How to Restore Old Photos at Home (2026): What Actually Works

Most guides to restoring old photos skip the part that decides the result: what the software is actually able to do, and what it is quietly inventing. This one doesn't. If you have a box of faded family photos and you want them to look good — not weird — this is the honest version.

Step 1: how you capture the photo matters more than the software

This is the step people rush, and it caps everything that follows. No AI can recover detail that was never captured.

• Scan at 600 DPI, not 300. For a small print — and old family photos are often tiny — 300 DPI gives the software almost nothing to work with. 600 DPI is the single biggest free improvement you can make.
• Turn off the scanner's own "restore" and "descreen" options. They smear detail before you've started. You want the flattest, most boring scan possible.
• No scanner? Use your phone properly. Daylight near a window, never direct sun, never flash — flash bounces straight off the print. Hold the phone parallel to the photo, not at an angle. Take the shot square-on so you aren't fighting perspective later.
• Photos in an album or behind glass: take them out if you safely can. Glass adds reflections that look exactly like damage to restoration software.

Step 2: understand what "AI restoration" actually does

Restoration is not one thing. It's three or four different jobs, and they have very different success rates. Knowing which is which is what stops you being disappointed.

Upscaling and sharpening — reliable

Modern upscalers genuinely reconstruct plausible texture rather than just enlarging pixels. Fabric, brickwork, hair, foliage — these come back convincingly. This part works, and it works on nearly every photo.

Faces — the big one, and the one to be careful with

Here's the thing nobody tells you: an upscaler cannot fix a face. It will make a blurred face into a bigger blurred face. Faces need a dedicated face-reconstruction model, which is a different tool doing a different job.

And that model does not "recover" the face. It rebuilds one that is statistically plausible given what's there. On a decent photo that's remarkable. On a very blurred one it can drift into someone who looks like your grandfather rather than being him — the tell-tale signs are skin that's too smooth and eyes that look glassy.

The practical fix: use a strength slider and turn it down. Around 60% usually keeps the real face while still recovering eyes and detail. Any tool that only offers full-strength face rebuilding will eventually give you a stranger.

Colourising black and white — good, but it is a guess

Colourisation is genuinely impressive now, and it is important to be clear about what it is: a good guess, not a memory. The model has learned what the world usually looks like. So it gets skin, sky, grass, wood and skin tones convincingly right — and it has no way of knowing that your grandmother's dress was red.

Expect believable, not accurate. For most family photos believable is exactly what people want.

One technical detail worth knowing, because it separates good tools from bad: a well-built colouriser predicts only the colour and keeps the original brightness detail untouched. If colourising makes your photo look softer, the tool is doing it wrong.

Cracks, folds and stains — still needs a human

This is where automatic tools quietly fail, and the reason is interesting. Damage detection works on shape: a scratch is a thin line, so the software looks for thin lines.

The problem is that a mouth is also a thin line. So is an eyelid, an eyelash, a nostril. Any automatic tool aggressive enough to remove a fold across a face is also capable of erasing the face's actual features — and once you erase an eyelash, the face-rebuilder is guessing from damaged evidence, which is where those glassy eyes come from.

The reliable approach is the one professional restorers use: automatic cleanup for dust and specks anywhere, and a manual brush for damage on faces, where a human decides what is a crease and what is a lip.

Step 3: the order you do things in

This matters more than people expect. The right sequence is:

1. Clean up dust and scratches first, at the original size
2. Colourise, if it's black and white
3. Rebuild faces
4. Upscale last

Why last? Because repairing damage on an already-enlarged photo means filling a hole four times as wide, and every repair tool turns wide holes into smudge. Fix the small original, then let the upscaler rebuild real texture over the repair. Same tools, completely different result.

What about the free online restorers?

They work, with three costs worth weighing: you upload family photographs to someone else's server, you're usually limited to a handful before paying, and the output is often watermarked or capped in resolution. For one photo that's fine. For a whole family album it stops being fine quickly — both on cost and on privacy.

Doing it on your own PC

This is why I built PhotoFix: an old-photo restorer that runs entirely on your own computer. No uploads, no account, no per-photo credits, and no cap on how many you do — which matters when the box under the bed has two hundred photos in it.

It does the four jobs above in the right order: dust and scratch removal, colourisation for black and white, face rebuilding with a strength slider so you can keep the real face, and 2×/4× enlargement. Damage on faces gets a repair brush, because as explained above, that judgement should be yours and not a shape-matching algorithm's. The interface is available in English, Romanian, Italian, Spanish, German and French.

It's a one-time purchase, not a subscription. See the shop — and if you'd rather just do the free part, the guidance above works with any tool you like.

FAQ

Can AI restore a photo that is very blurry? Partly. It will produce a sharp, plausible image — but the further it has to guess, the less it is really the person. Lower the face-rebuild strength on badly damaged photos.

Will restoring change my original file? It should never do so. Any tool worth using writes a new file and leaves the original untouched.

Is colourisation historically accurate? No. It is a plausible guess based on what the world usually looks like. Treat it as an interpretation.

What resolution should I keep? Keep the largest version you produce. You can always make a smaller copy for messaging; you can never go back up.

More from the workshop: how to write AI photo editing prompts that don't ruin the photo, how this one-person studio works, or the rest of the apps.