What AI photo restoration cannot fix
AI handles most common damage well — fine scratches, dust, fading, blurry faces, missing color. But it has clear limits. Knowing them upfront saves you a credit and prevents disappointment.
You always get a free watermarked preview before paying. We also offer a 24-hour refund. Still, the cases below tend to fail across every AI restoration service on the market, including ours.
1. Group photos with very small faces
School class portraits, wedding parties, large family gatherings — any photo where faces are smaller than a fingertip when the print is held at arm's length. The face-restoration models (CodeFormer, GFPGAN) were trained on close-up portraits. When the face is under roughly 80×80 pixels in the scan, the AI guesses too aggressively and produces a smooth, generic, often unsettling result. We show a warning on the upload page for this reason.
Better approach: crop the scan to one or two faces at a time, restore each separately, then assemble the result manually. Or accept that group photos restore the background and clothing well, but the faces stay roughly as the original.
2. Large physical damage and missing pieces
Big tears, lost corners, water stains that ate through the emulsion, chemical burns — the AI can only enhance what is visible. It cannot recreate missing image content from nothing. A tear through a face will be smoothed but the underlying features cannot be invented truthfully. Treat heavily damaged photos as a partial recovery.
3. Glare, reflections, and overexposure
When the original was photographed through glass, or under a strong flash, the bright spots are baked into the pixels. AI cannot differentiate “face” from “light blob on face” in those zones. Same for fully blown-out sky, white t-shirts that lost detail, or windows behind a subject. If detail was never captured, it cannot be restored.
4. Hands, uniforms, text, and small objects
Face-restoration models are excellent on faces and only faces. When they pass over hands, military insignia, jewelry, name tags, or printed text, they often introduce visible artifacts — extra fingers, melted buttons, garbled letters. The upscaler helps, but if the small object was unclear in the original, the result will be a clean-looking but fictional version of it. Genealogy use cases (medals, dates on documents) are particularly affected.
5. Colorization on texture-heavy backgrounds
The colorization model (DDColor) was trained on natural scenes — skin, clothing, sky, foliage. It struggles with woven walls, brick patterns, ornate carpets, or carved wooden surfaces, sometimes painting them in odd pink or orange tints. The subjects of the photo come out right, but the background may need a manual touch afterwards. If your photo has a heavily textured backdrop, expect a less natural color on that part.
6. Very dark or night scenes
Night-time photographs, dimly lit indoor portraits, and silhouettes can't be brightened without inventing detail. AI will brighten the overall image, but the faces and clothing in shadow will look painted-on rather than recovered. Daytime and bright indoor scenes give much stronger results.
How to get the best result
- Scan flat at 600 DPI or higher — phone snapshots add glare.
- Crop close to one or two faces for portraits.
- Start with the “subtle” face setting before trying “strong”.
- Disable colorization if the photo is already color.
- Use the free preview to judge — pay only when you like it.
Try restoration
Upload your photo and preview the result before paying. If the preview shows any of the failure patterns above, you can stop without paying.
Frequently asked questions
Can AI restore group photos with small faces?
Usually no. Face-restoration models (CodeFormer, GFPGAN) need faces around 80x80 pixels or larger to produce a believable result. School class photos, weddings, and large family groups end up with smooth, generic-looking faces. The fix is to crop one or two faces at a time and restore them separately.
Can AI rebuild a face if part of the photo is torn or stained?
No. AI can only enhance pixels that exist. A tear or stain through a face is smoothed but the missing features cannot be invented truthfully. Heavily damaged photos restore as partial recoveries, not full reconstructions.
Can AI remove glare or reflections from old photos?
Not reliably. If the original was photographed through glass or under a strong flash, the bright spots are baked into the pixels. AI cannot tell a face from a light blob on the face. Same for blown-out skies or windows behind a subject.
Why do hands and small objects look weird after restoration?
Face-restoration models are trained on faces, not on hands, jewelry, name tags, or printed text. They often produce extra fingers, melted buttons, or garbled letters on those areas. Genealogy details like medals or dates are particularly affected.
Does AI colorize textured backgrounds correctly?
Subjects look natural, but textured backgrounds (brick walls, ornate carpets, wood grain) sometimes pick up odd pink or orange tints. The colorization model was trained on natural scenes. Expect to manually touch up the backdrop if needed.
Can AI brighten a dark or night photo?
AI brightens the overall image, but faces and clothing in deep shadow look painted-on rather than recovered. Daytime and bright indoor scenes produce much stronger results than night scenes or silhouettes.
How can I avoid paying for a bad restoration?
Every photo gets a free watermarked preview before payment. If the preview shows any of the failure patterns above, stop without paying. There is also a 24-hour refund window after purchase.