Restore a scanned photo

Scanning well is half the work

Everything downstream depends on how many real pixels the scan contains and how clean they are. A restoration model cannot work with detail the scanner never captured.

• Set the resolution from the size of the original, not from a fixed habit. For a small print, the standard 10 by 15 cm snapshot, scan at 600 dpi: that gives you roughly 2400 by 3500 pixels, about 8 megapixels, which is enough to rebuild a face and to print the result larger than the original. For bigger originals, an A4 portrait, an 8 by 10, a full album page, 300 dpi is already enough, because the print is physically larger and 300 dpi across it still produces thousands of pixels. Aim for somewhere between 3000 and 4000 pixels on the long side and you will not go wrong.
• Ignore interpolated resolutions. If the hardware is 600 dpi optical, scanning at 2400 dpi gives you a much bigger file, a much slower scan, and not one extra piece of detail.
• Scan in colour even when the photograph is black and white. A 24-bit RGB scan records three channels instead of one, so paper tone, yellowing, foxing and stains are separable from the image, and you keep the option of working from the cleanest channel. Many scanners also apply their own processing in grayscale mode. Converting to neutral grey afterwards takes a second; undoing a grayscale scan is impossible.
• Save TIFF, or JPEG at maximum quality. The email and web presets in scanner software compress hard, and those artefacts are the first thing a denoiser preserves, because they look like structure. Airpic takes TIFF directly.
• Clean before you scan. Wipe the glass with a dry microfibre cloth and lift dust off the print with a blower or a soft brush rather than rubbing it, since an old emulsion can lift. Dust specks are the single most common thing people mistake for damage once the scan is on screen.
• Turn the automatic corrections off. Dust removal, sharpening, colour restoration and auto levels in scanner drivers all throw away information the restoration models could have used, and they do it before you ever see the file. The one exception is descreening, worth enabling when you are scanning a halftone image printed in a newspaper, a magazine or a yearbook.
• Lay the print flat and close the lid. A curled print produces a soft band along the curl; weight the corners if you have to. No scanner at hand? A phone photo works as a fallback: even indirect light, no flash on a glossy print, camera parallel to the photo, frame filled.

The restore tool accepts up to 20 megapixels, which comfortably covers a 600 dpi scan of a small print or a 300 dpi scan of an A4.

What AI can recover, and what it makes up

What the models genuinely do well: rebuild facial features that grain, blur or low resolution had smeared, the eyes, the lashes, the line of the mouth, the edge of the jaw; remove film grain and scanner noise without flattening the image; recover apparent sharpness; add plausible colour to a black and white photograph; and raise resolution so the result can be printed larger than the original was.

What they cannot do is recover information that was never recorded. A tear, a missing corner, a crease that cracked the emulsion, a highlight burnt out to pure white, a face turned away from the camera: there is no data in those areas, in any channel, at any bit depth. The model fills them with something plausible instead, a face that fits the shape of the head, skin that continues the texture around the gap. Plausible is not the same as true.

That has a practical consequence for family archives. A restored image is a reconstruction, not a document. Keep the raw scan alongside it and label which is which, because in twenty years nobody will remember which file was the original. Printing the restoration for a relative is a lovely thing to do. Using a photograph as evidence of what somebody actually looked like is a job for the scan.

Colourization is the clearest case of the same problem. The model picks colours that are statistically likely for the shapes it recognises: skin, sky, grass, wood, common fabrics. It has no way of knowing the dress was green or the eyes were blue. If you know, a colourized version that contradicts the family memory is worse than a good black and white one.

One rule of thumb about faces: if a face in your scan is only forty or sixty pixels across, there is barely anything to work from, and the model leans heavily on what it learned about faces in general. That is precisely where reference photos earn their place, and it is covered below. The guide on restoring old photos with AI works through more examples.

The right order: noise, then faces, then colour

If you run the steps separately, the order changes the result. Noise first, faces second, colour last.

Denoise first. Film grain and scanner noise are high-frequency detail, and a face model handed a grainy crop cannot tell grain from skin texture. It either bakes the grain into the reconstruction or smooths it away together with the real detail, and the geometry it infers from a noisy image is less accurate to begin with. A clean input gives it a clean signal. Denoise is a separate tool with its own strength slider, so you can run it first and feed the result into restoration.

