How to Restore Old Photos With AI (Without Changing the Face)
AI can bring a cracked, faded family photo back to life in one pass — but the whole job is to revive the person, not invent them. Good restoration repairs scratches, tears and water damage, lifts fading, sharpens a soft face and clears grain, and it leaves the subject looking exactly like who they were. The failure mode is a model that “beautifies” Grandma into a stranger — smoother, younger, wrong. Below is the exact order of operations for a faithful restore — scan clean, repair damage, recover the face, colourise honestly, then upscale — plus the honesty rules that keep it recognisably them.
What “restore” means — and where AI overreaches
Restoration and generation are two different jobs. Restoration recovers information that the photo already implies: a scratch runs across a cheek, but the cheek underneath is knowable, so filling it is honest. Generation invents information that was never there: a blurred eye has no true detail to recover, so a model that renders a crisp, confident eye is guessing — and a guess that changes the face is a failure, however sharp it looks. The line between the two is the entire skill. Everything below is designed to keep you on the recovery side of it.
| Damage | What AI can honestly fix | Where it overreaches |
|---|---|---|
| Scratches, creases, tears | Reconstruct texture from surrounding pixels | Redrawing a torn-off eye or mouth from nothing |
| Fading, yellowing, low contrast | Rebalance tone and lift the midtones | Inventing saturated colours the scene never had |
| Soft or blurry faces | Sharpen real edges already present | Hallucinating detail into a face too blurred to read |
| Film grain, dust, speckle | Denoise while keeping skin texture | Smoothing skin into a plastic, poreless mask |
| Black-and-white | Colourise with plausible, muted tones | Guessing exact eye or dress colour as if it were fact |
One rule keeps every restore honest: if the detail is not in the source, it is a guess. Recover what the photo implies; flag anything the model invents — especially a face too damaged to read.
Step 1 — Start from the cleanest scan you can get
The restore is only ever as good as the input, so spend a minute on the source before you touch a model. Scan the print flat at 600 DPI or higher rather than photographing it at an angle — a phone snap adds glare, keystone distortion and its own compression, all of which the model then treats as “detail” to preserve. Wipe dust off the glass, capture the whole print including a little border, and save a lossless copy (PNG or TIFF) so you are not restoring on top of JPEG artefacts. A clean, high-resolution scan gives the model real edges to work from instead of noise to guess at.
No scanner? Shoot the print flat in soft, even daylight, phone parallel to the surface, no flash. Even lighting beats a high-resolution scan riddled with glare — the model cannot tell a hotspot from a highlight.
Step 2 — Repair the damage before you touch the face
Work in the order a conservator would: structural damage first, cosmetic detail last. Have the model reconstruct scratches, creases, tears and missing corners by reading the texture around each break — a crease across a wall or a jacket is fully recoverable because the surrounding pattern tells the model what belongs there. Then lift the fading: rebalance the tone, recover the midtones, pull contrast back into a flat, yellowed print. Do damage repair as its own pass and check it before moving on, because a model asked to “fix everything at once” will trade a clean tear repair for a smoother, less faithful face.
- Scratches & creases: reconstruct from surrounding texture — this is pure recovery, safe to run aggressively.
- Tears & missing corners: fillable where context exists; a tear through a face is the one place to slow down and verify.
- Fading & yellowing: rebalance tone and lift midtones rather than cranking saturation, which invents colour.
- Stains & water damage: treat like scratches where the pattern behind them is knowable; flag where it is not.
Step 3 — Recover the face without changing it
This is the step that decides whether the restore is a success or a stranger. Sharpen the real edges that are already present — the line of a jaw, the shape of an eye, the set of a mouth — and denoise the grain sitting on top of them, but do not let the model redraw the face. The tell of an over-restored portrait is a face that is suddenly younger, symmetrical and smooth, with every pore gone: that is generation wearing restoration’s clothes. Keep the asymmetries, the wrinkles, the exact spacing of the features — those are the likeness. If a face is too blurred to recover honestly, it is better to leave it soft than to invent a confident, wrong one.
Beware the “beautify” drift: models trained on modern selfies will quietly slim a jaw, open the eyes and erase every line. Compare the output to the source at 100% — if you would not recognise them across a room, the model invented a new person. Dial it back.
