A family submitted a 1943 photograph of their grandfather — a 480×360 pixel scan with a diagonal fold crease, silver mirroring on the lower-left corner, and 40% of the face obscured by a water stain. After AI restoration: the crease was gone, the silver mirroring suppressed, and the face — including the eye and cheek hidden under the stain — was reconstructed. The family was moved. They were also warned: the reconstructed face features were statistically plausible given what was visible, not photographically accurate. The grandfather might have looked like that. He might not have.
This distinction — repair vs. invention — is the most important thing to understand about AI photo restoration.
What the Model Repairs vs. Invents
| Damage type | Operation | Accuracy |
|---|---|---|
| Dust and scratches | Noise removal | High — no content invented |
| Silver mirroring / foxing | Tone correction | High — reverses chemical shift |
| Fading / yellowing | Color normalization | High — predictable degradation pattern |
| Fold creases | Inpainting from adjacent pixels | Medium — blends seamlessly on flat areas, less so on faces |
| Torn edges | Outpainting / edge fill | Medium — invents content outside original frame |
| Obscured faces (>30%) | Face hallucination from model priors | Low — plausible but not accurate |
| Complete areas destroyed | Generative inpainting | Low — entirely invented based on context |
Colorization: What AI Knows and Doesn't
AI colorization has seen remarkable numbers: a DeOldify benchmark showed 87% of colorized images rated "natural" by human judges who could not see the original reference. But color is fundamentally ambiguous in a grayscale image. A blue dress and a red dress produce the same gray value. The model chooses based on statistical priors — what color is most common for that type of object. A sky is almost always blue. A 1940s car interior is probably brown. A woman's blouse in 1920 was probably white, cream, or gray — but it could have been red. The model will not know, and it will not say.
Best Practices for Archival Use
- Always keep the original unmodified scan alongside the restored version.
- Label restored images as "AI-restored" when sharing digitally — this is an emerging best practice in digital archiving.
- For faces with more than 50% damage, treat the reconstruction as an illustration, not a photograph.
- For professional archival projects, pair AI restoration with manual review by a photo conservator.