How to Colorize Black and White Photos (and Get Natural Colors)
How AI photo colorization works, how to get skin tones and period colors that look right, and when a colorized photo crosses into guesswork — with sources.
PhotoRestoration Team
A black-and-white photograph feels like a window with the glass painted over. The faces are your family's — the colors just are not. Colorization is the craft of putting them back, and in the last few years AI has turned it from a specialist's week of hand-painting into a one-click step. This guide explains how the model decides what color things should be, how to get results that look natural rather than painted, and where the honest limits are.
How does AI colorization actually work?
A colorization model is trained on millions of color photographs that were converted to greyscale. During training it learns the statistical relationship between greyscale patterns and their original colors: the gradient that means skin under northern light, the texture that means army wool, the way grass reflects versus how a painted wall does.
At run time, the model sees your photo's greyscale values and predicts a color for every pixel — plus a confidence for each guess. Sky is easy: it is almost always blue. Uniform backdrops and standard-issue uniforms are easy. The hard cases, covered below, are objects that exist in many plausible colors.
The output quality depends directly on the quality of the greyscale input, which is why professionals always restore the photo first and colorize second — scratches and fading get misread as color information if they are still in the image.
What about colors the camera never recorded?
This is the question worth asking honestly: where does "the" color come from, when the original recording never had any?
Part of the answer is research. Uniforms, medals, vehicles, school blazers, factory paint — much of the material world of the twentieth century was standardized, and those documented colors are knowable facts. When you colorize a 1943 portrait, the model is working in a world where the correct answers were recorded and still exist in archives and museums.
The other part of the answer is probability. The dress that could have been red or green gets one of them, based on the fashions the model has seen. For family use this usually does not matter — but if a photo is going to serve as a historical record, it is fair to say out loud that colorization is an interpretation of the greyscale original, and the monochrome scan should be preserved alongside it. The Library of Congress's personal digital archiving guidance treats any derivative of a scanned original this way: keep the source.
How to colorize a photo, step by step
On our colorize tool the flow is deliberately short:
- Start from the best scan you can make — 600 DPI, clean glass, print pressed flat. Our restoration guide covers the scanning checklist in detail.
- Repair damage before colorizing if the photo is scratched, torn, or stained. A stain the repair step would have removed otherwise becomes a color patch. The repair mode handles this in the same upload.
- Upload and choose the colorize mode. The preview is free, so you can see the model's interpretation before paying for the download.
- Check skin first, period details second. Faces are where the eye is most critical; after that, look at anything whose true color you know — the family car, the school uniform, the kitchen tiles.
- Keep both files. The colorized print for framing and sharing; the monochrome original as the archival record.
Can AI colorization be accurate?
Accurate enough that most viewers cannot tell the photo was ever black and white — and that is the right way to frame the question. Two honest statements sit side by side:
- Where reality was standardized (uniforms, skin under known lighting, manufactured goods), modern colorization lands on the true color most of the time.
- Where reality was arbitrary (the exact dress, the paint on the door), the model produces a plausible color, not a recovered one.
So: colorized family photos are fair visual memory, and museum-grade claims should be labeled as interpretations. What colorization is not, anymore, is the flat orange-and-teal look of the early tools — that reputation is a decade out of date.
Why do skin tones sometimes come out wrong?
Skin is the hardest class for three reasons: it occupies a narrow greyscale range that overlaps with wood, leather, and tan fabric; its true color varies enormously between people and lighting conditions; and every viewer is a built-in expert on what skin should look like, so a small error reads as a big one.
Three practical fixes:
- Fix the scan, not the result. A sepia or yellow cast biases the model; scanning per the checklist removes the bias at the source.
- Restore first. Faded areas carry corrupted brightness values, and the model maps corrupted values to corrupted colors.
- Avoid extreme lighting in re-photography. Harsh sunlight during a phone re-shoot blows out highlights, and blown highlights have no greyscale information to colorize.
Frequently asked questions
Is colorizing an old photo disrespectful?
Handled honestly, no — it is a way of making ancestors legible to descendants who scroll past anything monochrome. The respect lives in the details: keeping the unaltered scan, not pretending the colors are documented fact, and choosing a tool that renders skin and period colors conservatively.
Can a faded color photo be "re-colorized"?
Yes, and it is often the best fix for 1970s–80s prints whose dyes have shifted. Convert the scan to greyscale, then colorize — you are re-asking the model the color question with a cleaner input. The restoration tool handles the conversion in one pass.
Does colorization work on very old photos, like from the 1900s–1920s?
Yes, with a caveat: orthochromatic films of that era rendered certain colors — deep reds especially — much darker than they appeared in life. The model colorizes what the film recorded. Faces usually come out beautifully; a dark flowered dress may stay ambiguous, because the film genuinely did not capture its color.
How much does it cost?
The preview is free; full-resolution colorized downloads are a few dollars — the pricing page has the current credit packs. No subscription needed for a one-off project.