30 ChatGPT Old Photo Restoration Prompts: Repair, Colorize, Upscale Without Redrawing Faces (2026)

ChatGPT old photo restoration prompts are instructions attached to a scanned print that tell ChatGPT Images 2.6 to repair damage, correct fading, add color, or enlarge the image while recovering only detail that exists in the scan. Restoration goes wrong when the prompt leaves the model room to guess, and it guesses faces first. The 30 prompts below work in passes, damage, then color, then upscale, and every one forbids redrawing a face.

What Makes ChatGPT Old Photo Restoration Prompts Work

A restoration prompt asks the model to remove what time added and leave what the camera recorded. Without a selection, ChatGPT regenerates the whole frame from your scan plus the instruction, so "fix the scratches" can also tidy a grandmother's jawline or draw a face where a crease used to be. Every prompt below is written to stop that.

Recover only detail that exists in the scan. The model cannot read what the emulsion lost; it can only draw what it thinks was there. That line, repeated in every prompt, keeps the fill honest: a scratch is filled from the pixels around it, a soft face stays soft, nothing is invented.

Never reconstruct a missing face or person. Where a crease crosses a face, the prompt fills the line and leaves the face as scanned. Where a tear removes part of a person, the prompt leaves plain background. A reconstructed face is a stranger wearing a relative's clothes.

Work in passes: damage, then color, then upscale. Damage first, because fills need the original tones to match. Color second, because colorizing a scratched print bakes the scratch into the color. Upscale last, because it touches every pixel and should run once on a finished image. ChatGPT Images 2.6 is about 50% faster than the previous version, so three short passes cost little time.

Name the faces so they stay on the record. "Keep every face" is too loose to hold. Name what is visible: face shape, the set of the eyes, nose width, the line of the mouth, hair and hairline, expression, pose, framing, and proportions. For a black-and-white print add "tonal values," so skin is neither lightened nor darkened. The model preserves what it is told to notice.

Select the damage, describe the fade, and forbid the quiet edits. Open the scan, choose the Select tool, and highlight the scratch or stain so only that region is regenerated. OpenAI notes that highlights are not always precise and an edit can extend past the selection, so the keep list stays in the prompt even when you select. Fading, cast, and upscales touch the whole frame, so describe those and rely on the keep list. The "do not" lines, no skin smoothing, no sharpening, no invented detail, are negative prompting applied to restoration.

This post is a spoke of 40 ChatGPT photo editing prompts, which covers restoration in five prompts alongside backgrounds, relighting, and restyles; the 30 below go deeper on scans alone. If none fits your print, the AI prompt generator writes a restoration prompt with the keep list and pass order built in from a plain-English description of the damage.

Every prompt below assumes one workflow. How to restore an old photo in ChatGPT, in six steps:

Damage Repair (1–6)

Physical damage is where the Select tool earns its place. Highlight the scratch, crease, or stain, ask for a fill from the surrounding tones and grain, and keep every face outside the selection. These ChatGPT old photo restoration prompts repair the surface of the print and nothing underneath it; color and sharpening come later in their own passes.

1. Remove Scratches and Dust

Remove the fine scratches and dust spots across the [sky and the
wall behind the group] of this scanned print, filling each mark
from the surrounding tones, grain, and paper texture so no patch
shows. Recover only detail that exists in the scan. Keep every
face shape, features, tonal values, hair, expression, pose,
framing, proportions, and grain as scanned. Do not sharpen,
smooth skin, or colorize.

What you'll get: A clean surface with every face carried over as scanned.

Variation: Highlight the dusty sky with the Select tool so the faces are never regenerated.

2. Fill Creases and Fold Lines

Remove the [two horizontal fold lines across the middle] of this
print, filling each line from the tones and grain on either side
so the [wall, floor, and clothing] continue without a seam. Where
a line meets hair or a shoulder, fill the line only. Recover only
detail that exists in the scan. Keep every face, expression, pose,
framing, proportions, and grain as they are. No blur patch, no
sharpening.

What you'll get: Fold lines gone, with fabric and background continued instead of smeared.

