Bringing Old Family Photos Back to Life with AI: A Careful, Step-by-Step Guide

I spend most of my working week fixing photographs, and the single most emotional job is always the same: someone hands me a cracked, yellowed print of a grandparent and asks if anything can be done. In 2026 the answer is almost always yes. AI has made photo restoration something you can do on your phone in an afternoon. But the tools are also very good at quietly inventing things that were never in the original, and that is the part nobody warns you about. This guide walks through a process that gets real results without turning your great-grandmother into a stranger.

Start by scanning, not photographing

Every restoration lives or dies on the quality of what you feed the software. A phone snapshot of a print, taken at an angle under a lamp, gives the AI glare, shadows, and warped edges to fight through. Do it properly once and you save yourself hours.

• Use a real scanner if you can. Flatbed scanners at libraries and copy shops are fine. Scan at 600 DPI for a standard print, higher if it is small (a wallet-sized photo benefits from 1200 DPI).
• No scanner? Use Google's PhotoScan app (still free in 2026). It takes several shots and stitches out the glare automatically, which a single photo cannot do.
• Lay the print flat, fill the frame, and shoot in soft daylight near a window rather than under a ceiling light.

Save the raw scan somewhere safe before you touch it. You will want the untouched original later, and you never want to overwrite it.

Do the repairs in a sensible order

Restoration is really four separate jobs, and doing them in the wrong order makes each harder. Work through them like this.

1. Fix the physical damage first

Scratches, tears, missing corners, tape marks, and water stains should go before you do anything else, because sharpening and colorizing will only amplify them. This is where AI is genuinely excellent now. In Photoshop, select a damaged area with the lasso, open Generative Fill, and leave the text box empty or type something plain like "clean photo background, no damage." It rebuilds the missing texture convincingly. If you do not have Photoshop, Google Photos' Magic Editor can erase creases and stains on a phone, and the free web tool Photopea has a similar generative fill.

One rule I never break: on faces, work in small pieces. Repair a torn cheek by selecting just the cheek, not the whole head. The moment you ask AI to regenerate an entire face, it stops restoring and starts imagining.

2. Denoise and sharpen

Old film grain, scanner noise, and soft focus are a different problem from physical damage. The specialist tool here is Topaz Photo AI, which is what a lot of professionals actually use. Its "Recover Faces" feature is built for exactly this and lets you dial the strength down with a slider. Free alternatives include Google Photos' Photo Unblur and the enhance feature in the Remini app.

Remini deserves a specific warning. It produces dramatic before-and-afters that look amazing in a thumbnail, but it tends to over-smooth skin into a plastic, airbrushed finish and give everyone the same slightly generic modern face. For a distant background figure that is fine. For the main subject, use it at low strength or skip it.

3. Upscale if the photo is tiny

If your only copy is a small, low-resolution image (a crop from a group shot, say), an upscaler can multiply the pixels intelligently. Topaz does this, as does Magnific and the upscale option built into many editors now. Upscale after denoising, not before, or you just enlarge the noise.

4. Color and tone last

Save color for the end. For a black-and-white photo you want to keep as black-and-white, this just means fixing the yellowing and faded contrast: pull the print back toward true neutral grays. Lightroom, Google Photos, or the free Snapseed app all handle this with an auto-tone button that gets you 80 percent there.

Colorizing: powerful, and where honesty matters

AI colorization has gotten genuinely good. Palette.fm, Photoshop's colorize neural filter, and the colorize options inside Google's tools will turn a 1950s photo into believable color in seconds. It is also, by definition, guessing. The software does not know your grandmother's dress was green. It picks statistically likely colors.

My advice: colorize a copy, keep the black-and-white as the real record, and if any living relative remembers the actual colors, correct them by hand afterward. Nothing lands harder with family than "her coat was actually red, here it is right." And label colorized versions honestly when you share them, so nobody in 2050 mistakes an AI guess for documentary fact.

What AI still gets wrong, and how to catch it

This is the section I wish more tutorials included. AI restoration hallucinates, and the mistakes are subtle enough to slip past a quick glance.

• Invented teeth and eyes. When a face is soft, enhancers frequently add a full set of even teeth to a closed mouth, or open eyes that were shut. Zoom to 100 percent and check.
• Changed identity. Aggressive face recovery narrows noses, smooths jawlines, and de-ages people until they resemble a generic attractive stranger. Compare side by side with the original constantly.
• Wrong hands and jewelry. Rings, watches, and fingers get mangled or added.
• Text and patterns. Any writing, signage, or fabric pattern in the photo tends to turn into plausible-looking gibberish.

The defense is simple: keep the original open beside your edit and flip between them. If a change makes the person less like themselves, undo it, even if it looks technically "better." A slightly soft, honest face beats a crisp, fictional one.

A realistic 20-minute workflow

1. Scan the print at 600 DPI (or use PhotoScan). Save the original.
2. Open it in Photoshop or Photopea. Repair tears and stains with Generative Fill, one small area at a time.
3. Run Topaz Photo AI or Google Photo Unblur to denoise and gently sharpen. Keep face strength low.
4. Upscale only if the image is small.
5. Auto-tone in Lightroom or Snapseed to fix fading.
6. Optionally colorize a copy. Correct known colors by hand.
7. Compare against the original at 100 percent. Undo anything that changed who the person is.
8. Export a high-resolution copy for printing and a smaller one for sharing.

When to hand it to a person

AI handles maybe 90 percent of restorations well. The exceptions worth paying a human for: a face that is more than half missing, a photo where several people overlap in the damaged area, or an heirloom where getting it exactly right matters more than getting it done fast. For everything else, the tools on your phone in 2026 are better than what a professional lab could do a decade ago. Just keep one hand on the undo button and the original always in view.