How to restore old photos with AI
A scanned photo from the sixties, a WhatsApp capture of a grandfather's portrait, a tiny ID picture: each loses different information and is restored differently. Here is what AI really recovers, what it invents and how to keep the likeness of the person.
Start from the best copy you can get
Scan the print at 600 dpi if it is small (ID size, 9 × 13 cm) and at 300 dpi if larger; save as TIFF or PNG, not JPEG. If all you have is a phone photo of the print, take it in diffuse light, without flash, framing only the image. Avoid WhatsApp forwards: each one compresses again.
What each restoration step does
- Sharpness and cleaning: a super-resolution model (Real-ESRGAN) removes grain, scanner noise and compression and rebuilds textures of clothing, hair and background.
- Faces: a dedicated model (GFPGAN) rebuilds eyes, mouth, skin and hair from what is left. It is the step with the most “invention”, which is why Airpic offers a fidelity control to blend the reconstruction with the original and, since September 2026, matches the tone and blends the face seamlessly.
- Colour: DDColor colourises black-and-white photos plausibly (skin, sky, vegetation); the exact colour of a dress is a guess.
- Scratches and stains: paint over them with “Erase objects” (LaMa), which fills in from the surroundings.
When the face does not look right: reference photos
With heavily degraded inputs, any face restorer tends to produce a generic face. On Airpic you can add one sharp photo of the same person (up to three, from any period): the face is rebuilt guided by their identity with FLUX.2 klein, four candidates are generated, likeness to the references is measured and you choose. It works best with frontal, well-lit references; expression and pose come from the photo you are restoring, not from the reference.
Step by step on Airpic
- Upload the scan or capture (JPG, PNG, TIFF, HEIC).
- Leave enhance faces on; add colourise if the photo is black and white.
- Set the intensity: at 100 % the face is the reconstruction; between 60 and 85 % it keeps more of the original.
- Optional: add a reference photo and pick the candidate that looks most like the person.
- Compare before and after with the divider, view at 100 % and download.
What not to expect
AI does not recover what is not there: a 20-pixel face does not hold enough information to know the exact shape of the nose, and the result will be a plausible face, not the real one. It also cannot fully undo heavy motion blur or read erased text. For documents and text, use Unblur; for very grainy photos, Denoise before restoring.
FAQ
Can a very pixelated WhatsApp photo be restored?
Yes, within limits. Sharpness and colour improve clearly; the face is rebuilt, and to keep the likeness it helps to add a sharp reference photo of the same person.
Does AI restoration change the person's face?
It can when the input is heavily degraded. That is why there is a fidelity control that blends the reconstruction with the original, and restoration with reference, which guides the face with a photo of the same person.
At what resolution should I scan an old photo for restoration?
At 600 dpi if it is smaller than 10 × 15 cm and at 300 dpi if larger, as TIFF or PNG. Airpic accepts up to 20 megapixels in Restore.
How much does restoring an old photo cost on Airpic?
Previewing is free. Downloading uses your free download of the month or one standard credit (from €0.099); with a reference photo, one large-format credit (from €0.49).