Photo Restoration

When AI photo restore goes too far

Spot facial drift, honor consent, and keep untouched masters when generative edits invent a stranger.

Monochrome close-up of bundled old letters and a photograph.
Photography · Image credits

At a glance

  • Stop if eyes, teeth, age lines, or clothing weave no longer match the master scan.
  • Consent covers living subjects and family agreements—not only the upload checkbox.
  • AI outputs are derivatives: different filenames, never overwrite masters.

In plain language

Knowing when AI photo restore goes too far means comparing a generative output to the untouched master for identity drift, documenting consent, and refusing edits that invent biographical detail—so a sharper stranger never replaces a real relative.

Assumes: A lossless or carefully scanned master exists; AI tools are optional and consent-based; family sharing norms may restrict public posts.

A restored portrait should still look like the person in the emulsion. When generative tools invent smoother skin, modern teeth, or fashion-magazine eyes, you no longer have a conservation aid—you have a new fiction wearing your relative’s name.

Red flags (stop and compare to the master)

  • Eyes and brows changed shape, spacing, or gaze direction.
  • Age cues vanished (crow’s feet, nasolabial folds, freckles, scars that relatives remember).
  • Teeth or lips appear where the master shows a closed mouth or shadow.
  • Clothing weave or jewelry morphs into generic fabric and chrome.
  • Background props appear that were not in the scan.
  • The file overwrote the only master, or shares the master’s filename.

If any red flag appears, keep the master, discard or quarantine the derivative, and tighten prompts using AI restoration prompts that preserve facial features.

  1. Upload consent on HeirloomMath covers sending a resized copy to the disclosed AI provider after you tick the box—see the privacy policy.
  2. Living subjects (and often close family) deserve a say before public social posts of restored faces.
  3. Minors and sensitive contexts need extra care; when unsure, keep derivatives private.
  4. Label shares clearly: AI-restored derivative · compare to master.

Keep originals; name derivatives

Follow the photo digitization hub two-tier habit:

  • Masters stay untouched (often TIFF / lossless).
  • Restored files use different names (…_ai-restored.png) under /ai-restored/ or similar.
  • Re-run from the master if you dislike an edit—never “fix” the only copy in place.

Optional practice: inspect fade with histograms, plan PPI with the scan checklist, then try a careful restore in the photo & slide digitization planner only after consent.

Common questions

What is facial-feature drift?

When eye shape, brow, teeth, age cues, or skin texture shift toward a modern average face. If you would not swear the person is the same, reject or constrain the edit.

Is ticking the upload box enough consent?

It covers sending a copy to the disclosed provider. Living subjects and family publication still need ordinary interpersonal consent.

Where do I practice safer prompts?

Read AI restoration prompts that preserve facial features, then use the photo planner’s restore tool only after consent—and keep the master.

Put it into practice

Try it with your own collection

Sources & further reading

  1. FADGI – Technical Guidelines for Digitizing Cultural Heritage Materials
  2. Library of Congress – Personal Digital Archiving: Photographs

Information on this page is for educational archival preservation and historical genealogy research. Always test conservation handling on non-unique materials first, verify AI handwriting transcriptions against original county or NARA microfilm, and never use autosomal DNA statistics for clinical or legal parentage determinations. Nothing on this site is legal, probate, medical, or financial advice. Spotted an error? Tell us and we will review it under our corrections policy.

Back to the beginning ↑