What file should I upload — a scan or a phone photo?

A flatbed scan at 600 DPI beats a phone photo almost every time, because it captures the print evenly without glare or perspective. If you only have a phone, shoot in soft daylight near a window, hold the camera parallel to the print, and avoid the camera flash.

Can AI fix a photo that's ripped in half?

Mostly, yes. Tape the pieces back together on the back, scan it flat, and the model will reconstruct the seam. The bigger the missing area, the more the AI has to invent, so a clean tear restores better than a chunk that's gone entirely.

Is the colorization historically accurate?

It's an educated guess based on light and context, not a record of fact. Skies, skin, and grass come out believable. A specific dress or uniform color the model has no way of knowing may be wrong. Treat color as interpretation, not documentation.

What happens to my photos after I upload them?

Your images are processed to generate the result and are not used to train models. See our privacy policy for how long files are kept and how to delete them.

Guide

How to restore old photos with AI (and when it actually works)

Most guides on this oversell it. Here's the honest version: what AI restoration handles well, where it still falls short, and how to get a result you'd actually want to print.

A few years ago, fixing a cracked photo of your grandmother meant either learning the healing brush in Photoshop or paying a retoucher $80 and waiting two weeks. AI changed the math. A model trained on millions of damaged-and-repaired pairs can now read a scratch, understand what skin or sky should sit underneath it, and paint it back in seconds. That's the good news. The part nobody tells you is that the result depends almost entirely on what you feed it.

What it handles well

Surface damage is the easy win. Dust, fine scratches, foxing, the white crease from a photo that lived folded in a wallet — the model erases these cleanly because it has seen the same damage on thousands of training images. Faded contrast comes back too. A washed-out print from the 1970s gets its blacks and midtones restored without you touching a single slider.

Faces are where it earns its keep. Old film and slow shutters left a lot of portraits soft or slightly out of focus. Modern restoration models reconstruct eyes, the line of a jaw, the texture of hair — and the better ones keep pores and wrinkles instead of smoothing everyone into wax. If a face turns plastic, that's a sign the tool is doing too much.

Where it still struggles

When a large part of the image is gone — a corner torn off, a face half-missing — the AI isn't repairing anymore, it's inventing. Sometimes the invention is convincing. Sometimes it gives someone the wrong eye shape or smooths away a feature that made them look like themselves. I'd rather a restoration leave a small flaw than confidently rebuild a face into a stranger.

Text is the other weak spot. Handwriting on the back of a print, a sign in the background, a date stamped in the corner — models tend to turn these into gibberish that looks like letters. If the writing matters, restore the photo but keep the original scan for the words.

Try it on one photo first

Pick the photo that matters most and run it through Restory. You'll know in about a minute whether the result is worth doing the rest of the box.

Restore a photo

Restore, colorize, or both?

These are different jobs. Restoration repairs damage and sharpness and leaves a black-and-white photo black and white. Colorization adds color but doesn't fix scratches. For a damaged B&W print, running both gives the most dramatic result — but it's also the most interpretive, because the AI is repairing and guessing colors at once. If a photo is precious and you're unsure, do the restore first, look at it, then decide whether color adds something or just distracts.

Getting the best scan

This is the single biggest lever, and it costs nothing. The model can only work with the detail you give it.

Use a flatbed scanner at 600 DPI if you have one. It lights the print evenly and keeps it flat, which a phone can't do.
No scanner? Shoot near a window in soft daylight, never with the flash. Flash blows out highlights and bounces off the glossy surface.
Hold the camera straight above the photo so the edges stay rectangular. Skewed shots confuse the model about geometry.
Scan the print, not a framed copy behind glass. Glass adds reflections and a second layer of dust the AI will try to "fix."

Common questions

What file should I upload — a scan or a phone photo?: A flatbed scan at 600 DPI beats a phone photo almost every time, because it captures the print evenly without glare or perspective. If you only have a phone, shoot in soft daylight near a window, hold the camera parallel to the print, and avoid the camera flash.
Can AI fix a photo that's ripped in half?: Mostly, yes. Tape the pieces back together on the back, scan it flat, and the model will reconstruct the seam. The bigger the missing area, the more the AI has to invent, so a clean tear restores better than a chunk that's gone entirely.
Is the colorization historically accurate?: It's an educated guess based on light and context, not a record of fact. Skies, skin, and grass come out believable. A specific dress or uniform color the model has no way of knowing may be wrong. Treat color as interpretation, not documentation.
What happens to my photos after I upload them?: Your images are processed to generate the result and are not used to train models. See our privacy policy for how long files are kept and how to delete them.

None of this makes AI a replacement for a skilled retoucher on a museum-grade job. But for the shoebox of family photos most of us have, it gets you most of the way there in a minute, for pennies. Start with the one photo you'd be saddest to lose, and go from there.