Colorize a Black and White Photo

How it works

The photograph goes to DDColor, a model trained to predict plausible color for a grayscale image. It reads structure — shapes, textures, the relationships between regions — and assigns hues it judges consistent with what it recognizes.

The luminance of your original is preserved. The model is deciding hue and saturation and laying them over the tonal values the photograph already had, which is why a well exposed black and white image colorizes better than a flat one: the structure the model reads is the same structure your eye reads.

Nothing about the original file is altered. The colorized version is a new image.

The colors are a guess, and some guesses are safer than others

This is the part worth being clear about, because a colorized photograph looks like evidence and is not.

The model is reliable where the world is consistent. Skies are blue, foliage is green, skin falls in a narrow believable range, wood and stone and water sit where you expect. These are safe because there is a strong statistical answer and the model has seen it many thousands of times.

It is unreliable wherever the real answer was arbitrary. The color of a dress, a painted front door, a car, a book cover, a military uniform, a team strip, a flag — none of these can be recovered from a grayscale photograph, because the information was discarded when the shutter opened. The model will supply something confident and it has no way of being right except by chance.

So a colorized family photograph is a reasonable thing to hang on a wall and a poor thing to cite. If you need to know what color your grandmother's coat actually was, the answer is not in the file and no tool will find it there.

Where it tends to go wrong

Faded or yellowed originals confuse it. The model expects neutral grays, and a sepia scan gives it a color cast to interpret as well as tone. Neutralising the scan first — desaturating it fully back to gray — usually produces a noticeably better result than feeding it the sepia directly.

Large flat regions can come back patchy, because there is little structure to anchor a decision and the model drifts across the area.

Faces in a crowd get less individual attention than a single large face, so a group photograph often shows more skin tone variation between people than is plausible.

And anything the model misreads gets confidently colored as whatever it thought it was. A misidentified object is not a subtle error.

Getting a better result from a scan

Scan at the highest resolution the scanner offers and do not sharpen during scanning. Dust and scratches will be interpreted as structure and colored accordingly, so removing the worst of them first pays off.

If the print has a border, crop it. A white or cream border is a large flat region that gives the model nothing useful and can pull its interpretation of the rest.

Run it more than once at different crops if the result disappoints. The model is deterministic for a given input, so the way to get a different answer is to give it a different input.

What to do with the output

Keep both files. The grayscale original is the record; the colorized version is an interpretation of it, and conflating the two is how a guess ends up in a family archive as a fact.

If you are printing, check skin tones on paper rather than on screen. Colorized skin often sits slightly warm, which is flattering on a monitor and can tip towards orange in print.