Home › Prompt Library › Prompt to Restore & Colorize Old Photos

Prompt to Restore & Colorize Old Photos

A copy-paste prompt that repairs an old, damaged family photo — scratches, tears, fading — and optionally adds natural color, while keeping every face and detail true to the original.

Images · Works great in Any AI

Copy-ready prompt

Restore the attached old photograph. Work from the original image — do not replace anyone or invent people, objects or background that aren't already there.
Repairs needed:
- Remove [scratches / tears / creases / dust spots / water stains].
- Fix [fading / low contrast / yellowing] and recover detail in [faces / clothing / background].
- Gently sharpen and upscale for a clean, higher-resolution result.
Color: [keep it black-and-white and only repair damage] OR [add historically-plausible, natural color — realistic skin tones, muted period-accurate clothing and surroundings].
Hard rules:
- Do not change anyone's face, expression, age, hairstyle or body — keep every person exactly recognizable.
- Do not add or remove people, objects, jewelry or text.
- Keep it photorealistic and true to the era; no modern styling, no over-smoothing, no plastic skin.
- Preserve the natural film grain and texture where possible.
If any area is too damaged to reconstruct confidently, leave it plausible and understated rather than guessing at specifics. Ask me one question if the photo is unclear before you begin.

Want a version tailored to you?

Answer a few quick questions and the AI Photo & Headshot Generator builds a custom prompt from your exact details.

📷 Open the AI Photo & Headshot Generator

What AI restoration can and can't do

Old family photographs fade, crack, yellow and tear, and for decades the only fix was a skilled retoucher working for hours. AI image editors now do a convincing first pass in seconds — but they carry a specific risk that ordinary photo edits don't. When a model repairs a damaged face, it isn't uncovering the real face underneath; it is guessing what should be there. Guess too freely and you get a photo that looks restored but no longer looks like Grandma. The single most important thing you can do is tell the model to treat the original as the source of truth and to repair rather than reinvent.

Set your expectations honestly. AI is excellent at removing scratches, dust, creases and stains, at rebuilding contrast in a washed-out image, and at upscaling a small print into something you can frame. It is weaker at reconstructing detail that is genuinely gone — a face half-lost to a water stain, or text that has flaked away. In those areas the right instruction is not "reconstruct it perfectly" but "keep it plausible and understated," so the model fills gaps quietly instead of confidently painting in something wrong. The prompt above says exactly that.

Colorizing without lying about the past

Colorization is the most tempting and the most dangerous step. A black-and-white photo contains no color information, so every color the AI adds is an educated guess. The failure mode is over-saturated, modern-looking color — electric blue dresses, orange skin, lawns in a shade of green that didn't exist on that film. Ask instead for historically-plausible, natural color: realistic, slightly muted skin tones and period-accurate, desaturated clothing and surroundings. If the era or setting matters, name it ("1940s, rural, indoors"). And decide deliberately whether to colorize at all — sometimes the honest, moving choice is to repair the damage and leave the image in its original black-and-white. The prompt lets you pick either path.

Keep faces off-limits during colorization too. Adding color shouldn't be a license to smooth skin, brighten eyes, or subtly reshape features. The identity rules in the prompt apply to the colorized version just as much as the repaired one.

Source scans and working in stages

The quality of your restoration is capped by the quality of your scan. Photograph or scan the original at the highest resolution you can, in flat, even light with no glare, and capture the whole print. A sharp, high-resolution scan of a damaged photo gives the model real detail to work with; a blurry phone snapshot taken at an angle forces it to invent, which is where mistakes creep in. Clean the glass or the print gently first so you aren't asking the AI to remove smudges that weren't part of the original damage.

Finally, work in stages rather than demanding everything at once. Ask for damage repair first and check that faces are still faithful. Then, in a follow-up, request upscaling and sharpening. Only then, if you want it, ask for colorization — and review that separately. Staging makes each step easy to judge and easy to reject: if the color goes wrong you still have a clean black-and-white restoration to fall back on. Refine conversationally ("less saturated," "leave that torn corner subtle," "the face changed — keep it exactly as the original") rather than starting over and losing progress.

Why this prompt works

AI restoration tools tend to over-reach: they "improve" a face into a slightly different person, invent buttons and jewelry that were never there, or paint colors that look garish and modern. This prompt draws a hard line between repair (removing damage, recovering faded detail, upscaling) and invention (changing faces, adding objects, guessing at specifics). It locks every person's identity, asks for historically-plausible rather than vivid color, and tells the model to stay understated where the original is too damaged to know the truth — which is exactly what makes a restoration feel faithful rather than fabricated.