Variation: For a print folded into quarters, name all four lines and where they cross.

3. Rebuild a Torn Corner

The [upper left] corner of this print is torn away. Rebuild the
missing area from the surrounding [plain wall and curtain edge],
matching the print's perspective, tone, grain, and edge fade.
Recover only detail that exists in the scan. Keep every face,
pose, object, clothing, and the original framing as they are. If
the tear reaches a person, do not reconstruct any part of them;
leave plain background there. No sharpening.

What you'll get: A rebuilt corner that continues the room and stops at anyone it touches.

Variation: If the corner is gone entirely, ask for "a plain border in the paper's tone."

4. Lift a Water Stain

Remove the [brown water stain with a hard tide line across the
lower third] of this print. Restore the tones under the stain to
match the undamaged areas beside it, following the gradient of
the [floor and skirt], and dissolve the tide line without leaving
a lighter band. Recover only detail that exists in the scan. Keep
every face, expression, pose, framing, proportions, and grain as
they are. Do not sharpen, smooth skin, or colorize.

What you'll get: The stain gone and the tide line dissolved, with no lighter band left.

Variation: For patterned clothing, add "continue the [stripe] pattern at the same spacing."

5. Clear Mold and Foxing Spots

This print has [rust-colored foxing spots and gray mold speckles]
across the [background and lower half]. Remove each spot by
filling it from the tones and grain around it. For spots on
faces, fill the spot only; do not alter the skin, features, or
tonal values around it. Recover only detail that exists in the
scan. Keep every face, hair, expression, pose, clothing, framing,
and grain as they are. No smoothing, sharpening, or color shift.

What you'll get: A spot-free print with the skin around each spot untouched.

Variation: Run background spots and clothing spots as separate passes.

6. Fill a Crease That Crosses a Face

A [diagonal crease runs from the top right edge across the
forehead and left cheek of the woman on the right]. Fill the
crease line only, borrowing tone and grain from the skin on
either side. Do not redraw, sharpen, smooth, or reshape any part
of her face: face shape, eyes, brows, nose, mouth, tonal values,
hairline, and expression stay as the scan shows them, soft where
it is soft. Recover only detail that exists in the scan. Keep
everyone else as they are.

What you'll get: The line gone and the face it crossed unchanged, softness included.

Variation: Trace the crease with the Select tool at its exact width so nothing else is regenerated.

Fading and Cast (7–11)

Fading, yellowing, and lost contrast touch every pixel, so these are described rather than selected and lean on the keep list. Say what the print should look like afterward in concrete terms; "restore the colors" invites a modern, saturated look no print of that era carried. Run these on the repaired result, never on the raw scan.

7. Revive a Faded Print

This color print has faded to a flat, low-contrast look with weak
blacks. Restore the density: deeper blacks, clean whites, and
color as it looked when the print was new, with realistic skin
and the saturation of a print of its era. Recover only detail
that exists in the scan. Keep every face shape, features, hair,
expression, pose, framing, and grain as they are. Do not push
saturation, sharpen, smooth skin, or crop.

What you'll get: The print as it looked the year it was made, no filter on top.

Variation: Give known colors, "the car was pale green," to anchor the correction.

8. Neutralize a Yellow Cast

This print has an overall [yellow / orange] cast from age. Remove
the cast so whites read as white and grays as neutral, using the
[white shirt collar / white border] as the reference, and return
each person's skin to a natural tone without lightening or
darkening it. Keep a faint warmth. Recover only detail that
exists in the scan. Keep every face, hair, expression, pose,
framing, and grain. No smoothing.

What you'll get: Neutral whites and natural skin, the print's age still in the grain.

Variation: For a magenta or cyan shift, name that color instead.

9. Restore Contrast Without Crushing Detail

This black-and-white print has gone gray and flat. Restore a full
tonal range: true blacks in the [coat and shadows], clean whites
in the [collar and sky], smooth midtones in the faces, with no
blocked shadows or clipped highlights. Recover only detail that
exists in the scan. Keep every face shape, features, hair,
expression, pose, framing, grain, and the tonal relationship
between skin and clothing. No sharpening or smoothing.

