Restoring a family photo with AI feels like magic until the tool "fixes" your grandmother's eyes into someone else's. The model is not evil. It is filling missing pixels with statistically likely faces. For genealogy and memory, statistically likely is a failure mode.
The goal is not the sharpest possible image. The goal is a readable, respectful restoration that still looks like the people you know.
What AI restore is good at vs. bad at
| Job | AI usually helps | Risk if unsupervised |
|---|---|---|
| Scratch and dust removal | High | Over-smoothing skin texture |
| Color fade / exposure | High | Wrong era colors, plastic skin |
| Mild blur / focus | Medium | Fake detail, "waxy" faces |
| Missing corner / tear | Medium | Invented background objects |
| Heavily damaged faces | Low–medium | Identity drift (biggest risk) |
| Upscaling tiny scans | Medium | Hallucinated freckles/teeth |
If the face is half gone, treat AI output as a hypothesis, not a record. Compare to other photos of the same person before you share the "restored" version as truth.
A staged workflow that prevents most disasters
1) Digitize carefully before you "AI" anything
- Scan at 600 DPI when possible (phone photos of photos work in a pinch; use even light, no flash glare).
- Capture the whole print including borders: damage context matters.
- Keep the original file named
family-YYYY-source-original. Never overwrite it.
2) Run a conservative restore first
Use settings closest to "repair damage" not "enhance beauty." Turn off automatic face "beautify" if the tool has it. You want scratches gone, not a yearbook retouch.
3) Identity check (non-negotiable)
Open original and restore side by side at 100% zoom on faces:
- Eye shape and spacing
- Mouth line and teeth (if visible)
- Distinctive marks (moles, scars)
- Hairline
If two of those changed, reject the restore or mask the face and retry with lower strength.
4) Color decisions are historical, not aesthetic
Unless you have evidence (another color photo, a relative's memory of a blue dress), prefer gentle fade correction over vivid reinterpretation. Label colorized images as colorized when you share them. Future relatives will thank you.
5) Upscale last, not first
Upscaling early multiplies artifacts. Restore → verify → then upscale for print if needed.
Prompt / settings language that keeps tools in line
If the tool accepts a text prompt (many modern restore/edit models do):
"Restore scratches, dust, and small tears. Preserve exact facial identity and expression. Do not beautify skin. Do not change age. Do not add jewelry, teeth detail, or background objects that are not partially visible. Keep film grain moderate. No cosmetic retouching."
Negative instructions matter. Models default to "pretty" if you only say "restore."
Worked example: damaged 1970s print
Problem: Horizontal scratches across a group photo; faded color; one face slightly soft.
Good process:
- Scan original at high res.
- Scratch removal pass only.
- Compare faces to a second photo from the same year.
- Mild color recovery.
- Export web size + archive TIFF/PNG of the final.
Bad process:
- Phone snap in yellow indoor light.
- One-click "HDR restore + beautify + 4x upscale."
- Post to social as "the original restored!"
The bad process often looks more impressive in a feed and less trustworthy in a family archive.
Tool path on aihowto.pro
Start with the Restore Old Photo Guide to get a step-by-step plan for your specific damage type (scratches, fade, tears, water marks) before you touch pixels. When the restore is solid and you need a larger print, use a careful Image Upscale pass: still with the original nearby for comparison.
Ethics and family politics (yes, really)
- Consent: Living people in the photo may not want a public restore on Facebook. Ask.
- Truth labeling: Colorized or heavily reconstructed images should be labeled. Archives care; so do relatives who use photos for identity documents or obituaries.
- Do not "improve" bodies or faces to modern beauty standards. That is not restoration.
- Backups: Keep original + restored. Hard drives die; cloud accounts get locked.
Print and share checklist
| Step | Check |
|---|---|
| Side-by-side face review | Pass / fail |
| Filename includes "restored" or "colorized" | Yes |
| Original archived separately | Yes |
| Shared version resolution appropriate (not a 80MB email) | Yes |
| Caption notes year/place if known | Yes |
When to stop and call a professional
- The photo is one-of-a-kind and the physical print is still deteriorating (stabilize storage first)
- Legal or passport use (AI restores are the wrong tool)
- Severe mold/water damage on the physical media
- You need museum-grade documentation
AI is excellent for family-readable copies. It is not a substitute for conservation of a unique physical artifact.
Bottom line
Restore in stages. Verify identity like a skeptic. Upscale last. Label what you changed. The best AI photo restore does not look "stunning": it looks like your people, only clearer, with the damage gone and the truth left alone.