Restore an old photo, sharpen the faces

What it actually does

Three steps, and all three run in your own browser. First a tiny face-detection network is run over the picture: it says where the faces are and where the two eyes are — which is how we know how big the head is and how far it is tilted. Then a square is cut around each face and put through a super-resolution neural network on its own. Finally the finished face is composited back through a soft oval mask, so the edge of the cut is never visible.

The face is deliberately NOT given to the model at its full resolution, but smaller. That sounds backwards, and we measured it: the model was trained to undo a fourfold reduction, so it expects an input whose real information content is a quarter of the result. A blurry 300-pixel face is not 300 pixels of information — it is about 100 that have already been interpolated once. The quartered version produces a better picture, and thirty times faster.

What it does not do — and that is the point

It does not invent facial features. That is the most important sentence on this page. There are tools that “conjure” a complete, sharp portrait out of a twenty-pixel smudge — those do not recover the face, they draw a plausible one in its place, and the result regularly does not even resemble the person who was photographed. That cannot happen here: the model running on this page is a general-purpose upscaler, it has no notion that it is looking at a face, and it cannot add detail that leaves no trace in the pixels.

It does not retouch. Scratches, tears, creases, mould stains and faded colour all stay exactly where they are. This tool removes softness and compression noise, not physical damage. It does not colourise a black-and-white picture either.

Where face detection fails. On heads seen in profile, on very small faces (the back row of a group photo), in strong backlight, on the glare of a framed photo's glass, behind the reflection in a pair of glasses, and on motion-blurred shots. If we find no face we say so, and we do not hand back a “processed” picture that is in fact the same one.

Why there is no “AI face reconstruction” here

Because none of it can be used cleanly in a commercial product. We went through every well-known face-restoration model in the field, and the licence on the weights failed in every case: one is explicitly non-commercial (CodeFormer), one had its better version withdrawn by its own author citing “commercial issues” (GPEN), and one advertises an Apache licence while its own licence file lists the parts of it that carry “research or evaluation purposes only” terms (GFPGAN). A re-publisher cannot grant a right the copyright holder did not give.

The face detector we use is MIT-licensed and the upscaler is BSD-licensed — in both cases the grant covers the weights, not only the source code. This does less than what the competition promises, but it does exactly what we write about it. The full licence chain is traceable in our developer documentation.

Frequently asked questions