What you'll get: Deep blacks and open highlights, faces unchanged.

Variation: Upload a good print from the same set and add "match its contrast."

10. Keep the Sepia on Purpose

Restore this sepia-toned print and keep the sepia. Correct only
the fading: bring back the density of the dark browns and the
brightness of the cream highlights so the full range returns. Do
not convert to neutral black and white or colorize. Keep the warm
tone, every face, features, hair, expression, pose, framing, edge
vignetting, and grain as they are. Recover only detail that
exists in the scan. No sharpening, no smoothing.

What you'll get: The same sepia print with its range back.

Variation: If the card mount is in the scan, add "keep the mount and its embossing."

11. Recover a Blown-Out Scan

This scan is overexposed: the [faces and the white dress] are
nearly white with little detail. Bring the exposure down by about
[1] stop and recover tone in the bright areas only where the scan
still holds it, leaving pure white where it does not. Recover
only detail that exists in the scan. Keep every face shape,
features, hair, expression, pose, framing, and grain. Do not draw
in facial detail or sharpen.

What you'll get: Whatever the scan still held in the highlights, honest white elsewhere.

Variation: Rescan at a lower brightness first; no prompt recovers what the scanner never read.

Colorize (12–17)

Colorization is the most requested restoration and the most open to guesswork, because the scan records brightness and the model supplies the hue. These ChatGPT old photo restoration prompts narrow the guess: supply every color you know, ask for the muted palette of a print of that era, and keep the grain. Label the result as colorized wherever it travels.

12. Colorize a 1940s Family Photo

Colorize this black-and-white family photo from the 1940s with
realistic, period-plausible color: natural skin tones, [a dark
wool suit, a floral cotton dress, and a knitted cardigan], muted
greens outside, and the soft saturation of a print of that decade.
The light is [overcast daylight from the front]; keep the shading.
Recover only detail that exists in the scan. Keep every face
shape, features, hair, expression, pose, framing, and grain as
they are. Do not sharpen, repair damage, or smooth skin.

What you'll get: Believable period color over the original grain with nothing else touched.

Variation: Add "[NAME] had red hair and blue eyes" for anyone whose coloring you know.

13. Colorize a Wedding Portrait

Colorize this black-and-white wedding portrait. The dress is
[ivory satin], the suit is [charcoal], the bouquet is [white roses
with green foliage], and the backdrop is [a painted studio canvas
in soft gray-blue]. Give both faces natural skin tones under the
soft studio light from the [upper left], with no blush or makeup
added. Recover only detail that exists in the scan. Keep both
faces, hair, expressions, poses, framing, and grain. No smoothing.

What you'll get: A colorized wedding portrait in the colors you supplied, both faces unchanged.

Variation: If the dress survives, upload a photo of it and say "match the dress to this color."

14. Colorize a Uniform With the Insignia Kept Exactly

Colorize this black-and-white portrait of a person in uniform:
[UNIFORM COLOR] tunic, [TROUSER COLOR] trousers, [COLOR] cap
band. Keep every badge, button, ribbon, stripe, and insignia
exactly as shaped and positioned in the scan: do not redraw,
simplify, enlarge, add, or remove any of them, and do not guess
their colors unless I have given them. Natural skin tone. Recover
only detail that exists in the scan. Keep face shape, features,
hair, expression, pose, framing, and grain. No sharpening.

What you'll get: A colorized uniform with every insignia in its original shape and place.

Variation: When the ribbon colors are unknown, add "leave all badges and ribbons in their original gray."

15. Colorize a Street Scene

Colorize this black-and-white street photo with realistic,
period-plausible color: [warm red-brown brick, a gray stone curb,
dark painted shopfronts, and an overcast sky]. Give the people
natural skin tones and muted clothing colors, and keep any visible
signs and lettering exactly as written. The light is [flat
overcast daylight]. Recover only detail that exists in the scan.
Keep every person, building, vehicle, framing, and grain. No new
objects, no modern colors.

