Photo restoration prompts to paste into ChatGPT or Gemini
Copy-paste photo restoration prompts for ChatGPT, Gemini and Nano Banana: repair, colorize, group photos, faded prints and fixes for a changed face.
Compare it with your original
Free during early access · 2 restorations a day
Choose a photoUsed only to restore your photo. How we handle your photo
A photo restoration prompt works when it names only the damage that is really on the print, spells out what must not change, and asks for the picture only. Below are six prompts to paste into ChatGPT, Gemini or Nano Banana, then the rules behind them, drawn from a restoration instruction of ours that went through thirteen versions.
Used only to restore your photo. How we handle your photo
Can ChatGPT or Gemini restore old photos?
Usually yes, for ordinary damage. Give ChatGPT, or Gemini and its Nano Banana image models, a scan and a clear instruction, and scratches, stains, creases and fading typically come back cleaned up.
The catch is how they work: they draw a new picture based on yours instead of patching pixels in place. Every restoration is a redraw, and redraws go wrong first on faces and small details. A prompt cannot switch that off. It decides how much the model may invent.
Our numbers below come from image models we tested in our own lab, not from the ChatGPT or Gemini apps, so read them as direction, not as a guarantee.
Prompt to restore an old black-and-white photo without adding color
This is the base prompt for restoring an old photo in ChatGPT or Gemini; the others are variations of it. Attach your scan and replace each bracket.
Prompt: “Repair only the damage listed here in this old [black-and-white / sepia] photograph, and change nothing else. Damage: [list what you can see, such as a tear from the top edge, brown spots, fine scratches]. Keep its original tones and add no color. Every person stays exactly as they are: face proportions, eyes, expression, apparent age, hair, clothes and pose; the background stays too. Do not smooth skin, make anyone younger or prettier, or add people, objects or text. Where the photo is unclear, leave it soft instead of inventing detail. Keep the framing, and return the picture only, with no border.”
How to write a photo restoration prompt, in six parts
Open with the job, and keep the top short
The first sentences carry the most weight, and whatever you add above them pushes the rest down. In one test, a “repair, do not redraw” line added at the very top displaced the list of what must stay, and the face-similarity score on a portrait fell from 0.815 to 0.620 (1.0 is an identical face).
List only the damage that is really there
Name each flaw plainly: scratches, dust specks, a crease, a tear and where it runs, brown spots. Never call the whole photo damaged. On a clean photo, an early instruction of ours that did so made the model draw a tear that was not there, in 2 of 2 runs. Naming damage only where it exists gave none in 3 runs.
List what must stay
Spell out the face proportions, the eyes and where they look, expression, apparent age, hair, clothes, pose and background. Long and concrete beats short and polite: a six-line prompt of ours matched the face less closely than a long, specific one.
List what must not appear
Forbid the extras by name: smoothing, a younger or prettier person, new people or text, any border. Leave out “modern” and “DSLR quality”. Asking for the look of a new camera went wrong another way: pose, background and face moved, and the similarity score was 0.595 against 0.761 for repair-first wording.
Say what to do where the photo is unclear
Ask for unclear areas to stay soft instead of invented. Fit the skin instruction to the print: on a badly cracked face, “keep pores and blemishes” made the model read cracks as skin and scatter dark specks across the cheeks. “Clean, even skin where the surface is destroyed” fixed it.
Change one thing at a time, and run it three times
The same prompt returns a different picture on each run, so we never trust fewer than three. Change one line between attempts, judge by eye, and keep your original scan untouched.
Prompt to restore and colorize an old photo
State the color request in the first sentence and repeat it in the last. In our tests a sepia portrait came back monochrome in 3 of 6 runs when the request sat mid-prompt. Moved to the front and repeated at the end, it failed in none of 5 runs. Five runs is a small sample.
Prompt: “Colorize this old black-and-white photograph and repair only the damage listed here, while every person stays exactly who they are. Damage: [list what you can see]. Give each person lifelike skin that differs from person to person, never the same orange or pink on everyone. Read the colors of clothing and surroundings from the tones and light in the photo, keep them slightly muted, and if the photo gives no clue to a color, use a plain, common one. Do not change any face, pose or background, do not smooth skin, and do not invent detail. Return the picture only, with no border. The final result must be in full color, not black and white or sepia.”
If the print is clean, delete the damage sentence.
Prompt for a phone photo of a print or an album page
A phone photo needs a prompt that also asks for the crop and the straightening; the model has no reason to remove the table, fingers or album page unless you say so.
Prompt: “This is a phone photo of a printed photograph lying on [a table / an album page]. Return only the photograph itself, straightened and filling the frame edge to edge, with the table, fingers, album page and paper border cut away. Do not cut into the picture: everyone and everything inside it stays. Remove glare, and repair only the damage listed here: [list what you can see]. Every person keeps their face proportions, eyes, expression, apparent age and pose. No frame, margin or shadow around the result.”
Shoot straight-on in soft daylight, without flash, and check all four edges of the result: removing a border can trim real picture too.
Prompt for a group photo with many faces
Group photos are where prompts help least. In our tests, three unrelated instructions and two reference strategies all failed the same way: with more than one face in the frame, likeness did not hold.
