How to Restore Old Photos With AI Without Losing the Person
A practical guide to restoring old family photos with AI: what to scan, which damage AI can repair, how to check likeness, and what to keep in the archive.
Published
Jump to the editor ↓Short answer: Scan at 600 dpi or higher, keep the untouched file, and let AI repair dust, scratches, creases and fading. Restoration is strong at rebuilding texture around damage and weak at inventing a face that is not in the scan. Check the eyes, nose, jaw and hairline against the original, keep a restored black-and-white copy as the archive, and colorize only after that. The old photo restorer opens the editor with a restoration prompt already written.
A good restoration keeps the original composition and the person’s features; the repair fills damage instead of inventing a new relative.
A family photograph usually has two kinds of damage. There is the damage you can see: a torn corner, a scratch across a face, a crease where the print was folded. And the damage you cannot see: a scan too small to hold the detail you want, or fading that erased the shadow under the chin.
AI restoration helps with the first kind. It reads the pixels around the damage and rebuilds texture, tone and edges, but it cannot recover what is not in the file. Knowing which problem you have is the difference between a good restoration and a portrait of a stranger.
Scan before you restore
The model can only repair what it can see, so the scan sets the ceiling.
- Use a flatbed scanner at 600 dpi or higher. A print scanned at 300 dpi has already lost the texture around eyes, fabric and paper grain.
- Save as TIFF or the highest-quality JPEG you can. Compression adds blocky artifacts that look like damage and can be copied into the repair.
- Keep the untouched scan. Treat it as the negative. Every experiment should start from this file.
- Include the whole print. Cropping in the scanner removes context the model could use to match paper texture.
If the print is torn, scan each piece separately; do not tape the pieces on the scanner glass, because tape adds an artifact that has to be removed later.
What AI repair handles well
Restoration models are best on damage that sits on top of a scene they can infer, which covers most of what families have in a shoebox.
| Damage | What usually happens | What to check |
|---|---|---|
| Dust, specks, hair on the scan | Removed cleanly | That freckles or jewellery were not erased |
| Scratches over sky, walls or clothing | Texture is rebuilt from nearby pixels | That the texture does not repeat in a pattern |
| Creases across background areas | Tone and edges reconnect | That a straight line does not survive across a face |
| Fading and colour casts | Contrast and tone come back | That the period look is kept rather than modernized |
| Large missing areas, especially faces | A plausible shape is generated | That it still looks like the person |
Work in stages, small damage first. Repair the background before the face and save a copy between stages, so a bad result costs one step rather than the whole job.
What restoration cannot do
- A missing half is a guess. If one side of a portrait is gone, the model invents an eye, an ear and a jawline. The result can look convincing and still be wrong.
- Motion blur and out-of-focus faces stay blurred. Sharpening adds contrast but cannot recover features that were never recorded.
- Small prints hold little detail. A passport photo from 1952 scanned at 600 dpi may only contain a few hundred pixels across the face.
- Text and inscriptions are unreliable. Names, dates and handwriting are often better left untouched.
- Colour was never recorded. A faded sepia print cannot tell the model whether a dress was blue or green.
If a service promises to bring back a missing face, ask what source it used. The honest answer is that it generated one.
Check the likeness before you print
The most common failure is not a visible scratch; it is a face that quietly drifts. Models sometimes smooth skin, widen a jaw or invent detail in a dark eye socket. At thumbnail size this looks better; at print size it looks like someone else.
Compare the restored face with the original scan at 100% and look at four things in order:
- Eyes. Shape, spacing, eyelid crease and the direction of the gaze.
- Nose. Width at the bridge and the base, plus the shadow under the tip.
- Jaw and chin. The outline should match the original silhouette, not a tidier one.
- Hairline and ears. These are where models most often invent detail.
Also check permanent marks: a scar, a mole, a chipped tooth, since those are what a beautifying model tends to remove.
If the likeness drifts, regenerate rather than paint over the result:
Restore this scanned family photograph. Repair cracks, scratches, dust and torn edges. Keep the person’s exact facial features, age and expression. Keep the original composition, clothing and background. Do not add people, text, borders or watermarks.
Run the photo through a second model if the first softens the face. EzCreate lets you switch between hosted models such as GPT Image 2, Seedream, Nano Banana Pro and Flux 2 in the same editor.
Restore first, then colorize
Colour is the step people want first and should do last. A colorizer guesses a dress, uniform or wall from tone and context, and does better on a clean image. Restore first, keep the black-and-white file as the archive copy, and treat the colour version as an interpretation. Our guide to the photo colorizer covers period research and honest labelling, and the image extender handles frames that are too tight.
Keep an archive you can trust
A folder with four files covers most projects: the untouched scan, the restored black-and-white version, the colorized version, and a note describing what was repaired.
EzCreate accepts JPG, PNG, WebP or AVIF up to 10 MB. You can upload a photo and open the editor for free; sign in to claim the daily preview credits. Larger runs, downloads and 2K or 4K output use purchased credits. Results are deleted after 7 days, so download what you want to keep. There is no EzCreate watermark.
For printing, a 600 dpi scan of a 4 × 6 inch print is roughly 2400 × 3600 pixels, enough for a similar-sized reprint; enlarging a small original to poster size will show softness.
Frequently asked questions
Can AI restore a photo that is torn in half?
It can rebuild small missing areas, but a large missing half is a guess. Scan both pieces, restore the easy damage first, and compare the rebuilt area with any other photo of the same person.
What resolution should I scan at?
Scan at 600 dpi or higher and keep the untouched file. A higher-resolution scan gives the model more real detail and gives you a better archive copy.
Will AI change the person’s face?
It can. Restoration models sometimes smooth skin or invent detail in dark areas. Compare the eyes, nose, jaw and hairline with the original and regenerate with a more conservative prompt if the likeness drifts.
Should I colorize before or after restoration?
Restore first, then colorize. Repairing cracks and fading gives the color model a cleaner image, and you can keep the restored black-and-white file as the archive copy.
Can I restore a photo when I only have a phone picture of the print?
Yes, with lower expectations. A phone photo adds perspective distortion, uneven light and sensor noise, all of which the model has to separate from real damage. Photograph the print straight on, fill the frame, avoid flash, and keep still.
Sources
- Library of Congress: Care, Handling, and Storage of Photographs
- U.S. National Archives: How to Preserve Family Archives
Try EzCreate now
Upload a scan to the old photo restorer, describe the damage you want repaired, and compare the result with the original at full size before you print or share it.