What you'll get: A colorized street with every sign still spelled as scanned.

Variation: Name the city and decade so vehicle and paint colors fit the era.

16. Colorize a Child Portrait

Colorize this black-and-white portrait of a child: natural skin
tone, [HAIR COLOR] hair, [EYE COLOR] eyes, [a pale blue romper
with a white collar], under [soft window light from the left].
Keep the softness of the original: do not sharpen the face, add
eyelashes or detail, smooth skin, or make the child look older or
younger. Recover only detail that exists in the scan. Keep face
shape, features, hair, expression, pose, the background, framing,
and grain as they are. Muted print palette.

What you'll get: A gently colorized child portrait with the original softness kept.

Variation: If the eye color is unknown, add "leave the eyes a neutral gray-brown."

17. Colorize a Group With Known Colors Supplied

Colorize this black-and-white group photo of [NUMBER] people with
these known colors: [the man on the far left wore a brown suit;
the woman second from left had a green dress; the child in front
wore a red jumper; the car was black]. For everything I have not
named, choose muted period-plausible colors. Natural skin tones,
nobody made lighter or darker. Recover only detail that exists in
the scan. Keep every face, hair, expression, pose, position,
framing, and grain. Do not redraw or swap anyone. No sharpening.

What you'll get: A colorized group with your known colors right and the rest plausible.

Variation: Correct a miss with "change only the [dress] to [navy]; everything else stays."

Sharpen and Upscale (18–22)

Upscaling touches every pixel, so the Select tool cannot protect a face here and the keep list does all the work. Ask for up to 2K, the output ceiling, soft where the scan is soft, and run it last on the repaired and corrected result. For phone photos rather than scans, ChatGPT photo enhance prompts covers blur, noise, and compression.

18. Enlarge a Wallet Print to 2K

This is a scan of a small wallet-size print, about [2 by 3
inches]. Enlarge it to up to 2K, same aspect ratio, as a faithful
enlargement: clean edges on clothing, hair, and background,
recovering only detail that exists in the scan. Leave the face
soft where the print is soft; do not invent eyelashes, pores, or
texture. Keep face shape, features, tonal values, hair,
expression, pose, framing, and grain. No halos.

What you'll get: A larger file that holds at 5 by 7 inches without a drawn-in face.

Variation: For a bigger print, take the 2K result to a dedicated upscaler.

19. Sharpen a Soft-Focus Portrait

This studio portrait was shot in soft focus and scanned slightly
blurred. Bring the hair edges, the [lace collar], and the backdrop
texture back to clean definition where the scan shows them, and
leave the face at the softness the photographer chose. Do not
sharpen, redraw, or add detail inside the face. Recover only
detail that exists in the scan. Keep face shape, features, hair,
expression, pose, framing, tonal values, and grain. Up to 2K.

What you'll get: Crisp edges around a face left as photographed.

Variation: For a print, allow "a slight overall lift, the face no sharper than the collar."

20. Rebuild a Newspaper Halftone

This is a scan of a newspaper halftone, made of dots with coarse
midtones. Convert it to smooth continuous tone by blending the
dot pattern into the tones it represents, without adding detail
the dots do not carry. Faces will stay soft and low in detail; do
not draw features in. Keep the composition, every face shape,
hair, expression, pose, framing, and any caption lettering.
Recover only detail that exists in the scan. Up to 2K.

What you'll get: Smooth tone at the level of detail the halftone carried.

Variation: Add "keep the printed caption below the photo, spelled exactly" to include it.

21. Clean a Slide or Negative Scan

This is a scan of a [35mm color slide / color negative, already
inverted to a positive]. Remove the dust specks and fine
scratches, neutralize the [magenta] cast from age, and restore
natural density without pushing saturation. Keep the film grain
rather than a clean digital look. Recover only detail that exists
in the scan. Keep every face shape, features, hair, expression,
pose, framing, and all other colors. Up to 2K. No sharpening.

What you'll get: A clean, neutral frame with its grain intact and nothing drawn in.

