Restore Old Photos With AI: From Shoebox to Screen

Every family has the box: a wedding photo from the seventies with a crease through the groom's face, a grandparent's portrait gone sepia and soft, a group shot that's the only picture of someone who's no longer here. AI can repair all three, often in minutes.

The part people get wrong is the order. Most of the quality in a restored photo is decided before any AI touches it, in how you got the print into your phone or computer. Here's the process end to end.

Step 1: Capture the print properly

This is the step that limits everything after it. A model can only repair detail it can see.

• Use a flatbed scanner if you can borrow one. Scan at 600 dpi for a normal print, higher for something small you'll enlarge. Save as PNG or TIFF, not JPG.
• No scanner? Use your phone in daylight. Place the print flat near a window, not in direct sun. Turn the flash off, because it produces a hotspot and kills detail.
• Shoot straight down. Hold the phone parallel to the print so the edges stay rectangular, and fill the frame with the photo.
• Kill the glare. Glossy prints reflect everything, including you. Angle the print slightly or shade it with a sheet of paper.
• Capture the back too if there's writing on it. Names and dates are worth more than the restoration.

Step 2: Repair in the right order

Work in this sequence, because each step feeds the next:

1. Crop and straighten to the photo's edges.
2. Remove damage: creases, tears, dust and stains.
3. Sharpen and denoise to recover detail.
4. Fix the colour: correct the yellow or blue cast, or colourise if you want to.
5. Upscale last, once the image is clean, if you plan to print it large.

Doing it the other way round, upscaling first, enlarges the damage along with everything else.

Step 3: Pick the tools to restore old photos

Google Photos for the quick fix

If the photo is just soft or slightly damaged, start with what's already on your phone. Google's photo editing page describes Photo Unblur as improving "out-of-focus blurry photos, both new and old," and Magic Eraser removes unwanted objects. What your account includes depends on your device and plan, so check before assuming. For many family photos, this is enough.

Topaz for serious restoration

Topaz Labs builds the tools professional restorers lean on, with dedicated photo restoration and colourisation workflows alongside upscaling and denoising. Its pricing page currently advertises a subscription covering all its apps at $399 a year, with individual apps such as Topaz Photo and Topaz Gigapixel sold separately. Prices and bundles change, so read the page before buying. This is the choice when you're restoring a whole album, not one picture.

Free web tools for one-offs

Cleanup.pictures erases a stain, a crease or a distracting object in the browser for free. Canva handles cropping, basic enhancement and export if you're assembling a photo book. Krea and other image tools can upscale, though results vary with how damaged the source is.

For more options, see our best AI image tools and the free AI image tools list.

About colourising black-and-white photos

Colourisation is the feature families love most and the one to be most careful with. The AI doesn't know what colour your grandmother's sari was; it guesses from what usually appears in similar photos. The result is plausible, not true.

Two rules keep it honest. Keep the black-and-white original as the canonical version, and tell people the colour is reconstructed when you share it. If someone in the family remembers the real colours, a few manual corrections afterwards beat any model.

What AI still can't do

• Invent a face that isn't there. If an eye is torn away, the AI will produce an eye, not their eye. For faces, repair conservatively.
• Recover detail from a tiny, blurry scan. Upscaling guesses. Re-scan at higher resolution instead.
• Know what's real. Generated detail can be convincing and wrong, which matters if the photo is evidence of something.

That's why the one unbreakable rule is: never overwrite the original. Keep the untouched scan, work on a copy, and save the restored version as a new file. Future tools will be better, and they'll need the original to work from.

Sharing the results

Once a few photos are restored, put them somewhere the family can actually see them: a shared album, a printed set for an older relative, or a short video with the names and dates read out. The scan is the preservation; the sharing is the point.

FAQ

What is the best way to restore old photos with AI?

Scan or photograph the print at high resolution first, then repair damage, sharpen, correct colour and upscale last. Google Photos handles light fixes free; Topaz is the tool for serious restoration work.

Can AI remove creases and tears from a photo?

Yes, object-removal tools such as Cleanup.pictures handle creases, stains and small tears well. Large damage across a face is where results get unreliable, because the AI invents the missing detail.

Should I colourise old black-and-white photos?

If you like, but treat the colour as a reconstruction rather than a record. The AI guesses colours from similar images. Keep the black-and-white original and say the colour was added when you share it.

How do I scan old photos without a scanner?

Lay the print flat in indirect daylight, turn the flash off, hold your phone parallel to it and fill the frame. Photograph the back too if it has names or dates written on it.

Does upscaling actually recover detail?

No. Upscaling generates plausible detail that wasn't in the file. For an old print, a higher-resolution scan beats any amount of upscaling.