Restoring old photographs
A cracked print from grandma's album, a faded card from 1945, a crease running right across a face. A neural network really does repair this kind of damage, and the result can be striking. But before you start, it is worth understanding one thing about what exactly you will get. It decides both the order of work and how to treat the finished file.
What restoration actually does
Where the photo has a crack or a crease, there is no image at all. The neural network does not "reveal" what is hidden under the damage; it draws in whatever fits there.
If the crease ran through the background, that is exactly what you want, and nobody will ever notice the difference. If it ran through a face, you will get a plausible face, and it may not quite be that face.
Order matters
Restoration is the one job in this book where the sequence changes the result. Every pass redraws the whole photo, so invention piles up: whatever was done earlier, the next step takes as the truth. Hence the rule: one job at a time, not "fix everything" in a single request.
The order it is done in
Restoration is the one job in this book where order changes the result. Every pass redraws the whole photo, so invention piles up: whatever was done earlier, the next step takes as the truth.
Everything that follows works from this file. A flat scan of the print at 600 DPI gives a noticeably better result than a phone photo of it: the model takes moiré, glare and a tilted surface for part of the picture and starts "restoring" them.
What to do: Scan it, or photograph it straight from above in even, diffused light without flash. Shoot at the largest size and do not send the file through a messenger first, because that recompresses it.
And one rule throughout: keep the original and the result of each step as separate files. When, three passes in, the face turns out to have drifted, you will want something to go back to.
What repairs well and what does not
| What is on the photo | How reliable | What to say in the request |
|---|---|---|
| Fading, yellowing, low contrast | Excellent. It is drawn out of the file; nothing is invented | "Restore natural contrast, remove the yellowing, do not draw anything in" |
| Scratches and cracks across the background | Excellent. The background is rebuilt from neighbouring areas, and nobody can check it | "Remove the scratches and creases, restore the background, keep the film grain" |
| A crease or crack across a face | Risky. Part of the face will be invented, and the result looks convincing either way | "Do not change the facial features, restore only the damaged area" |
| Overall softness and grain | Moderate. Film grain is part of the photo, not a defect; remove it and you easily get a plastic picture | "Keep the film grain, do not make it glossy" |
| A torn-off corner, a missing part of the frame | Invented entirely. Sometimes beautiful, but no longer a record of its time | "Only extend the background, do not add people or objects" |
| Colourising black and white | The colour exists nowhere; it is composed entirely. A muted palette looks more truthful than a bright one | "Natural muted colours, like film stock of the period" |
Scan, do not photograph
The cheapest improvement to the result happens before any processing. A flat scan of the print works noticeably better than a phone photo of it: the model takes moiré from the paper, glare from the lamp and a tilted surface for part of the image and starts "restoring" them.
And do not send the file through a messenger before processing: it recompresses the photo, and some of what you later ask to bring back is lost right there.
On photos of people who are gone
Restoring a family archive is the most common reason to take on this job, and it is worth saying out loud what usually goes unsaid. Restoring a photo means letting a model redraw the face of a person nobody can ask any more.
- Do not throw away or replace the original. The restored file is a copy with guesswork in it; the original is the record. Keep both, and label which is which.
- Show the result to people who remember. A relative who knew the person by sight notices a substitution where you will not, and that is the only check available.
- Do not ask it to "make them younger", "add a smile" or "open their eyes". That is no longer restoration but a different image of a different person, and in a family archive it will sooner or later start being taken as real.
- For an archive, a publication or research, a restored photo is only acceptable with a note that it was processed.
Attach the scan and start with the first step: light and colour.
Further reading
The general mechanism and the line between "reveal" and "invent": improve photo quality with AI. If the photo is also blurred: removing blur from a photo. If you need a large print: enlarging a photo. More on colourising black and white in improve and restore a photo. And if you would like to see the same photo move, that is the next job over: bring a photo to life.
Frequently asked questions
How do I restore an old photograph with a neural network?
Will the face come back as it was?
Is free online photo restoration realistic?
A scan or a phone photo?
Can I remove grain from an old photo?
Should I colourise a black-and-white photo?
Can this recover photos deleted from my phone?
In the Twelver chat you can clean up an old photo step by step (light, damage, colour, size) and keep the intermediate result at each one. Starter tokens arrive on your first visit, and signing up adds more.
Try it yourself
Everything in this guide runs inside Twelver
One chat for text, images, video, music and voice — no separate services or subscriptions.
Open Twelver chat