How to Colorize Black and White Photos With AI

A grey value is not a colour with the saturation turned down. It is a brightness reading that dozens of different colours could have produced, which is why the same mid-grey could have been a forest green, a brick red or a navy blue on the day.

So the model is not recovering anything. It is betting on what a scene like this one usually looks like, and it bets well on some things and badly on others. The pattern is consistent enough to plan around.

Order of operations first, because it decides most of the outcome: repair, then colour, then grade. Colour applied to a damaged scan paints the damage, and a crease that was merely visible becomes a crease with its own hue.

1. 1

   Fix the physical damage first

   Tears, creases and stains go before anything else. A repair pass costs the same 13 credits as a colour pass and it changes what the colour pass has to work with.

2. 2

   Neutralise any cast

   Sepia and yellowing are not monochrome. If the scan has drifted amber, the colorizer reads amber as data and pushes everything warm. Ask for a neutral black and white first.

3. 3

   Open the colorizer and load the clean file

   The photo colorizer takes the photo in the Photo box, the ratio chips beneath it and nothing else. There is no colour picker and no per-region control, so steering happens in the Prompt box.

4. 4

   Name what you know

   Garment colours, hair colour, the season, the light source. Anything you state is treated as fact; anything you leave out is guessed.

5. 5

   Run, then compare against the monochrome

   Keep both files open side by side. The question is not whether the colour version looks nice, it is whether it contradicts anything the grey version told you.

Run two makes the argument for the order above. Same tool, same ratio, same price, but the input is the file that came out of the restoration pass rather than the raw scan.

Both runs went through GPT Image 2. Nothing about the colour model changed between them. The only variable was the quality of the file going in, which is the whole point.

Two failure modes turn up constantly, and they have opposite causes.

• **Grey, lifeless skin.** The model hedged. It happens on low-contrast scans where the face has little tonal separation, and on group shots where faces are small. The fix is a tighter crop or a repair pass that restores contrast before colour is attempted.
• **Orange or sunburnt skin.** The input was not neutral. A sepia print, a yellowed scan or a warm phone capture pushes the whole palette hot, and skin shows it first because that is where you are looking.
• **Waxy, poreless skin.** Not a colour problem at all. That is over-processing from a previous pass, and colour is merely revealing it.

The correction sequence that works: neutralise, colour, then grade. Ask the pass for neutral, natural skin under the light source that was actually in the room, and name that source. Window light, a bare bulb and a flash produce three different skin renderings, and the model will use whichever one you mention.

People also arrive here from the reverse question, how to make color photo black and white in photoshop, and then wonder why converting back and forth loses something. It is the same asymmetry: throwing colour away is lossless bookkeeping, putting it back is invention.

The Prompt box is the only steering wheel. There is no eyedropper, no swatch panel and no way to paint a region, so precision comes from writing precisely.

Colorize naturally for a 1954 summer afternoon in England. The woman's dress is pale yellow, the man's suit is charcoal, the front door behind them is dark green. Keep skin tones neutral, keep the print's grain, do not change either face.

Four things are doing work in that sentence. A date and place set the palette era. Two garment facts remove the least reliable guesses. A skin instruction blocks the warm drift. And a face-protection clause stops the pass restyling the people while it is busy with the coat.

• **State the era, not just the decade.** Colour films of the 1950s, 1970s and 1990s rendered differently, and naming the period nudges the palette toward it.
• **Give a light source.** Overcast, low sun, tungsten, flash. This does more for realism than any colour name.
• **Correct one thing per run.** If the coat came out wrong, run again mentioning the coat. Rewriting the whole prompt changes everything else too.
• **Keep a written note of what was known versus guessed.** Six months from now nobody remembers which was which, and that note is the difference between a family record and a nice picture.

There is a real argument here and it is not settled. Colour makes distant people legible to modern eyes, which is why the practice spread. It also overwrites the photographer's actual choice with a machine's statistical average of similar scenes.

Conservation bodies are consistent on one point: the object is what needs protecting. Preservation guidance from NARA (2026) and from the Getty (2026) is about the physical print, and a derivative made from a scan never touches it. The exposure is elsewhere. An unlabelled colour version, two shares later, becomes what people believe the scene looked like.

Colour it if you like. Just never let the colour version travel without a sentence saying where the colour came from.

— Maya Chen

• **Keep the monochrome master.** Always, and in the same folder.
• **Label the derivative in the file name.** Something as blunt as `_ai-colour` survives being emailed around.
• **Say what was known.** If the dress colour came from a relative, that belongs in the caption. It is the only part of the colour that is evidence.
• **Do not colorize other people's archives on their behalf.** Institutions have their own policies, and a helpful upload can create a lot of work for a curator.

You can. The search for how to colorize black and white photos free tends to end at a stamped output, a tiny quota or a size ceiling. Each line below comes from the tool's own page, checked 7 September 2026.

One thing to know before you shop on price: the grant in the image generator is shared across tools, so the repair-then-colour sequence this article recommends fits inside a single new account without paying anything. That matters more than the headline number, because colorizing an unrepaired scan wastes the pass either way.

The handset is fine. The capture is the risk. Almost every disappointing phone colorization traces back to the copy shot rather than the model.

1. 1

   Photograph the print properly

   Shoot in indirect daylight with no flash, holding the handset square to the paper, and use the camera at full size rather than screenshotting the gallery.

2. 2

   Kill the cast on the copy shot

   A phone auto-white-balancing against a sepia print will make it worse. If the result looks amber on screen, it will look amber to the model.

3. 3

   Work in the browser

   The tools open in mobile Safari or Chrome, so there is nothing to install and the file goes straight from the camera roll.

4. 4

   Judge it on a bigger screen

   Skin tone decisions made on a phone at arm's length rarely survive a laptop. Colour is the one thing a small bright screen flatters most.

If the print is in a frame, take it out. Glass adds reflections that the model interprets as highlights, and a highlight in the wrong place becomes a colour in the wrong place.

Five tells, and none of them is about colour accuracy. They are about consistency, which is what the eye actually checks.

• **Uniform saturation.** Real scenes have dull corners and bright centres. When every surface is equally vivid, the picture reads as painted.
• **Clean colour over dirty grain.** The monochrome file has grain; the colour layer usually does not. Ask for the grain to be preserved and the mismatch goes away.
• **Period-wrong hues.** Neon brights on 1940s clothing, or a modern white that no fabric of the era held after a season of washing.
• **Lips redder than any film stock rendered them.** A recognisable habit of colorization models, and the fastest way to date a colorized image to the year it was processed.
• **Backgrounds that agree too much.** Every leaf the same green, every brick the same red. Real materials vary within themselves.

The fix for four of those five is to run again asking for lower overall saturation, retained grain and varied material tone. The fifth, period-wrong colour, is not fixable by prompting harder. It needs a fact from outside the photograph, which is where a relative's memory or a written record beats any amount of processing.

Anyone deciding how to colorize black and white photos without ai should know that the manual route is still the accurate one: masks, layer modes and hand-picked hues in a full editor produce a defensible result because every colour was a decision. It is also a day's work per photo, which is why almost nobody does it for a shoebox.