Colorizing old photos with AI: a guide and practical tips

Colorizing a black and white photo used to be a craft: someone painted by hand, layer by layer, deciding every tone. Today an AI does it in seconds — and surprisingly well — because it has learned from millions of real photos what color things usually are: skin under a certain light, the sky at a certain hour, a period uniform.

Understanding that "usually" is the key to using these tools well.

What AI gets right and what it invents

The model does not discover the original color: it infers it. That works very differently depending on the element:

| Element | How reliable the color is | | ------------------------------------ | ---------------------------------------------------------- | | Skin, sky, foliage, wood | High: the physics of light leaves real clues in the grays | | Brick, stone, paths | High | | Neutral clothing (suits, uniforms) | Medium: plausible, not documentary | | One specific dress, one specific car | Low: the red it shows you is a guess | | Flags, coats of arms, insignia | Low — and here an invented color can be a historical error |

The practical consequence: a colorized photo is an interpretation, perfect for bringing a family portrait back to life, and not proof that "the house was green". If the exact color matters — for historical documentation, say — you need a source, not an AI.

Prepare the photo before colorizing

The most common mistake is colorizing a damaged photo as-is. Color is applied on top of whatever is there: if there are stains, scratches or yellowing, you get a damaged photo… in color, often with muddy tones where the damage confused the model.

The order that works:

1. Restore it first: repair scratches, stains and fading while it is still black and white.
2. Check the restoration (faces and text above all).
3. Colorize afterwards, on a clean base.

Two separate steps beat asking for everything at once: each process works on a coherent image instead of competing with the other.

Tips that make a difference

• Start from the best scan you can get. A 600 dpi scan in good light gives the model more texture clues, and texture guides color. (How to scan properly: our restoration guide.)
• Sepia photos are fine too. You do not need to convert them to pure black and white first; the process handles the warm tone directly.
• Groups and portraits work better than distant landscapes. The more recognizable each object is, the better the model infers its likely color.
• If a color comes out wrong (it happens: a garment tinted an impossible shade), run it again. The inference has a random component and a second pass may choose better.
• Always compare against the original before sharing: check that people's features are still intact. Colorizing should not change a face — if it did, discard the result.

What to expect from the result

A good modern colorization produces natural skin, believable skies and clothing consistent with the period. Where the machine still shows is in small details that carry meaning — a badge, a logo, a complex print — and in very crowded scenes.

For what most of us actually want — seeing our grandparents in color again — the state of the art is more than good enough. Upload your photo to colorize photo and see for yourself.