Fix Old Photo Quality — Pro, on our server
“Bad quality” is a bundle, and unpacking it is most of the work. Put an old photograph on screen at 100 per cent and you can nearly always separate what is wrong with it into six things. The faces are soft, from missed focus or from the print itself. The whole image has faded and drifted towards orange, magenta or yellow. There is visible grain or noise across the flat areas. There is physical damage: a crease, a tear, a water stain, a missing corner. The pixel count is small, so it falls apart when printed at any size. And on top of all that there are often compression artefacts, added not by 1971 but by a messaging app in 2019. Those six have almost nothing in common except that they all read to the eye as “old photo”.
Each defect has a different owner, and this tool owns exactly one of them. Soft and blurred faces are what happens here: RestoreFormer++ rebuilds each face and blends it back, at the same output size, with an identity check on every one — the restoration hub explains that check. A global colour cast is a levels or white-balance adjustment in any editor you already own, done locally in seconds. Grain has its own page at grainy old photos, and it is worth reading before you try to remove it, because grain is not always damage. Physical damage goes to the scratch remover or the stain remover, where you paint a mask over the damage and the surrounding scene is filled in — that is inpainting, a different model doing a different job. Size belongs to the upscaler. And compression artefacts have no fix at all beyond finding a less-mangled copy of the file.
The order matters, mostly because it stops you paying twice. Do the free local work first: crop the scanner bed away with the crop tool and pull the colour cast out in your own editor, because every AI step downstream gets a better input for nothing. Then brush out physical damage, then restore the faces, then add colour if the photograph is monochrome, and enlarge at the very end. Enlarging first is the classic expensive mistake: an upscaler magnifies faithfully, which means it magnifies the stain and the grain along with everything else, and you then have a larger file to run every later step against. If you would rather see the whole sequence written out around a single photograph, the family archive page works through it.
What each step costs, and where the file goes. The AI steps are Cloud HD, our Pro-only, opt-in server-side feature set: a plain face restoration is one ordinary run, a restoration with colour chained on is one HD run, and each object-removal pass is one HD run. Pro Monthly includes 150 runs a month, Pro Annual 200, of which up to 10 may be HD runs; the allowance resets on the 1st in UTC, does not roll over, and a refused or failed run is never charged. Any photograph you send is encrypted in transit, uploaded to our server for that run and deleted when the run finishes, and in all cases within 24 hours, processed in the United States on Cloudflare R2 storage and a Modal GPU, and never used to train anything. Cropping, converting, compressing and on-device upscaling never upload at all.