Old Photo Restorer

The three levels of old photo restoration

For a portrait where the face is damaged, Full is usually right: the model rebuilds features from surrounding structure, and doing it half-way leaves a visible seam.

How to restore an old photo in three steps

1. 1Scan at the highest resolution you can600 dpi for a print up to A5, 300 dpi at minimum. A phone photo of a photo works if you shoot straight on, avoid reflections, and fill the frame with the print.
2. 2Pick a level and uploadJPG, PNG or WEBP up to 4 MB. If the scan is larger than that, crop to the print itself first: borders and mounts waste resolution and confuse the repair.
3. 3Judge the face first, then the damageA restored photo should still look like the same person. If the face has drifted younger or smoother, drop a level or add "keep the original features, do not beautify" to the note, and run it again.

What restoration can and cannot recover

Restoration can rebuild what has a pattern: creases that run across a jacket, a scratch through a plain background, missing corners of an empty sky, speckles of mould on a wall. It can also recover tone, because fading is a global shift and contrast can be stretched back.

It cannot recover what is gone entirely. A face worn away by chemical damage, a hand cut off by a torn corner, or text that has faded to nothing has no structure left, and anything the model draws there is invention. That distinction matters when the photo is evidence, an heirloom for identification, or a historical record: repair is honest, invention is not.

Old photo renovation: the workflow that works

• Scan or photograph the original print, never a photocopy of a photocopy.
• Crop to the print edge before uploading, so the model works on the picture only.
• Restore at Full if there is visible damage, Light if it is only faded.
• Upscale the restored file afterwards, which is free here and gives you a print-ready size.
• Colourise only after restoration and upscaling, so colour lands on clean structure.
• Keep the untouched scan: restoration is an interpretation, and the original is the record.

Common mistakes that ruin a restoration

The first is over-sharpening. Old prints are soft by nature, and forcing sharpness produces halos around every edge. The second is over-smoothing: denoise too hard and the grain that makes a photograph feel like a photograph disappears, leaving a waxy surface. The third is beautifying faces, which is the single most common failure of AI restoration: it makes everyone look twenty-five and related to each other.

The fourth is colourising a damaged scan. Colour on top of scratches turns the scratches into coloured streaks. The fifth is restoring a low-resolution file and expecting print quality; the pixel count has to be there before the repair can be detailed.

Free AI old photo restorer: cost and quality

The page is free to use and the model call costs credits, shown next to each model before you run it. New accounts start with free credits, which covers several images, and the cheapest model is genuinely capable on light damage. The premium image to image models hold small faces, fabric pattern and handwriting noticeably better, which matters on a portrait you plan to print.

A sensible budget approach for a family archive of twenty photos: restore everything on the cheapest model at Light to see which ones actually need more, then spend on the three or four that carry the most meaning.

Printing, archiving and limits

Output keeps your aspect ratio and renders at up to 1344 pixels on the longest edge. That is enough for a screen and a small print; for anything larger, run the restored file through the free upscaler to 4096 pixels, then print at 300 dpi. Restored files are usually worth keeping as lossless PNG rather than JPG.

Uploads are capped at 4 MB, and images are sent only for the duration of the request; results are stored so you can download them, with no public gallery and no email capture.

Assessing damage before you restore

Look at the print in raking light before scanning: it reveals creases that are invisible head-on, which is exactly what a restoration model needs to see. Sort the damage into three groups. Surface problems, such as dust, foxing and light scratches, are cosmetic and respond to the light setting. Structural problems, such as tears, missing corners and deep creases, need the full setting because the model has to rebuild geometry. Chemical problems, such as silvering, mould or a face worn away by damp, are where the underlying information is gone and no setting will bring it back honestly.

That sorting tells you what to expect. A print with surface damage only will come back looking close to new. A print with a damaged face will come back looking plausible but not identical, and that difference should be disclosed if the photo is an heirloom people will compare against memory.

Scanning an old print properly

A flatbed scan at 600 DPI is the gold standard for a print up to A5, and it captures surface texture that a phone camera flattens. Where no scanner is available, a phone works if you follow four rules: shoot straight on with the camera parallel to the print, use diffuse daylight rather than a flash, fill the frame with the photo, and use a tripod or rest the phone on a stack of books.

Two details make a large difference. Remove the print from a glossy album page if you can, because the plastic adds reflections the model reads as damage. And scan in colour even if the photo is black and white, since the paper tone carries information about fading that a greyscale scan throws away.

Archiving the restored file

Save the restored image as PNG rather than JPG for the master copy, so repeated exports do not add compression artefacts to skin and sky. Keep three files per photo: the untouched scan, the restored version, and any colourised variant, with the date of each in the filename. Memory is unreliable, and in ten years the difference between the original and the restoration will not be obvious to anyone who did not do the work.

If the photo is going into a family book or a frame, also record what you know about it, because restoration preserves the image and not the context. A name, a place and an approximate year on the back of the file are worth more than an extra pass of sharpening.

Group photos, wedding albums and damaged corners

Group photographs are the hardest restoration case because there are several faces, and a model that rebuilds one convincingly can leave another looking drawn. Restore the whole frame first, then check each face at 100 percent; if one has drifted, crop that face into its own file and run a second pass, then composite it back. It is more work than a single click and it is the difference between a usable print and an uncanny one.

Torn corners and missing edges are easier than they look, because the model only needs to continue a background. Give it a hint about what should be there, such as the wall or the tablecloth, and reconstruct. For album pages where the print is glued down, scan the page and crop to the print edge before restoring, so the mount is not interpreted as damage.

Keep going with the rest of the toolkit