Step 4 — Colourise black-and-white honestly (optional)
Colourising a monochrome portrait is the most seductive and the most dishonest step, so treat it as interpretation, not fact. The model has no way to know the true colour of an eye, a dress or a wall — it produces a plausible palette, not the real one. Ask for muted, period-appropriate tones rather than saturated modern colour, keep skin natural, and accept that you are making an educated guess. Colourise when it helps a family connect with a face; keep the crisp black-and-white version too, because it is the one that is actually true to the moment.
| Colourise it when… | Keep it black-and-white when… |
|---|---|
| A family struggles to relate to a monochrome ancestor | The photo is a formal or historical record |
| You want a warm, shareable keepsake | Accuracy matters more than warmth |
| Clothing and setting give strong colour cues | The scene gives no honest colour signal |
| You keep the true B&W original alongside it | The grain and tone are part of the story |
Step 5 — Upscale and export at print size
Restoration and resolution are separate jobs — do the repair first, then enlarge the clean result so you are upscaling detail, not amplifying damage. A repaired print at 800px looks fine on a phone and falls apart on a wall, so finish with an upscaling pass to reach 4K or true print dimensions. Export a lossless master (PNG or TIFF) for archiving and framing, plus a right-sized JPEG for sharing. For the mechanics of resolution and how to avoid an over-sharpened, crunchy enlargement, see how to upscale an image with AI; for choosing a general cleanup tool, the best AI photo enhancer roundup goes deeper.
Restore, then upscale — never the reverse. Enlarging a damaged scan first just makes the scratches and grain bigger and harder for the repair pass to read as damage rather than detail.
Do it in Photo Revival Studio
A raw image model can do any one of these steps, but you have to re-type the whole discipline — “repair scratches, keep the face identical, muted colours, no beautify” — into every generation and hope it holds the line. ReelWand’s Photo Revival Studio carries that restraint as a permanent, server-side style DNA: the repair-don’t-reinvent rule, the keep-the-likeness constraint and the honest-colour bias are assembled into every request, so the model recovers your relative instead of replacing them. Drop the scan in, and session memory means your next instruction — “warmer skin,” “leave the grain” — iterates on the last render instead of restarting. When a face needs enlarging for a frame, hand the clean result to the Upscale Engine for a print-size finish. Image generation runs on a credit system priced so restoring a whole shoebox of photos stays cheap.
Drop in a cracked, faded scan and bring the person back — without changing their face.
Restore a photo in Photo Revival StudioFrequently asked questions
How do I restore an old photo with AI without changing the person’s face?
Repair the damage — scratches, tears, fading — as its own pass, then sharpen only the real edges already in the face rather than letting the model redraw it. Keep the asymmetries, wrinkles and exact feature spacing, because those are the likeness. Always compare the output to the source at 100%: if the person looks younger, smoother or symmetrical, dial the restoration back.
Can AI colourise a black-and-white photo accurately?
AI produces a plausible colour palette, not the true one — it cannot know the real colour of an eye, a dress or a wall, so it is making an educated guess. Ask for muted, period-appropriate tones and keep skin natural for the most believable result. Treat colourisation as interpretation and keep the true black-and-white original, which is the version that is actually faithful to the moment.
What resolution should I scan an old photo at before restoring it?
Scan flat at 600 DPI or higher and save a lossless PNG or TIFF, not a JPEG, so you are not restoring on top of compression artefacts. A clean, high-resolution scan gives the model real edges to recover instead of noise to guess at. If you only have a phone, shoot the print flat in soft even daylight with no flash — even lighting matters more than raw megapixels.
Should I restore or upscale an old photo first?
Restore first, then upscale. Enlarging a damaged scan before repair just makes the scratches and grain bigger and harder for the repair pass to read as damage. Once the print is repaired and the face recovered, run an upscaling pass to reach print or 4K dimensions for framing.
Can AI recover a face that is completely blurred?
Not honestly — if a face is too blurred to read, there is no real detail to recover, so any crisp result the model produces is invented rather than restored. It is better to leave a soft face soft than to render a confident, wrong one that no longer resembles the person. AI restoration recovers information the photo implies; it should never fabricate a face that was never captured.
Put it into practice
Specialized image agents carry the craft this guide describes. Pick an available agent and start creating.