Restore Old Photos with AI: A Step-by-Step Prompt Workflow

Restoring a damaged family photograph is the most emotionally loaded AI edit there is, and the one with the strictest rule: the people in it must remain exactly who they were. A restoration that invents a plausible face is not a restoration, it is a replacement. The workflow below keeps that line intact by making the model fix one category of damage at a time.

Pass One: Physical Damage

Start with the things that are objectively missing or broken — scratches, creases, tears, dust and torn corners. Nothing tonal, nothing cosmetic. Small, verifiable repairs first.

Pass 1 — Damage: Repair physical damage in the uploaded photograph: scratches, creases, dust, spots and torn edges. Reconstruct missing areas using only the surrounding detail. Preserve every facial feature, expression, hairstyle, garment and background element exactly as photographed, including the original grain. Produce a clean scan of the same photograph. Do not sharpen faces, do not change tones, and do not invent details that are not implied by the surrounding image.

Pass Two: Tone and Contrast

Old prints fade, yellow and lose contrast. Correct that as a photographic operation, not a stylistic one: recover the density in the blacks, restore a white point, remove the colour cast from ageing. Keep the original medium's character — a 1950s print should still look like a print.

Pass 2 — Tone: Correct fading and colour cast in the uploaded restored photograph. Recover black density and a clean white point, remove the yellow ageing cast, and restore natural mid-tone contrast. Preserve identity, grain structure and the character of the original print. Produce an accurate, archival-quality tonal restoration. Do not add saturation, do not apply modern digital sharpening, and do not alter any facial feature.

Pass Three: Colourising, If You Must

Colourisation is interpretation, so treat it that way. Give the model any facts you know — eye colour, the blue of a uniform, the red of a front door — and ask for period-appropriate restraint everywhere else. State explicitly that skin tones should be natural for the subjects' actual complexion, and that unknown items should stay muted rather than guessed brightly.

Pass 3 — Colour: Colourise the uploaded black-and-white photograph with restrained, period-accurate colour for the 1950s. Known facts: the woman's dress is dark green, the car is cream. Use natural, believable skin tones and muted colours for anything uncertain. Preserve every facial feature, expression and detail of the original exactly. Produce a subtle, documentary-style colourisation. Do not saturate, do not modernise the palette, and do not alter faces or clothing shapes.

Pass Four: Gentle Detail, Then Stop

A light clarity pass can help a soft scan, but this is where restorations go wrong. Ask for detail recovery 'only where detail exists in the original', and never ask the model to 'enhance' a face. If the eyes were soft in 1957, they stay soft — that softness is the photograph.

• Scan or photograph the print as flat and evenly lit as you can; a better input needs less restoration.
• Work on a copy, and keep every intermediate version.
• Show the result to someone who knew the person before you print it — they will catch a wrong face instantly.
• For badly damaged areas over a face, accept a partial repair rather than an invented reconstruction.

Four passes, each with a narrow job and a strict preserve clause. It takes ten minutes longer than asking for a miracle in one message, and it is the difference between a restored memory and a convincing stranger.