Variation: For an uninverted negative, add "invert to a positive first, then correct."

22. Restore a Passport-Era ID Photo

This is a scan of a small [ID / passport-style] photo, about [35
by 45 mm], on a [plain light] background. Enlarge it to up to 2K.
Clean the dust and the [corner of an ink stamp at the lower right]
by filling from the background, and leave the face exactly as the
scan shows it: same face shape, eyes, nose, mouth, tonal values,
hairline, and expression, soft where the scan is soft. Recover
only detail that exists in the scan. Do not sharpen the face or
smooth skin.

What you'll get: A clean enlargement of a tiny photo, same face.

Variation: If a stamp crosses the face, fill the stamp line only or keep it as a record.

Group Photos (23–26)

Group photos multiply every risk: more faces to drift, more chances the model swaps or blends two people, and usually someone half-cropped at an edge. Count the faces in the result against the scan before you accept anything. These ChatGPT old photo restoration prompts apply equal restraint to every face, and no face is ever completed from a guess.

23. Restore a Group With Equal Restraint on Every Face

Restore this scanned group photo of [NUMBER] people: remove the
scratches, dust, and the [crease across the lower half], and
correct the fading. Treat every face with the same restraint: keep
each person's face shape, eyes, nose, mouth, tonal values, hair,
glasses, expression, pose, and position as they are, and do not
swap, blend, redraw, or sharpen anyone. Recover only detail that
exists in the scan. Keep framing, clothing, and grain. No smoothing.

What you'll get: A cleaner group where everyone is still exactly themselves.

Variation: If one face drifts, highlight that person's clothing and background and rerun the fill there.

24. Leave a Half-Cropped Person as Plain Background

At the [right] edge of this print a person is cut off by the
frame, and the [tear / damage] removes part of them. Do not
reconstruct the missing part of that person or complete their face
or body. Where the damage reaches them, fill with plain background
in the print's tone and grain, and leave the visible part of them
as the scan shows it. Recover only detail that exists in the scan.
Keep everyone else, the framing, and the grain. No sharpening.

What you'll get: The edge repaired without a guessed face or shoulder.

Variation: To lose the partial person instead, say "crop to [4:5]; do not regenerate content."

25. Fix One Dark Face Without Touching the Others

In this group photo the [person second from the left] is in
shadow and their face is much darker than the others. Lift the
exposure on that person's face and shoulders by about [1] stop to
match the rest, leaving it soft where the shadow hid detail. Do
not change their face shape, features, hair, or expression; alter
their tonal values only by the exposure lift. Recover only detail
that exists in the scan. Leave every other person, the
background, framing, and grain as they are. No smoothing.

What you'll get: One face matched to the group, everyone else untouched.

Variation: Highlight only that person with the Select tool so the rest is never regenerated.

26. Re-Level a Tilted Group Photo

This group print was scanned at an angle and the [floor line /
wall edge] runs downhill to the right. Rotate the image until it
is level, then crop to the largest clean rectangle that keeps
every person fully in frame. Do not stretch, warp, or regenerate
content to fill the corners; crop them away. Keep every face,
hair, expression, pose, proportions, tonal values, and grain.
Recover only detail that exists in the scan. No sharpening.

What you'll get: A level frame with every person intact and no invented corners.

Variation: Rescan the print squared to the glass first; rotation always costs some edge.

Print and Share (27–30)

The last pass prepares the restored file for the wall, the album, or the family group chat. These prompts add a border, a layout, or a caption and change nothing inside the photo. Give sizes in inches and the aspect ratio, ask for up to 2K, and carry the label: a colorized or repaired copy is marked as one.

27. Print-Ready 5 by 7 With a Border

Prepare this restored photo for a 5 by 7 inch print. Place the
image on a [white] canvas at a 5:7 ratio with an even border of
about [half an inch] on all sides, scaling the photo to fit
without cropping or stretching it. Keep the photo itself exactly
as it is. Output at up to 2K. No sharpening, no vignette, no
text, no rounded corners.

What you'll get: A print file with a clean border and the photo untouched.

