ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›OLD PHOTO RESTORATION PROMPTS
ARTICLES · HOW-TOBEGINNER

Old Photo Restoration Prompts: Copy-Paste Prompts That Keep the Face

Copy-paste prompts to repair scratches, fix faded colour and colourise old photos without changing faces, plus how to check results against the original.

READ11 min
WORDS2,190
SECTIONS8
TYPEHOW-TO
CHECKED12 OCT 26

An old photo restoration prompt tells an AI image tool to repair damage (scratches, tears, fading) while keeping the people exactly as they were. The prompts below name what to fix and list what must not change. Keep the original and compare every result against it.

One warning: chat image tools such as ChatGPT and Gemini do not clean your photo. They draw a new picture that looks like it, so faces can change.

◈ IN PLAIN TERMS

A restorer with a brush works on your actual photo. An AI image model works more like a skilled artist copying it: it paints a fresh picture, fixing the damage as it goes. At a glance the copy looks the same. Up close, a nose, an ear or a smile may be slightly different. That is why every prompt here repeats “keep the face exactly as it is”, and why the last step is always checking the copy against the original.

TL;DR — THE SHORT VERSION
  • Keep the original file untouched. Work on copies.
  • Start with the faithful prompt: fix damage, change nothing else, do not beautify or modernise.
  • One job per round: scratches first, then colour, then anything else.
  • Colourising is guesswork, and “make it look recent” invents detail.
  • Check faces against the original, and say it was AI-restored when you share it.
◈ THE QUICK LIST: 8 RESTORATION PROMPTS, BY JOB
  1. The faithful restoration start here: repair only the damage and keep everything else exactly as it is.
  2. Scratches, creases and tears fill each damaged area so it matches what is right next to it.
  3. Faded colour recovery correct a yellow, orange or pink cast without repainting anything in new colours.
  4. Colourising a black-and-white photo natural, muted colours; the colours are guessed, so name any you know.
  5. Faint or blurry faces gently clearer faces, each person kept recognisably the same.
  6. Group photos state how many people there are, and keep everyone in place with the same face.
  7. "Make it look like it was taken yesterday" the tool invents new detail, so treat the result as an artwork based on the photo.
  8. The follow-up line for each later round: change only one thing and keep the rest as in the last image.

Then check it against the original.

Before you start

The plain point: a good copy of the original, a backup, and permission if the photo shows someone else.

  • Get a clean copy. Scan the print, or photograph it flat in daylight with no glare. Crop off the table or album page.
  • Never overwrite the original file. Every AI result is a new file.
  • Think about whose face it is. OpenAI’s ChatGPT guide asks you to confirm you have permission to use a real person’s likeness.1 Google’s Gemini help says editing images is not available to users under 18.2

Why AI restoration can change a face

The plain point: the tool redraws the whole picture each time, and small differences add up over several rounds.

OpenAI says so in its prompting guide: repeated edits can still change details you meant to keep, so restate what must stay the same and inspect each result.3 Its advice for edits is to say “change only X” and list the details to keep, and for anything that must stay exactly as it was, to paste the approved fix back into the original rather than trust the prompt alone.3 The prompts below follow that pattern.

TAKEAWAY

A prompt lowers the risk of a changed face. It cannot remove it. The check at the end is not optional.

The restoration prompts

The plain point: start with the faithful prompt, then use one other for a single extra job. Change the parts in [brackets].

1. The faithful restoration (start here)

USE WHEN: you want the photo repaired, not updated
Restore this old photograph. Repair only the damage: scratches, dust spots, creases, small tears and stains. Keep everything else exactly as it is: every face, expression, hairstyle, body shape, pose, clothing, background, framing and the original lighting. Do not beautify, slim, smooth skin, whiten teeth, add makeup or change anyone’s age. Do not modernise the clothes, hair or setting. Keep the original grain, softness and tone ([black and white / sepia / faded colour]). If part of a face is missing or unclear, leave it soft rather than inventing new detail.

2. Scratches, creases and tears

USE WHEN: the print is cracked, creased, torn or spotted
Repair the physical damage in this photo: the [crease across the middle / tear in the top left corner / white scratches / water stain]. Fill each damaged area so it matches what is right next to it, with the same grain, sharpness and tone. Change only the damaged areas. Do not change any face, expression, pose, clothing, background or the overall brightness.

