ColorizeOldPhotos
Research

How accurate is AI colorization? What it knows, what it guesses, and how to check

Why a black and white photo does not contain its own colors, which parts of an AI colorization you can trust, and where to find real color references from the same era.

By the ColorizeOldPhotos editorial team

· 7 min read

AI colorization can look convincing, and that is exactly the problem. A natural-looking result is not evidence that the colors are correct. This article explains why, shows an example from our own tool, and describes how to find out what colors were really there.

A black and white photo does not contain its colors

A monochrome photograph records how bright each part of a scene was, not what color it was. Many different colors have the same brightness, so the information needed to recover the original is simply not in the picture. Researchers call this an ill-posed problem. The DDColor paper, whose model powers our colorize-only mode, frames colorization as having many possible outputs and builds in a color-refinement stage to cope with that.

What a model does is use what it has learned from many color photographs to predict what is most likely: skies are blue, foliage is green, skin has a familiar range of tones. That prediction can be excellent for common things and arbitrary for everything else.

What tends to be reliable, and what is a guess

Our assessment from reviewing results, not a measured accuracy figure.
ElementHow much to trust itWhy
Skin and hairFairly reliableFaces are the most common subject in training photos, and skin tones fall in a narrow range.
Sky, water, grass, trees, stone, woodMostly reliableNatural materials have a limited range of plausible colors.
Clothing, uniforms, flags, vehiclesA guessThese can be almost any color, and period dyes and paints varied.
Signs, labels, packagingA guessPrinted colors are design choices, not physics.
Small objects and background detailA guessToo little information for the model to infer.

Same photograph, two answers

The woman in a leather jacket and cap, colorized with the colorize-only mode: the jacket appears dark blue-greenThe same woman, colorized with the restore and colorize mode: the jacket appears brown
The same 1942 portrait colorized in two modes. The jackets differ (dark blue-green on the left, brown on the right), and so does the neon sign (yellow and blue letters, then warm white). The right-hand input was the damaged copy that we repaired in the same pass. Neither jacket color is known to be right. Photograph: Library of Congress, public domain.

Both versions look natural. They cannot both be correct, and no one looking only at the black and white photograph could say which is closer. This is the clearest demonstration we know of why a plausible color is not a verified color.

How to find the real colors

  1. Ask the people who were there. A relative may remember the color of a dress, a car or a front door. Their memory beats a model.
  2. Look at real color photographs from the same time and place. Color photography existed earlier than most people expect. The Library of Congress holds about 1,600 color photographs made for the U.S. Farm Security Administration and Office of War Information between 1939 and 1944, most of them 35mm Kodachrome slides. They show everyday clothing, vehicles, buildings and interiors in real color.
  3. Check records for the specific thing. Museums, regimental associations and collectors document the colors of uniforms, insignia and vehicles for specific units and years.
  4. Look for physical objects. If the dress or the medal still exists, the real thing settles the question.

How to present a colorized photo honestly

  • Say it is colorized. A caption such as “colorized with AI; colors are estimated” is enough.
  • Keep the black and white original next to it, or at least keep the file.
  • Do not use a colorized image as evidence of what something looked like. Use it as an illustration.
  • If you verified a specific color, say what the source was.

For the practical steps of colorizing a photo, see how to colorize black and white photos.

Sources

  1. Kang et al., “DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders” (ICCV 2023)
  2. Library of Congress, FSA/OWI Color Photographs: About this collection

About this guide

Written by the ColorizeOldPhotos editorial team. ColorizeOldPhotos is operated by windflash, founded by Jason Guo, who since 2005 has started a number of projects spanning AI applications, website development, game development and operations. Read more on the about page.

Found a mistake or an out-of-date claim? Email support@colorizeoldphotos.com and we will fix it and note the change here.

Try it on your own photo

Upload a scan and see the result in about half a minute. Previews are free and need no sign-up.

Keep reading