Variation: Swap the border to "warm off-white (#F5F0E6)" for a sepia or black-and-white print.

28. Before-and-After Side by Side

Create a before-and-after image. Place the original scan I
uploaded first on the left and the restored version I uploaded
second on the right, both at the same size and aligned, with a
thin [white] gutter between them on a [white] canvas at a 2:1
ratio. Do not alter either photo in any way. Add the small labels
"Original" and "Restored" below each, spelled exactly, in a plain
sans-serif. Up to 2K.

What you'll get: A clean comparison that shows the work without editing either side.

Variation: Change the right-hand label to "Colorized" when that is what the copy is.

29. Photo-Book Page With a Readable Caption

Lay out this restored photo on a [square / 8 by 10 inch] photo
book page with a [cream] background: the photo centered in the
upper two-thirds, and below it the caption "[NAMES, PLACE, YEAR]"
in a plain serif, dark gray, spelled exactly as written. Add a
smaller line reading "Colorized from a black-and-white print" if
this copy is colorized. Keep the photo itself unchanged. Up to
2K. No decorations.

What you'll get: A page with the photo and a caption that identifies and labels it.

Variation: Ask for "two photos stacked with one caption each" for a before-and-after spread.

30. Framed Memorial Print, Done Gently

Prepare this restored portrait as a memorial print. Keep the
photo exactly as restored, with no further sharpening, smoothing,
or retouching. Place it on a [soft white] canvas at a [4:5] ratio
with a generous even border, and below the photo add "[FULL
NAME]" with "[YEARS]" beneath it in a quiet serif, dark gray,
spelled exactly. No glow, vignette, flowers, or soft-focus effect.
Up to 2K.

What you'll get: A quiet, print-ready memorial of the person as photographed.

Variation: Leave out the text when the frame will carry the name.

Tips for Better ChatGPT Old Photo Restoration Prompt Results

Scan once, scan well, and keep the original. Scan flat to the glass at the highest resolution your scanner offers, with every auto-enhance and dust-removal option off, and save that file where nothing overwrites it. The model can only keep what the scan holds. The scan is the record; every restoration is a copy.

Download each pass, and fix small misses with a comment. Later edits build on the previous result, so save the damage pass before the color pass and the color pass before the upscale; if a face drifts in pass two, restart from the pass-one file. When a result is close, comment on the image to point at the one spot that is off instead of pasting a new prompt; ChatGPT Images 2.6 keeps the rest consistent across rounds.

Say "soft where the scan is soft." A scanned face from a small print has no pores or eyelashes in it. If the prompt allows sharpening, the model draws them in and the face becomes someone else's. A slightly soft grandmother who is your grandmother beats a sharp one who is not.

Do not use a saved reference image to fill a lost face. ChatGPT Images 2.6 can save a reference image of a person and reuse it in later chats, which is useful for generating new pictures of someone and wrong for restoration: a face rebuilt from a different photo is a composite, and the album will not say so. Where the scan has lost a face, fill the damage line and leave the face as it is.

Pick the right tool for a strict fill. Gemini's Nano Banana is stronger at local edits where nothing else may move, such as a single crease fill on an otherwise perfect print; the 50 best Nano Banana prompts collection covers that side. ChatGPT is stronger for the fade, cast, and colorize passes, and for any layout with a caption.

Generate Custom ChatGPT Old Photo Restoration Prompts

Your print has its own damage, its own faces, and its own colors that someone still remembers. Describe them in plain English to the image prompt generator and it writes the full restoration prompt, keep list and pass order included; the Template Builder saves a family's keep list so you can paste it into every print from the same album.

For every other kind of edit, the hub post 40 ChatGPT photo editing prompts covers backgrounds, relighting, restyles, and wardrobe. If you came looking for restoration "codes," ChatGPT photo editing commands explains why there are none and what to type instead, and ChatGPT couple photo prompts picks up where the wedding colorization leaves off, with then-and-now layouts for anniversaries.

Scan high, keep the original, work in passes, and never let the model draw a face.