In a drawer somewhere is a photograph that exists nowhere else. It was taken on film, printed once, and scanned years later at whatever resolution the shop offered, usually 300 DPI, which for a small print is barely more than a screen thumbnail. Everyone who sees it says the same thing: that is amazing, can you make it bigger? For most of the last thirty years the honest answer was no. That answer has changed, and the workflow is worth knowing properly.
Step one: the scan is the master
Everything downstream depends on scan quality, so this is the step where shortcuts are expensive. Scan at the flatbed's maximum optical resolution, not the interpolation setting, which fabricates pixels rather than capturing them. Do not auto-correct away damage; dust and scratches are better handled selectively than by aggressive global processing that softens the whole image. And never scan directly into a compressed format: a lossless master file is the archive, and every derivative is disposable.
For prints too large for the flatbed, phone camera scanning has quietly become acceptable: straight down, even lighting, the phone on a small tripod or stack of books. The discipline is the same: capture the maximum resolution available, keep the file lossless, and resist the temptation to apply filters before the AI pass. Reconstruction works from raw evidence; preprocessing destroys evidence.
Step two: repair before you enlarge
Order matters. Dust spots and scratches that are invisible at scan size become glaringly obvious at 4×, because the network sharpens them along with everything else. Fix physical damage first: clone out scratches, spot-lift dust, and straighten the frame. These are five-minute operations in any editor, and doing them small means doing them accurately.
Fading is worth a separate pass. Old prints fade unevenly, magentas first and then yellows, and while reconstruction restores detail brilliantly, it has no way to know what color the sky was in 1962. A global color and contrast correction on the scan gives the network truer input, and the difference in the restored result is dramatic. Think of it as feeding the machine the best possible evidence rather than asking it to compensate for bad evidence.
Step three: right-size the upscale
Match the enlargement to the purpose. A frameable 8×10 from a small scan typically needs about 4×, enough to reach print resolution without pushing reconstruction past where its priors get speculative. Digital display needs less: 2× covers most uses, and smaller scales preserve more authentic texture. The temptation is always to go bigger, and the discipline is that reconstruction quality and enlargement factor trade against each other; a modest scale on a good scan outperforms an aggressive scale on a poor one.
Zoom to 100% on every face before exporting. Human face perception is the most sensitive instrument you own: it catches reconstruction drift that no metric will flag, and it catches it instantly.
Step four: verify honestly
Restoration is reconstruction, and reconstruction occasionally invents. The failure modes are predictable: fine text on documents becomes plausible but wrong; dense patterned fabric may emerge with a weave that was not in the original; very blurry faces can drift toward a generic version of themselves. None of this makes restoration wrong; it makes verification mandatory. Flip between original and restored at 100%, inspect the faces, and where the restored version has drifted from the person you remember, back off the scale or disable face enhancement for that image.
- Compare at 100% zoom, never at fit-to-window, because preview scale hides every artifact.
- Faces, hands, and text are the three drift-prone zones; check all of them.
- When in doubt, keep two exports: the aggressive restoration and the conservative one.
- The original scan is sacred; every operation happens on a copy.
Why this work should never leave your machine
There is a specific, underweighted privacy dimension to archival restoration. Family photographs contain people who cannot consent: children now grown, relatives now gone. They reveal homes, addresses, possessions, calendars of family gatherings. They are frequently tied to documents: a passport photo beside a passport scan is identity data in aggregate. Sending a shoebox of these to a free cloud service is a decision almost nobody would make consciously, yet the convenience of those services makes it the default.
The sensitivity compounds across a collection. One old photograph is nostalgic; forty of them, uploaded to the same account, sketch a family tree, a decade of addresses, and a set of faces that biometric systems can match today. This is not a theoretical exposure: archived imagery is exactly the material data brokers pay for and the material that surfaces in breaches years after the user forgot the upload happened. Local processing does not merely reduce this risk; it makes the category of exposure impossible by removing the transmission step entirely.
Local processing removes the trade entirely. The scans stay on the drive they were made on, the restoration runs on the machine in front of you, and the only network traffic is the one-time download of the network files. For this particular genre of image work, private by definition and sentimental by nature, that is not a feature comparison. It is the only acceptable architecture.
The output that makes it worth it
The payoff is concrete, and it is not subtle. A 4×6 print from 1975, scanned at modest resolution, typically lands around 800 pixels wide, enough for a phone screen but short of everything else. Restored at 4×, it arrives at 3200 pixels: a frameable enlargement, a viable canvas for a family slideshow at reunion resolution, a scan good enough that the great-grandchildren will see the faces rather than the blur. The photograph was always that detailed. The print, the scan, and the decades conspired to hide it, and it takes one drop into the right tool to get it back.
A final note on quality thinking
The practical rule of thumb for restoration is a simple one, and it should be written on the wall: trust the edges, verify the faces. Edges are where reconstruction genuinely creates detail from statistics, and they usually look right; faces are where a viewer keeps looking the moment you catch their eye, and where drift is most visible. A scan that looks beautiful in the corners but slightly off in a facial close-up is not done yet. The two checks sidestep each other, and together they produce work you can trust with your family.
