How to Restore Old Photos with AI Without Changing the Faces

Photos that are several decades old usually have the same problems: scratches, creases, water stains, flaked-off patches and yellowing, and many are black and white. A photo studio would charge for a repair and take days. Fixing it by hand in Photoshop would take hours.

Take this old photo, for example: two boys standing in front of some bushes, with scratches, water stains and flaked-off patches over most of the picture. We tried image-to-image AI on it: upload the photo, paste one prompt, generate once. This is what came back (left: before, right: after).

The scratches and stains are gone and the black-and-white photo is now in color. Both faces and expressions are the same, and period details like the stand-collar jacket, the cloth shoes and the bare feet are still there.

How to do it

1. Open an AI image tool that supports image to image (we used the CVY.AI image generator). Sign in, switch to Image to Image, pick GPT Image 2 and upload the old photo.
2. Paste the prompt below and edit it for your photo.
3. Put the result next to the original. If something is off, change the prompt and try again. For example, if a face changed, make the "keep the facial features" line stronger.

The prompt

Restore this old photograph with the following requirements:
- Remove scratches, creases and stains.
- Reduce noise and blur, and improve clarity.
- Correct yellowing and color casts. If it is black and white,
  colorize it naturally with skin tones, clothing and scene colors
  that fit the period. Avoid oversaturation.
- Keep the person's facial features, expression and identity
  exactly the same. Only repair the photo: no beautifying, no face
  slimming, no change to how the person looks.
- Keep the period and setting: keep the existing clothing,
  hairstyles, furniture, background, signs, vehicles and
  recognizable objects. Fill missing areas so they fit the original
  period, place and everyday setting.
- Do not add anything that does not fit that period, region,
  identity or setting.
- For unclear hands, objects, backgrounds and missing areas, repair
  conservatively based only on what is visible. If you cannot tell
  what something is, leave it naturally soft or low in detail
  instead of inventing a specific object.
- Keep the original composition and aspect ratio.

Why the prompt is written this way

If you only write "restore this old photo", the usual result is a sharper picture of someone who no longer looks like the person in it. Two things cause that:

• AI beautifies faces by default: smoother skin, a slimmer face, bigger eyes. Fine for a selfie, wrong for an old photo. So the prompt lists each of these one by one: no beautifying, no face slimming, no change to how the person looks. "Keep their identity" on its own is too vague.
• AI fills unclear areas with modern-looking things: blurry hands, clothing and objects in a decades-old photo can come back looking modern. So the prompt says to keep what is already there and not add anything from another period.

The last rule, "if you cannot tell, leave it soft", is the one people forget. Without it, the AI has to invent something whenever it is unsure. With it, it is allowed to leave that area blurry.

Add one line for your kind of damage

• Colorizing black and white: ask for natural, period-appropriate colors and realistic skin tones, not oversaturated. Otherwise colors can look plastic or neon.
• Heavy scratches, creases or missing corners: ask it to remove the damage and fill the missing areas. Whatever it fills in is a guess (see below).
• Very blurry or low resolution: ask for more clarity, and add "do not invent facial details". The blurrier the face, the more likely it turns into someone else.
• A phone photo of a print, with glare or a color cast: ask it to correct the color and remove glare and distortion.

A few things we learned

• If the photo has a lot of problems, do it in two passes: first remove the damage and sharpen it, then colorize it separately.
• The same prompt gives a slightly different result each time, so generate a few and pick the best.
• Anything that was blurred or missing in the original is the AI's guess, not what was really there. In the example above, the bushes in the background are much sharper than in the original, and those details simply were not in the photo. A restored photo is great for an album, a print or sending to family, but it is not a record of how things actually looked.

To try it on your own family photos, paste the prompt above into image to image on CVY.AI.