The restoration hub sets out the check in full.

Where the photograph goes while this happens. Face restoration is part of Cloud HD, the Pro-only, opt-in feature set that runs on our GPU because the model is far too large for a browser. Your file is encrypted in transit, uploaded to our server for that run and deleted when the run finishes, and in all cases within 24 hours. It is processed in the United States — Cloudflare R2 for storage, a Modal GPU for inference — and it is never used to train anything. If the photograph is black and white and you want colour as well, the colorizer handles that as a separate step, and every other tool on MiniPx runs on your own device.

How it works

  1. Give the face as many pixels as you can: Scan the print at a higher DPI, or crop tightly to the person, before uploading. Face size in pixels is the single strongest predictor of a good result.
  2. Upload the photograph: JPEG, PNG or WebP up to 25 MB and 24 megapixels. The detector locates every face in the frame.
  3. Let each face be rebuilt and checked: RestoreFormer++ reconstructs each face at high resolution, blends it back, and compares it with the same face in your original.
  4. Compare against the original at full size: Open both files side by side and zoom in. Keep the untouched original; the restored PNG is a derivative, not a replacement.

Frequently asked questions

How large does a face have to be before the model can rebuild it?
There is no published threshold, but the pattern is consistent: the more pixels a face occupies, the better the reconstruction and the more likely it is to pass the identity check. A head that fills a good part of the frame restores convincingly. A head a few dozen pixels wide often does not, and that is exactly the case the identity check exists to catch.
Does it work on a face in profile or turned away from the camera?
Frontal and near-frontal faces restore best, because that is the geometry the model saw most during training and the geometry the identity check reads most reliably. A strict side profile carries fewer of the landmarks both stages depend on, so expect a weaker result. A face turned fully away from the lens is not a face the detector will find at all.
What about glasses, a hat brim or a hand across the chin?
Partial occlusion is common in old photographs and usually survives. The model rebuilds what it can see and blends it back into the surrounding image, so a hat brim stays a hat brim. Heavy glare on spectacle lenses is the awkward case: there is no face underneath the reflection to recover, so the reflection comes back as a reflection.
Does it restore a child’s face as well as an adult’s?
Yes, with the same caveat that applies to everyone: it depends far more on how many pixels the face occupies than on whose face it is. Children in old group photographs tend to be smaller in frame and further back, which is the real reason they sometimes restore less well than the adults standing beside them.
Can I restore one face and leave the rest of the photograph alone?
Every face the detector finds is rebuilt on a run. If you want only one, crop to that person first with the on-device crop tool and upload the crop — you get a restored crop back, which you can then place beside the untouched original. Cropping in also makes the face larger in the file, which usually improves the result.
Is a photo of my grandmother stored anywhere after restoration?
No. It is uploaded to our server for that run and deleted when the run finishes, and in all cases within 24 hours, and it is never used to train anything. The download link for the result expires after a few minutes. This is the Cloud HD path; the rest of MiniPx never uploads your image at all.