If a tear runs through a face, the tool has to guess what was under it. Check that face hardest.

3. Faded colour recovery

USE WHEN: a colour photo has faded or turned yellow, orange or pink
This colour photo has faded and has a [yellow / orange / magenta] colour cast from age. Correct the colour cast and restore natural contrast, as the photo would have looked when first printed. Keep the colours the photo already has. Do not repaint clothes, hair, eyes or objects in new colours. Do not change any face, pose, clothing, background, framing or grain.

4. Colourising a black-and-white photo

The colours are guessed. A black-and-white photo records how light or dark each spot was, not its colour, so a red dress and a green dress can be the same grey.Reasoning, 12 Oct 2026 — a black-and-white photograph stores brightness only, so colour added afterwards is an estimate, not a recovery; the site’s explanation, not a vendor claim. If you know a real colour, put it in the prompt.

USE WHEN: you want a colour version of a black-and-white photo
Colourise this black-and-white photograph with natural, muted colours that suit [the 1950s]. Known colours: [the woman’s coat was navy blue; the car was dark red]. For everything else, choose quiet, believable colours and avoid bright, saturated tones. Keep everything else exactly as it is: every face, expression, pose, clothing shape, background, framing, grain and sharpness. Do not beautify or modernise anything. Add colour only.

Keep the black-and-white restoration too, and say the colours were added by AI when you share the colour one.

5. Faint or blurry faces

USE WHEN: the photo is fine but the faces are faint or soft
Gently improve the clarity of the faces in this photo, keeping each person recognisably the same. Keep each face’s exact shape, eyes, nose, mouth, ears, expression, age, wrinkles, freckles, glasses and hairline. Do not make anyone look younger, thinner or more attractive. Where detail is missing, keep it slightly soft instead of inventing sharp new detail. Do not change anything outside the faces.

A face that suddenly looks crisp and “perfect” is usually a new face.

6. Group photos

USE WHEN: there are several people, especially small faces at the back
Restore this group photo of [number] people. Repair only the damage: [scratches / fading / creases]. There are exactly [number] people. Keep every person in the same place, with the same face, expression, height, pose and clothing. Do not add, remove, merge or move anyone. Do not make the faces look alike. Keep the background, framing, lighting and grain as they are. Where a small face is unclear, leave it soft.

7. “Make it look like it was taken yesterday”

An old photo does not contain the detail a modern camera records, so the tool must invent it: skin texture, eyelashes, fabric weave. The result is a new picture based on the old one.Reasoning, 12 Oct 2026 — detail the original never recorded can only be generated; the site’s explanation, consistent with OpenAI’s note that edits can change details you meant to preserve (Image prompting, read 12 Oct 2026).

USE WHEN: you accept that new detail will be invented
Make this old photo look as if it was taken recently on a good modern camera: sharp, clean, natural colour, no grain or damage. Keep the same people with the same faces, expressions, ages, poses and clothing, in the same place in the frame, with the same background. Do not beautify anyone or change their age. Keep it natural and true to life rather than glossy or retouched.

Treat this version as an artwork based on the photo, not as how the moment “really looked”.

8. The follow-up line for every later round

OpenAI’s ChatGPT guide recommends adjusting one thing at a time so other details do not drift.1

USE WHEN: asking for one more fix on a result
Now change only this: [remove the spot on the left sleeve]. Keep everything else exactly as it is in the last image: every face, expression, pose, clothing, background, colour and grain. Do not redraw or improve anything else.

If a face has drifted after a few rounds, go back to your original instead of fixing the fix.

TAKEAWAY

Faithful first, one job per round, the same “keep” list every time. Label colourised and “made recent” versions for what they are.

Check it against the original

The plain point: put the original and the result side by side, zoom in, and look for anything new.