Prompt: “Repair only the damage listed here in this old group photograph, and change nothing else. Damage: [list what you can see]. The number of people stays the same, and each person stays in the same place, size and pose. Do not change any face. Where a face is too small or too damaged to show detail, leave it soft and neutral instead of drawing new features. Do not beautify anyone or add people, objects or text. Return the picture only, with no border.”
This lowers the risk and does not remove it. Check every face against the original, small ones first. If one is wrong, crop that row from the original scan and restore it alone.
Prompt for a faded color print
For a faded color print, ask for the cast to be corrected and nothing recolored. Dyes fade at different speeds and leave the whole picture tinted, so name the tint you see.
Prompt: “This is a faded color print with a [red / orange / yellow / magenta] tint. Correct the tint and the lost contrast so it looks like a fresh print, and repair only the damage listed here: [list what you can see]. Do not recolor anything: clothing, walls and sky keep their hues, and only the fade is undone. Skin looks natural and differs from person to person. Do not change any face, pose or background, and do not smooth skin. Return the picture only, with no border.”
If your print’s colors are healthy, leave the color lines out: a correction nobody needs invites a change.
Prompt to restore old photos without changing the face: what helps
No prompt can promise it. Our guide to faces changed by AI restoration explains why: the model redraws the whole picture and fills thin evidence with a plausible face. The rules above lower the risk, and these follow-ups help when a face still changes.
Follow-up prompts for a face that changed
The face is no longer the person
“The face in your result is not the face in my original. Redo the picture with the original face: same face shape, same eyes and spacing between them, same nose, mouth, ears, hairline, age and expression. Change nothing else.”
Smoothed skin, a younger look
“The person looks smoother and younger than in my original. Where the original shows skin, match it: its lines, its uneven tone, its age. Change nothing else.”
One face in a group changed
“Only the [man in the back row, second from the left] looks different from my original. Redo that face from the original and leave everyone else exactly as they are.”
A single portrait that keeps drifting
Attach a tight crop of the face from your scan as a second image and add: “Image 2 is a close-up of the same person. Use it only as a reference. Do not paste it into the result.” In our lab this made single-portrait results far more consistent. On group photos it made them worse, so skip it there.
Two follow-ups did not help
Start a new chat with the original and the full prompt. Every pass is another redraw, so changes can pile up.
How to check a restored face against the original
- Put both versions at the same size and flick between them, at 100% zoom.
- Eyes: the gap between them, their shape and where they look.
- Mouth: a closed mouth stays closed, with the same corners.
- Hairline and ears: where the hair starts at the temples, and the shape of each ear.
- Age and skin: nobody looks younger or smoother than the original shows.
- In a group photo, repeat all of this for every face, small ones first.
What no prompt can fix
- A guaranteed likeness. Wording lowers the risk; only a side-by-side check settles it.
- A sharper face where there are no pixels. A face a few pixels wide comes back cleaner, not sharper.
- The true colors of a black-and-white photo. Color is a plausible reading, so keep the original.
- A file as large as your scan. A chat app’s picture often has fewer pixels than a good scan, so keep the scan.
- A missing piece, truthfully. What fills a hole in a face or a torn corner is invented.
When a prompt is enough, and when a dedicated tool helps
A prompt is enough for a few photos, when you like steering the result and will check every face yourself. It is also the better route to a result that stays black and white, which a dedicated tool may not offer.
A dedicated tool helps with a box of photos, or when you would rather not write and tune instructions. Epoha runs an analysis pass first, builds the instruction for that photo, and lets you hold the photograph to see your original so you can compare.
A chat-app prompt and Epoha, side by side
| A prompt in a chat app | Epoha | |
|---|---|---|
| What happens first | You read the print and list the damage | An analysis pass finds faces and damage, and warns about faces that are too damaged or too small |
| Black-and-white result | Yes, if you ask for it | No: black-and-white and sepia always come back in color |
| Phone photo of a print | Ask for the crop in the prompt | Margins, table and fingers removed; tilt straightened |
| Steering the result | Unlimited follow-ups | One pass; upload again for a new variant |
| Checking the face | You, by eye | You, by eye, holding the photo to see the original; no automatic check |
Questions about restoring old photos with ChatGPT and Gemini
How do I restore old photos with ChatGPT or Gemini?
Attach your best scan, paste the prompt that matches your photo and fill the bracket. Compare the result with the original, change one thing at a time, and keep the best of three runs.
Why does ChatGPT change my face when editing photos?
It redraws the whole picture, and a face is where a small error shows most. With little to go on, such as damage across the face or a tiny face, it invents a plausible one instead of yours.
What is the best prompt to restore old photos?
The one written for your photo: it names only the damage that is there, lists what must stay and forbids the extras. Start from the base prompt above.
Can I restore and colorize an old photo in one prompt?
Yes, with the color prompt above. Or repair first and colorize second, which leaves a repaired black-and-white copy at the cost of one more redraw.
Should I add “4K” or “DSLR quality” to the prompt?
Better not. Those words ask for a new photograph rather than a repaired one, and no wording adds detail the print never held.
Compare it with your original
Upload your photo, then hold the restored photo to see your original and check the faces one by one before you decide anything.
Free during early access · 2 restorations a day
Used only to restore your photo. How we handle your photo