Most families have a shoebox somewhere: small black-and-white prints with soft corners, a few creases, and names written in pencil on the back. AI colorization tools have made it easy to give these photos a second life. Getting a result that looks natural, though, depends far more on what you do before and after you press the button than on the tool itself.

This is the workflow I use, from digitizing the print to judging whether the colors are believable.

Step 1: Pick the right photos first

Not every photo is a good candidate. Start with the ones most likely to work well:

  • Portraits and small group shots. Skin, hair and eyes are what colorization models handle best. Photos from roughly the 1920s to the 1970s tend to come out natural.
  • Outdoor scenes. Sky, grass, water and stone have consistent colors, so landscapes and buildings usually look convincing.
  • Sharp, reasonably exposed prints. A model can only add color to detail that exists.

Leave for later:

  • Tiny contact prints or heavily compressed images found online. There's little information to work from.
  • Large group photos where every face is a few pixels wide, unless you can scan them at high resolution.
  • Photos where exact colors matter (a uniform, a flag, a family car) and you don't have a source for the real colors. AI will guess, and guesses can be wrong.

Step 2: Digitize the print properly

Input quality changes the output more than any setting.

If you have a flatbed scanner:

  1. Clean the glass and gently dust the print with a soft brush.
  2. Scan at 600 DPI. For very small prints (wallet size), go higher.
  3. Save as PNG or high-quality JPG. Avoid repeated re-saving as JPG.

If you only have a phone:

  1. Lay the print flat next to a window in even daylight.
  2. Turn off the flash. Flash causes glare spots that a model may treat as part of the image.
  3. Hold the phone parallel to the photo to avoid keystone distortion.
  4. Shoot several frames and keep the sharpest.

Then, for both:

  • Crop away the table, album page and border.
  • Straighten the image so edges are square.
  • Keep the untouched scan as your archival master. Always work on a copy.

Step 3: Decide: colorize only, or restore first?

This is the decision people skip, and it makes a big difference.

  • Clean scan, no damage? Use a pure colorization pass. It adds color and leaves the picture structure as it is, which keeps it faithful to the original.
  • Scratches, creases, stains, fading? Repair first. If you colorize a cracked print, you get a cracked print in color, and sometimes the cracks pick up odd tints.
  • Faded color print (the orange/magenta 1970s look)? That's a restoration job rather than pure colorization.

As an example, Colorize Old Photos offers exactly these three modes: colorize only, restore only, and restore + colorize. Its colorize-only mode uses a dedicated colorization model that keeps the structure as is, while restore + colorize uses a generative editing model that removes damage and adds color in one pass. That trade-off applies to most tools: generative repair can fix more, but it's also more likely to subtly change details, so check faces carefully.

Step 4: Run it and compare side by side

Upload your scan and run the mode you picked. A good tool lets you drag a before/after slider. Use it. Look at:

  1. Faces first. Do skin tones look like skin, evenly across the face and hands? Are eyes and teeth free of strange tints?
  2. Edges. Is color bleeding across boundaries, such as sky into hair or a shirt color spilling onto the background?
  3. Large flat areas. Walls, sky and clothing sometimes come out muddy brown or greyish. That's the model hedging between colors.
  4. Identity (for restoration). Compare the restored face with the original at 100% zoom. The person should still look like the same person. If a jawline, smile or eye shape has changed, use a less aggressive mode.

Step 5: Re-run instead of fighting a bad result

Colorization models are usually not deterministic. If a dress comes out an odd color or a car looks wrong, simply run it again. A second attempt often gets the detail right. This beats hours of manual masking in an editor for most family photos.

If you know the real color of something (Grandma's coat was red, she talked about it for years), trust that over the model. You can fix one object in a photo editor with a hue adjustment layer and a mask.

Step 6: Save, label and share responsibly

  • Keep three files: the original scan, the AI result, and any edited final version.
  • Put "colorized with AI" in the filename or caption. Future relatives deserve to know the colors are an estimate, not a record.
  • Write the names, place and year in the file metadata or a sidecar text file while people who remember are still around.

Bonus: turning it into a print

A restored photo makes a great gift. A few practical notes:

  • Check the output resolution before ordering a print. Around 2000 px on the long side is comfortable for a 5×7 in print and still okay for 8×10.
  • Matte paper suits old photographs better than glossy.
  • A simple wood or brass frame keeps it on a shelf instead of in a drawer.

A note on privacy

Family photos are personal. Before uploading to any online tool, read how long it keeps your files and whether they are used for training. For example, Colorize Old Photos states that uploads and results expire 24 hours after upload and are never used to train AI models. Look for a statement like that from whatever tool you pick.

Quick checklist

  • [ ] Choose a sharp portrait or outdoor shot to start
  • [ ] Scan at 600 DPI (or flat, glare-free phone photo)
  • [ ] Crop and straighten; keep the original untouched
  • [ ] Damaged? Restore + colorize. Clean? Colorize only.
  • [ ] Check faces, edges, flat areas and identity
  • [ ] Re-run if a detail looks wrong
  • [ ] Label it as AI-colorized

That's it. Start with one photo of someone the family will recognize immediately. It's the quickest way to see what works before you tackle the whole shoebox.