How to customize it

Replace the bracketed choices — which damage to fix, and whether to colorize — before sending, and name the era if you know it.
Scan the original flat at the highest resolution you can; sharp detail in means a faithful restoration out.
Work in stages: repair first and check the faces, then upscale, then colorize separately so you can reject any step.

Example output

Sample only

Filled-in prompt (a creased, faded 1950s portrait):

"Restore the attached old photograph. Work from the original — do not replace anyone or invent anything that isn't already there. Remove the vertical crease down the middle, the dust spots, and the yellow staining in the top corner. Fix the heavy fading and recover detail in the faces and the patterned dress. Gently sharpen and upscale. Add historically-plausible, natural color — realistic muted skin tones and period-accurate 1950s clothing colors. Do not change anyone's face, expression or age. Do not add or remove people or objects. Photorealistic, true to the era, keep natural grain, no over-smoothing."

Typical result: the crease and stains are gone, the image has proper contrast again, faces are sharp and clearly the same people, and the color is soft and believable rather than vivid — a photo you could print and frame, with the flaking text in one corner left plausibly understated rather than reinvented.

Prompt variations to try

Repair only — keep it black and white

Restore the attached black-and-white photograph but keep it black-and-white. Only repair damage: remove scratches, creases, dust and stains, fix fading and low contrast, recover detail in faces and clothing, and gently sharpen and upscale. Do not colorize. Do not change any face or add or remove anything. Preserve natural film grain. Stay true to the original.

Colorize a clean black-and-white photo

The attached photo is undamaged black-and-white — do not repair it, only add color. Apply historically-plausible, natural color: realistic slightly-muted skin tones and period-accurate, desaturated clothing and surroundings. Keep every face, expression and detail exactly as it is — colorizing must not reshape, smooth or brighten anyone. Photorealistic and true to the era, not vivid or modern.

Heavily damaged photo — conservative reconstruction

Restore this severely damaged photo. Repair tears, missing corners and large stains, and recover what you can of the faces. Where an area is too damaged to know the truth — especially faces and text — reconstruct it plausibly and understated rather than inventing specific details. Do not replace anyone with a different person. Keep it photorealistic, black-and-white, and true to the original. Flag any area you had to guess at.

Common mistakes to avoid

Letting the model reinvent faces. Restoration should recover the real person, not generate a new one. Always state do not change anyone's face, expression or age so damaged faces are repaired, not replaced.
Over-saturated colorization. Vivid, modern color instantly looks fake. Ask for historically-plausible, natural, muted color and name the era if you know it.
Asking it to invent missing detail. Where the original is destroyed, "reconstruct perfectly" produces confident wrong guesses. Say "keep it plausible and understated" instead.
Feeding it a poor scan. A blurry, angled phone photo forces the AI to guess. Scan the original flat, in even light, at the highest resolution you can before restoring.
Doing everything in one shot. Combining repair, upscale and colorize in a single request makes errors hard to isolate. Work in stages and check faces after each step.

Frequently asked questions

Will AI restoration change what the people actually looked like?

It can, if you let it — repairing a damaged face means the model is guessing at missing pixels. That is why the prompt locks identity: keep the face, expression and age exactly, and tell it to stay understated where detail is truly gone. With those rules, the people stay recognizable.

Is the color historically accurate?

No — a black-and-white photo has no color data, so every color is an educated guess. Ask for period-accurate, muted, plausible color rather than treating it as fact. If accuracy matters (a specific uniform or dress), state the real color yourself so the AI uses it instead of guessing.

Should I colorize at all, or just repair?

Both are valid. Repair-only keeps the photo honest to how it was taken; colorization can make a scene feel present and alive. Because the prompt lets you choose, a good approach is to keep a clean black-and-white restoration and try a separate colorized version, then decide.

How good does my scan need to be?

As good as you can manage — the result is capped by the input. A sharp, high-resolution, evenly-lit scan of the whole print gives the model real detail to restore. A dim, angled, low-resolution photo forces it to invent, which is where restorations go wrong.

Tip: replace the parts in [square brackets] with your own details before you send. The more specific you are — audience, tone, goal, constraints — the better the AI output.