27 September 2026

 

Eight women stand in a line, each holding a parasol. It's probably a scene from the late 1920s. They wear pale dresses and hats, with one hand on a hip. Behind them are two columns topped with enormous decorative plants. It looks very much like a performance, although I do not know what the occasion was. My grandmother is fourth from the left.

Her face is small and vague, but I recognise it. I recognise her expression, too. I have seen other photographs of her with my mother, and I know how slender she was when she was young. She was outgoing and loved a party. It is easy to imagine her enjoying whatever brought these women together.

What I wanted was a better look at her face. I tried improving the photograph myself, asked AI to help and sent it to professional photo services. None gave me a faithful, clearer portrait. The AI result was the strangest: it appeared to show more, but the face it produced did not look like my grandmother.

How can a photograph become clearer and less accurate at the same time?

 

Start with the best scan of the old photo

 

Sometimes the detail you want is present in an old photograph but hidden by a poor copy. A print may be faded or scratched. A picture taken of it with a phone may have glare or uneven lighting. A file passed around the family may have been compressed repeatedly.

In those cases, going back to the source can help more than running another enhancement tool. If the negative survives, that is worth investigating. Otherwise, use the original print if you can. Scan it or photograph it square on, with even light, and keep an untouched digital copy. Google’s PhotoScan app offers a way to capture prints while removing glare. For digital preservation, the Library of Congress favours the highest resolution available from the source, without artificially enlarging the master file. 

Then try the simple edits first. Brightness and contrast adjustments may bring out tones that were difficult to see. Dust or a scratch can sometimes be removed. You can compare each attempt with the untouched scan and decide whether it has revealed something or merely changed the picture.

My photograph has another problem. It shows eight women from head to foot. Their faces take up only a tiny part of the frame. Even with an excellent scan of the print, there may be little facial detail to work with.

 

What AI photo restoration can and cannot do

 

AI upscaling can make a small image much larger and more pleasing to look at. Different tools use different methods: Adobe’s current Photoshop options, for instance, distinguish between preserving existing detail and adding new, creative detail.

That difference becomes crucial with a face. If the image does not clearly show the shape of an eye or mouth, a generative tool can supply a plausible one. It has learnt what human faces tend to look like. It has no personal memory of the woman standing second from the left.

Researchers are still working on how to restore very poor facial images while preserving the subject’s identity. Reference photographs of the same person can help reduce uncertainty, but even specialist methods cannot guarantee a faithful likeness.

I have other photographs of my grandmother, which give me a way to judge the result. The clearer AI face fails that test. Someone who had never seen her might accept it without a second thought. To me, it is a stranger in her place.

 

How to improve old family photos with AI

 

AI tools can still be useful. They may make a damaged picture easier to enjoy, particularly when repairing a background, removing a blemish or preparing an image for display. The question is how much accuracy you need from the result.

Start with the best scan you can make and keep the original file. Try gentle corrections before generating new detail. If you use AI facial restoration, compare the eyes, mouth, hairline and expression with the source and with other photographs of that person. Be especially wary of a dramatic transformation from a face that was barely visible to begin with.

Save and label the versions separately: original scan, edited copy, AI reconstruction. That way, someone opening the family folder years from now will know which image records what the camera captured and which shows a modern attempt to interpret it.

I may learn more about my grandmother’s photograph by looking beyond her face. Perhaps a relative will recognise the venue or costumes. Perhaps a programme or newspaper clipping will tell me what the women were performing. For now, I can see the parasols, the matching poses and my grandmother’s familiar expression. The picture tells me quite a lot, even in its imperfect state.

I have two versions of it. In the one at the top of this page, my grandmother’s face is vague, but I recognise her expression. In the other, below, her face is clearer but I no longer recognise her. I’m going to bin it.