Restore faces second. The face step works on an aligned crop of each face and blends the rebuilt version back into the photograph, while the rest of the image is upscaled at the same time. Worth knowing: running a face model twice does not double the quality. Each pass pulls the face further towards the model's idea of a face, so two passes drift further from the person, not closer.

Colour last. A colourization model decides what is skin, hair, fabric, foliage and sky by reading shapes and edges. Give it a noisy, smeared input and those regions are ambiguous, so the colour bleeds: blotches on cheeks, a jacket that changes shade halfway down, grass creeping up a wall. On a clean, sharp grayscale image the regions are unambiguous and the colour lands where it belongs. Colorize is also available on its own, with natural and vivid modes, if you want to do that step by itself.

Finally, restraint beats maximum settings. On a badly degraded original a face model at full strength produces a clean, sharp, convincing face that is not quite the person. Pulled back, it keeps more of the original geometry and expression and still reads as restored. Somewhere between 60 and 85 percent is a sensible place to start, and it costs nothing to compare before you commit.

Restoring in Airpic does all three in one pass with checkboxes, which is the right choice for a reasonably good scan. Split the steps when the scan is noisy or the colour comes out wrong.

When reference photos are worth it

There is a specific failure everyone hits eventually. You restore a badly degraded portrait, a small print, a photo of a photo, a screenshot forwarded through a messaging app, and what comes back is a clean, attractive, entirely generic face. The hair is right, the shape of the head is right, and it is not your grandmother. That is not a bug: with almost no information in the original, the model has nothing to anchor on and falls back on the average face it learned.

Reference photos supply the anchor. Give the tool one to three photographs of the same person and the reconstruction is guided towards that identity instead of towards the average. One reference is often enough; three is the maximum.

What makes a good reference:

• The same person, face clearly visible and reasonably sharp.
• A similar age to the photo you are restoring, if you have one. A different pose and different lighting are fine, helpful even.
• Avoid sunglasses, heavy shadow across the face, extreme angles, and group shots you would have to crop hard.
• A modern phone photo makes a perfectly good reference.

The fidelity slider controls how much of the rebuilt face is blended back over the original. At the top the reconstruction dominates: maximum detail, but the expression can be smoothed away, a smile flattening, a squint opening. Lower down, more of the original geometry, expression and skin tone survives, and the result looks like a restored photograph rather than a new one. With references, around 85 percent is a good starting point.

The tool produces several candidates and scores each one for similarity to your references. Treat the score as a guide rather than a verdict, particularly if you only supplied one reference photo. Look at the candidates and pick the one that looks like the person you remember, because nobody else can make that judgement for you.

Step by step in Airpic

1. Scan as described above: 600 dpi for a small print, 300 dpi for a large original, in colour, saved as TIFF or maximum-quality JPEG.
2. If the scan is grainy, run it through denoise first and download that result.
3. Open restore old photos and drop or paste the file. It takes JPG, PNG including alpha and 16-bit, WebP, TIFF, HEIC and AVIF, applies the EXIF orientation and keeps the colour profile. Up to 20 megapixels.
4. Tick improve faces. Tick colourize as well if the photo is black and white and you want colour; leave it off if you would rather keep it monochrome.
5. Choose the output scale. Doubling the pixel dimensions is what you want if the result is going to be printed.
6. If the face is badly degraded, add up to three reference photos of the same person.
7. Set the fidelity, process, and compare the before and after in the viewer. Open the 100 percent view and look at the eyes and the mouth; that is where a reconstruction gives itself away.
8. Pick a candidate if several were produced, then download.
9. Keep the original scan. Always.

For a large print, put the finished file through the upscaler afterwards: it goes up to 600 megapixels and has a mode aimed specifically at large-format printing. If the file is then too big to email, compress is free to process and download.

No account and no subscription. Processing and seeing the result are free; full downloads are covered by a free allowance each month and then by credits that do not expire. Files are deleted an hour after you download them and one day at the most, they are never used to train anything, and the GPUs run in the European Union, which matters more than usual when the photographs are of your own family.

FAQ

Keep reading