Enhance & Restore Old Photos with AI

Why old photos need more than an upscaler

A modern photo that is simply too small has one problem: resolution. An old photograph usually has four at once — grain from the film stock, colour shift from decades of chemical ageing, physical damage such as creases and scratches, and softness from the original optics or the scanning process.

A general-purpose upscaler treats all of that as texture to enlarge. The result is a bigger picture of a damaged photograph. Restoration means separating the damage from the image and rebuilding what sits underneath.

The restoration chain

ImageUpg's restore path runs specialists in sequence rather than asking one model to handle everything:

• SeedVR2 performs the deep reconstruction pass, rebuilding structure lost to age and compression.
• CodeFormer then works specifically on faces. Faces are the reason restoration is hard — we are extraordinarily sensitive to them, and a generic model that produces acceptable fabric will produce an uncanny relative.
• A Topaz pass finishes the image, resolving fine detail after the heavy repair work is done.

Treating faces as their own problem is the single biggest difference between restoration output that feels like your grandmother and output that feels like someone who resembles her.

Scanning matters more than the software

If you are working from physical prints, the scan sets the ceiling for everything that follows:

• Use optical resolution, not interpolated. Scanner software often advertises inflated numbers achieved by the same averaging trick that fails in image editors. 600 DPI optical beats 4800 DPI interpolated.
• Scan flat and clean. Dust becomes permanent detail once reconstruction runs — the model has no way to know a speck was not part of the photograph.
• Save as TIFF or maximum-quality JPEG. A quality-70 scan discards detail before restoration ever begins.
• Do not sharpen in the scanner software. Scanner sharpening adds halos that the restoration model then faithfully preserves.

What restoration can and cannot recover

Being straight about the limits saves disappointment. Restoration handles grain, moderate fading, compression damage, soft focus from scanning and general loss of contrast well. It reconstructs faces convincingly when enough structure survives.

It cannot recover a face that was never resolved in the original, fix severe motion blur, or restore detail from areas that are completely blown out or crushed to black. In those regions there is no information hiding underneath — the model would be inventing a person rather than restoring one.

For heavily damaged images, expect to run more than one attempt at different settings. Restoration involves estimation, and the same file can produce meaningfully different results between modes.

Working with family archives

If you are digitising a box of prints, do the whole batch at consistent scanner settings before restoring anything. Mixed scan quality across an album is far more visible than any individual imperfection, because the eye compares neighbouring images rather than judging each alone.

Frequently asked questions