Old photo restoration prompts and the risk of invented details

There is a particular kind of magic in feeding a torn, faded family photo to an AI and getting back something crisp and colourful. There is also a particular kind of danger, because the model did not recover the lost information — it guessed it. A scratch across a cheek becomes new skin the model imagined. A face softened by age and damage becomes a sharp face the model designed.

AI restoration is not archival restoration. An archivist stabilises what survives and is scrupulous about not adding what is gone. A generative model is built to fill gaps convincingly. For a photograph that is also a historical record of real people, that difference matters a great deal.

What "restore" actually asks the model to do

When you tell a model to restore a photo, you are asking it to look at the damaged pixels and produce undamaged ones in their place. Where damage is small and surrounded by clear detail — a speck of dust on a plain wall — the guess is safe. Where damage sits on top of irreplaceable detail — a tear through an eye, fading across a whole face — the model has nothing real to recover from, so it invents a plausible replacement. Plausible is not the same as accurate.

Write prompts that restrain the guessing

The safest restoration prompts are conservative by design. They tell the model to repair what it can verify and to leave alone what it cannot:

• "Repair dust, scratches, and creases in the background and clothing; do not alter facial features."
• "Recover faded contrast and correct the colour cast, but do not sharpen or redraw soft faces."
• "Where detail is missing, leave it soft rather than inventing new detail."
• "Do not change identity, age, expression, or the number of people in the photo."

A face that is genuinely soft in the original should stay soft. The moment a restoration makes a blurred face crisp, you are no longer looking at a recovered photograph — you are looking at a new portrait the model authored from a hint.

Colourising is interpretation, not memory

Adding colour to a black-and-white photo feels like restoration, but the original never recorded colour at all. Every hue the model chooses — the dress, the walls, someone's eyes — is an educated guess. Restrained, period-plausible colour can be lovely. Just label it honestly to yourself: it is an interpretation laid over the record, not a fact the photograph remembers.

Always keep the untouched original

The single most important habit in AI restoration has nothing to do with prompting: keep the original scan, untouched, and treat the AI version as a separate, derived file. The scan is the record; the restoration is one reading of it. If you ever need to know what the photograph actually showed, you will be glad the evidence still exists alongside the interpretation.

Key takeaways

• AI does not recover lost detail — it invents a plausible replacement for it.
• Restoration is not archival: a generative model fills gaps, an archivist refuses to.
• Conservative prompts repair verifiable damage and leave irrecoverable detail soft.
• Colourising a black-and-white photo is interpretation, not recovered fact.
• Always keep the untouched original scan as the record; the AI version is one reading of it.