  1. Open both at the same size and flick between them.
  2. Count the people. Nobody added, removed or merged.
  3. Check each face: eyes, nose, mouth, ears, teeth, glasses, hairline. Ask someone who knew the person if it still looks like them.
  4. Check hands, jewellery, buttons, and any writing on signs or uniforms. Small details get redrawn.
  5. Check the background: windows, trees, cars and walls should match.
  6. If something important changed, start again from the original with a smaller request, or fix only the damaged area with a selection tool.

The habits that help you spot AI-generated images help here too: look at small physical details, not the overall impression.

TAKEAWAY

If the person who knew them says “that is not quite her”, believe them over the screen.

How to use these prompts in each tool

The plain point: attach the photo, paste the prompt, keep each request to one job.

ToolHow to pasteWorth knowing
ChatGPTHow: attach the photo to the message box and paste the prompt. You can also select one area and describe the change for it.1Worth knowing: repeat anything that must stay fixed in every request.1
Gemini (Nano Banana)How: upload the image from your device, then type the prompt and submit.2Worth knowing: Gemini images carry a hidden SynthID watermark and a visible one.4
Photoshop Generative FillHow: select only the damaged area and generate. The prompt can be left blank.5Worth knowing: the fill goes on its own layer, and your original stays editable underneath.5
Other toolsHow: any tool that takes an uploaded photo and a written instruction can use these prompts.Worth knowing: a “strength” or “denoising” setting limits how far the result moves from the original; upscale prompts explains it.

For a whole-photo clean-up a chat tool is quickest. For one tear, a selection tool touches less of the picture. Midjourney’s help pages blocked our reader on 12 Oct 2026, so this page does not describe its editor.

Sharing a restored photo honestly

The plain point: say it was restored with AI, and keep the original with it.

  • Label it: “restored with AI”, or “colourised with AI; colours are guesses”.
  • Keep the original alongside it, so anyone can compare.
  • Leave watermarks in place. They tell others AI changed the picture. See Content Credentials and AI watermarking.

Questions people ask

Will AI photo restoration change the face?

It can. Chat image tools redraw the whole picture, and OpenAI says repeated edits can still change details you meant to keep.3 Use the faithful prompt, one change per round, and compare against the original.

Are AI-colourised colours accurate?

Not reliably. A black-and-white photo stores only light and dark, so the tool picks colours that seem likely. Put any colours you know in the prompt and label the result as colourised.

Can AI fix a torn or missing part of a photo?

It can fill the gap, but what it puts there is a guess from the surroundings. That works well for a wall or sky. Through a face, check closely, or ask someone who knew the person.

Where to go next

SOURCES
  1. 1OpenAI, ChatGPT Learn, Image generation, undated, read at source 12 Oct 2026“Use likenesses with care.” and “confirm that you have permission to use their likeness.” ↑a“Adjust one element at a time so the composition and other important details do not drift.” ↑b“Attach a reference image to the composer when you want ChatGPT to edit it” and “You can also select a specific area of an image and describe the change for that area.” ↑c“Repeat any requirement that must stay fixed.” ↑d
  2. 2Google, Gemini Apps Help, Generate & edit images with Gemini Apps, undated, read at source 12 Oct 2026“Editing images isn’t available to users under 18.” and “Be sure not to violate others’ copyright or privacy rights.” ↑a“Upload an image and ask Gemini to make edits.” ↑b
  3. 3OpenAI, Image prompting, undated, read at source 12 Oct 2026“Repeated edits can still change details you intended to preserve. Restate those constraints and inspect each result.” ↑a“For edits, say ‘change only X’ and list the details to preserve” and “If a region must remain pixel-identical, composite the approved edit into the original image” ↑b“Repeated edits can still change details you intended to preserve.” ↑c
  4. 4Google, Nano Banana 2 – Gemini AI image generator & photo editor, undated, read at source 12 Oct 2026“an invisible SynthID watermark, as well as a visible watermark to show they are AI-generated.” ↑
  5. 5Adobe, How to use Generative Fill in Photoshop, undated, read at source 12 Oct 2026“Or leave it blank and Photoshop will fill in the selection based on the surroundings.” ↑a“a Generative Layer will be created in your Layers panel for nondestructive editing of your original image.” ↑b

The comparison on this page (the artist copying a photo), the prompts and the checking steps are this site’s own, not claims from these sources. None of the prompts carries a recorded test date.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE