Colorize Black and White Photos with AI

How the AI infers color from a grayscale photo

A black-and-white photo already carries more information than it looks like it does. Every shade of gray encodes a real brightness value, and brightness is a strong clue to color — skin, hair, fabric, sky, and foliage each reflect light in fairly predictable ranges. OldtoLife's model has learned those patterns from a huge number of real color photographs, so when it sees a mid-gray tone in the shape of a face, it can estimate a plausible skin tone, and adjust for shadows, highlights, and the apparent lighting of the scene.

Context matters as much as tone. The model recognizes what it's looking at — a jacket lapel, a curtain, a stretch of lawn, a straw hat — and uses that recognition to narrow down likely colors. A rounded shape at head height with a lighter texture reads as hair rather than a hat, and gets colored accordingly. The same logic applies to backgrounds: a textured area behind a portrait is more likely painted wall or drapery than sky, and the AI colors it that way instead of defaulting to blue.

Keeping the result believable, not cartoonish

The easiest way to ruin a colorized photo is to oversaturate it. Cheap colorizing filters tend to push every color toward the same handful of bright, flat hues, which makes a portrait look like a poster instead of a photograph. OldtoLife aims for the opposite — muted, natural tones that sit close to what film and prints from the period would actually have shown, with enough variation in skin, hair, and fabric that nothing looks uniform or painted-on.

Where the AI genuinely can't know a color — the exact shade of a dress fabric, or a car in the background — it chooses a plausible, understated option rather than an attention-grabbing guess. That restraint is deliberate. A colorized family photo is meant to feel like a real memory, not a stylized graphic, and a slightly conservative color choice almost always reads as more convincing than a bold, wrong one.

Why colorizing works best after restoring damage

If a photo has scratches, stains, tears, or heavy fading, colorize it after you restore it, not before. Damage on a black-and-white photo isn't just visually distracting — it changes the tonal information the AI uses to infer color. A water stain across a cheek reads as an unusual brightness value in that region, and the colorization model has to guess whether it's shadow, skin, or damage. Removing that damage first gives the model a clean, accurate map of tone to work from.

Running Restore before Colorize also avoids compounding errors. A scratch that gets colored along with the skin around it is much harder to fix afterward than a scratch removed cleanly beforehand. For a badly faded print, running Restore first also recovers contrast, which gives the colorization pass sharper edges and clearer tonal boundaries — the two effects reinforce each other rather than fighting.

• Scratches and stains distort the tone the AI reads as color
• Faded, low-contrast prints give the model less to work with
• Restoring first recovers edges the colorizer relies on
• Do Restore, then Colorize, for the cleanest end result

What colorization can know for certain — and what's a best guess

Some colors are well-constrained by physics and biology, so the AI gets them right with high confidence: sky is some shade of blue or gray, healthy skin falls in a narrow range, grass and leaves are green, wood and brick are warm browns. These are the areas where colorized results tend to look immediately convincing.

Other colors simply aren't recoverable from a grayscale image, because the original information was never captured. The exact color of a dress, a car, or a painted door is genuinely unknowable from tone alone — the AI has to choose something plausible for the era and setting rather than reproduce a fact. This is true of any colorization method, AI or hand-tinted, and it's worth keeping in mind: a colorized photo restores the feeling of color, and gets the constrained details right, but small guessed details shouldn't be read as a factual record of what someone was wearing.

Getting the best colorization results

A little preparation goes a long way toward a natural-looking result. The clearer and more damage-free the source photo, the more accurately the AI can read tone and texture, and the more convincing the color it produces.

Resolution matters too. A higher-resolution scan or photo gives the model finer detail to read — individual strands of hair, the weave of a fabric, the grain of a wooden chair — and finer detail generally leads to more precise, less generalized color choices.

• Restore scratches, stains, and tears before colorizing
• Use the highest-resolution scan or photo you have
• Sepia and yellowed black-and-white photos both work well
• If faces are soft or blurry, run Enhance for sharper detail
• Compare with the before/after slider before saving