Why faces change during AI restoration
A restoration model must decide which pixels are damage and which belong to the person. When a face is small, blurred, scratched, or partly missing, the source may not contain enough evidence for an exact answer. The model can produce a plausible eye or mouth that looks convincing without being historically correct.
Strong face enhancement can also replace natural film grain, pores, wrinkles, and soft focus with a modern portrait look. The safest goal is not maximum sharpness. It is the smallest change that makes the original easier to see.
Identity details to protect
- Overall face shape, jaw, chin, and distance between features
- Eye spacing, eyelids, eyebrows, expression, and direction of gaze
- Nose width, mouth shape, teeth, smile lines, and facial asymmetry
- Hairline, ears, glasses, jewellery, moles, scars, and other distinctive marks
- Age cues such as wrinkles and skin texture rather than a beauty-retouched face
A face-first restoration workflow
- Start from the best source. Scan the print at 600 DPI in colour when possible. A compressed message attachment gives the model less real information to preserve.
- Keep an untouched master. Duplicate the scan before any crop, contrast change, restoration, or colourization.
- Run one moderate repair. Fix scratches, fading, and noise before adding stronger upscaling. Repeated enhancement compounds invented texture.
- Compare at matching zoom. Place the original and result side by side at the same size. A sharper image can feel more accurate simply because it is easier to see.
- Choose the conservative version. If two outputs both clean the damage, keep the one that changes fewer identity details.
What natural restoration should look like
A faithful result can still contain film grain, soft edges, wrinkles, uneven lighting, and the visual character of the original process. Those are not automatically defects. Remove the scratch across an eye; do not turn the entire face into a studio headshot.
RestoreBunny is designed for natural-looking repair, but every generative tool has limits. Use the before-and-after preview, and read our editorial policy on inferred detail before using a restoration as part of a family archive.
Warning signs that a result went too far
- The person looks younger or has a different expression
- Eyes, teeth, glasses, or jewellery appear where the source is unreadable
- Skin has no texture while clothing and the background remain grainy
- Hair, ears, fingers, text, or uniforms have changed shape
- The output looks like a modern camera portrait rather than a repaired print
FAQs
Can AI restore an old photo without changing the person?
It can preserve identity well when enough facial information remains, but no generative restoration can guarantee exact detail in a blurred or missing area. Compare the result with the source and use a lighter pass if features drift.
Why do restored faces sometimes look waxy?
Strong denoising and face enhancement can remove pores, wrinkles, and film grain. Start with moderate restoration and avoid repeating enhancement on an already enhanced output.
Which facial details should I compare?
Check face shape, eye spacing, eyelids, mouth, teeth, hairline, ears, moles, glasses, jewellery, and expression at matching zoom levels.
Sources
Preservation and technical guidance reviewed for this article.
- arXiv: Old Photo Restoration via Deep Latent Space Translation
- U.S. National Archives: Digitizing Family Papers and Photographs
Related
Ready to restore your photos?
Upload an image, choose your settings, and preview results with a before/after slider.