← All guides · 2026-09-26 · Updated 2026-09-28 · PhotoUnfade team

How to Restore Old Family Photos with AI (Step-by-Step Guide)

Old family photos do not just fade — they scratch, yellow, crease, spot, blur and tear. For decades the only fix was a professional retoucher charging $25-$75 per photo and taking two to seven days. In 2026, AI restoration does roughly 90% of that work in under a minute, for a few cents. This guide covers what AI can genuinely fix, where it fails, and how to avoid the mistake that ruins more restorations than any technical limitation.

What is AI photo restoration?

AI photo restoration uses a generative image model trained on millions of before-and-after pairs of damaged and clean photographs. Given your damaged scan, the model predicts what the undamaged original looked like: it removes scratches and stains by filling them from surrounding context (a technique called inpainting), rebalances shifted color channels, recovers tonal range in faded highlights and shadows, and sharpens soft detail. It is not a filter — a filter applies the same math to every pixel, while the model understands what a face, a shirt collar or a sky is supposed to look like.

That understanding is both its superpower and its danger, and the danger is what the rest of this guide is about.

What damage does AI restoration actually fix?

Ranked from most reliable to least, based on how the technology works:

The last row is the one that matters most, and almost no marketing page will tell you about it. More on that below.

What is the biggest mistake AI restoration tools make?

Over-restoration. Aggressive models "improve" faces until your grandmother looks like a stranger with plastic skin, an uncanny fixed stare, and teeth that have detail the original blur never contained. This is the single most common complaint in reviews of the large consumer photo apps: "it doesn't look like my mother anymore."

Why does this happen? Generative models are trained to produce pleasing, high-quality images. When detail is genuinely lost — a soft, faded face — the model fills the gap with an average face drawn from its training data. The output is sharper and prettier, but it is no longer the specific person in your photograph. That is not restoration; it is repainting.

The fix is a workflow choice, not a better model: use a tool that offers a subtle, conservative mode and shows you the original beside the result, so you decide how far is too far. That is precisely why PhotoUnfade defaults to Subtle and puts a before/after slider on every result.

How can I tell a faithful restoration from a repaint?

Three checks catch almost every bad result, and you do not need any expertise:

1. Compare at full zoom, not the thumbnail. Look at the faces side by side at 100%. Downscaled previews hide identity drift.
2. Check skin tone and complexion. A faithful repair removes the yellow cast of age but keeps the person's actual skin tone. If a pale relative comes back looking tanned or flushed, or the photo was black-and-white and came back in color you did not ask for, the tool over-processed.
3. Look at the eyes and teeth. Over-sharpening shows up there first. If detail appeared that could not have been recovered from the original, it was invented.

We put our own output through exactly these checks with instruments rather than eyeballs — skin hue deviation, invented-color percentage, sharpening overshoot. The full numbers, including two failures we found and fixed in our own pipeline, are in our pixel-by-pixel face accuracy test.

How do I restore a photo step by step?

1. Digitize at the highest quality you can. Scan at 600 dpi, or photograph the print flat and evenly lit, filling the frame. This is the step that decides everything: AI cannot recover detail that was never captured. A sharp scan gives the model real information to work with; a blurry phone photo of a glare-covered print caps the result no matter how good the algorithm is. Our digitizing guide covers scanner settings, file formats and how to photograph prints without a scanner.
2. Upload and start with Subtle mode. It removes damage — scratches, fading, yellowing, stains — while strictly preserving faces, expression, pose and composition.
3. Compare with the slider. Drag it back and forth. If damage remains on a heavily damaged photo, run it again in Full mode, which reconstructs more aggressively.
4. Download the full-resolution result. It is print-ready; families routinely print restored photos at 8×10 and larger for albums and frames.

When should I use Subtle mode versus Full mode?

• Subtle (the default): fading, yellowing, light scratching, dust, minor softness — and any photo where the faces are the point and the damage is moderate. It changes the least, which is exactly what you want for a portrait of someone you knew.
• Full: heavy staining, water damage with tide lines, significant tearing, severe blur, or a photo so damaged that a conservative pass leaves it looking untouched. It reconstructs missing areas from surrounding context, which means more repair — and more risk of the model guessing. Always review the side-by-side carefully in Full mode.

If you are unsure, run both. Subtle first, then Full, and keep whichever looks more like the person you remember.

What about photos that are truly destroyed?

If roughly 30% or more of a face is missing, no AI can genuinely recover it — it will invent one. For irreplaceable heirlooms in that condition, the best workflow is a hybrid: run the AI pass first, then send both the AI result and the original scan to a professional retoucher. The AI pass cuts their manual work substantially, and many now quote lower for a pre-restored input. A professional can also reconstruct a missing face from other real photos of the same person, which is reconstruction rather than invention — a difference that matters enormously for genealogy and memorial work. We cover the cost and quality trade-offs in detail in our honest AI versus professional comparison.

How much does AI restoration cost?

Professional retouchers typically charge $25-$75 per standard photo and $100-$200+ for severe damage, with a two-to-seven-day turnaround. AI restoration runs in under a minute at a few cents per image, or a flat subscription. For a shoebox of 200 photos, that is the difference between an afternoon project and a $5,000-$15,000 decision — which is exactly why most flooded, faded shoeboxes historically got thrown away.

If you want to judge your own photos before committing anything: PhotoUnfade gives you three free restorations at full resolution, with no account, no credit card and no watermark.

How should I handle the specific damage I have?

• Water damage, tide lines, stuck-together prints or mold — this one is urgent, because wet photos can fuse permanently within days. See our water damage first-aid guide before doing anything else.
• Yellow or orange cast, washed-out contrast, gray blacks — see why photos turn yellow and how to fix them, which explains why the order of corrections matters.
• Stopping future damage — restoration does not stop the original print from continuing to deteriorate. See our preservation guide for archival materials and the environment numbers that actually matter.

How do I preserve the restored result?

Digital files do not fade, but they do get lost. Save the restored copy in at least two places — cloud plus a local backup, following the 3-2-1 rule (three copies, two media types, one offsite). Keep the highest-resolution file you have as your master; do not repeatedly re-save JPEGs, since each save discards a little more data. If you print, use archival pigment inks and acid-free, PVC-free materials. And store the original print flat, cool, dark and stable — a beautiful restoration does not stop the physical photograph from continuing to decay in the shoebox.

Restoring a shoebox is one of the few family projects that gets easier every year and impossible eventually. The photos that are merely faded today may be fused, molded or fully blank in a decade. Start with the three faces you most want to see clearly again — it takes about a minute each.