Colorized by AI: which colors you can trust and which are guesses

The first time you see a black-and-white photo of your grandparents come back in color, the reaction is usually two things at once. Delight, because their faces suddenly feel closer. And doubt, because one question is hard to shake: was that dress really that color, or did the machine make it up? That question deserves a straight answer before you colorize a box of family photos, not after a print is already framed on the wall.

The short answer: plausible, not original

The AI guesses. But the guess isn't random, and knowing how it guesses tells you which parts of the result to trust.

A black-and-white photograph keeps only light and dark. Whatever color was in front of the camera was never recorded, so it isn't hiding anywhere in the file. What a colorizer does is look at the shape and context of each area — this is skin, that's sky above a horizon, those are leaves — and choose a color that's plausible for that kind of area. It answers "what color is this kind of thing usually?" It has no way to answer "what color was this thing that day?"

Researchers at UC Berkeley put it plainly in their 2016 paper on the problem: "Given a grayscale photograph as input, this paper attacks the problem of hallucinating a plausible color version of the photograph." They also call the task "clearly underconstrained" — there are many right-looking answers and no way to tell from the gray tones which one was real. In a "colorization Turing test," they showed people a real color photo and an AI-colorized one, and their method fooled people on 32% of the trials. That's a measure of how convincing the colors are, not how correct they are, and it says nothing about any particular tool, ours included.

Colorize black-and-white photos runs into the same limit: it picks plausible colors from what it sees, and it has nothing to adjust — no style, no color picker, no strength setting. The photo you choose is the whole input. So the useful question isn't whether the colors are right — some will be, some won't — but where in the picture the guesses tend to land close.

Fourteen parts of a photo, from most to least trustworthy

The table runs from the areas you can usually trust to the ones you should question, and it deliberately has no numbers. Colorization isn't a measurement, so scoring it would only look precise. There are three levels, in words: usually close, depends on the photo and often off, and that's expected.

Part of the photo	What AI usually gives	How close	Why	If it's off
Skin	Warm skin tones, from light to deep	Usually close	Usually believable, but the exact shade is still a guess: lighting, film and fading all shift it	Don't correct a shade from memory alone; ask someone who knew the person
Hair	Black, dark brown or gray, following how light each part reads	Usually close	Light and dark in the print point the right way for most hair	If gray hair comes out brown, keep the black-and-white version of portraits that matter
Daytime sky	Blue, paler toward the horizon	Usually close	A daytime sky has only a few possible looks, and its place at the top of the frame gives it away	A dusk sky can come back as midday blue; mention it when you share
Clouds	White, light gray in the shadows	Usually close	Clouds have almost no color of their own, only light and shade	Rarely needs anything
Grass and trees	Greens, yellower in strong sun	Usually close	Plants keep to a narrow palette	An autumn or drought scene turned lush green is worth a line in the caption
Water	Blue to green, following the sky	Usually close	Water takes its color from the sky above it and what lies beneath	Muddy river water turned clear blue: treat it as scenery, not evidence
Wood (porches, fences, furniture)	Warm browns	Depends on the photo	Bare wood is predictable; painted wood could be any color	Note the real color in the caption rather than repainting the photo
Painted walls, siding and brick	Off-white, gray, pale yellow, brick red	Depends on the photo	Brick and weathered siding are predictable; paint is a choice someone made	Crop tighter on the people and try again, or let the wall be wrong
Leather and shoes	Browns and black	Depends on the photo	Most shoes of an era fall in a narrow range, but not all of them	Usually too small to matter
Cars and chrome	Metallic gray for chrome, some body color for the paint	Depends on the photo	Bare chrome is easy to recognize; the paint was picked at a factory or by an owner	Check it against what the family remembers about the car
Plain clothing	A neutral color that suits the scene	Often off, and that's expected	Clothing color was chosen by the person wearing it, and the print keeps no trace of that choice	Ask the family, then write the color in the caption
Patterned fabric	The pattern keeps its shape; the colors are anyone's guess	Often off, and that's expected	Every patch of a pattern is a separate color choice, with nothing natural to anchor it	Keep the black-and-white version alongside, especially if the dress is the reason you love the photo
Uniforms (military, school, team)	A uniform color that looks right for the era	Often off, and that's expected	Uniform colors are set by regulation or by a school, and nothing in the gray tones carries that rule	Treat the color version as a reference, not as the record
Flags, insignia and printed shirts	Loose patches of color, sometimes bleeding into the pattern	Often off, and that's expected	Their colors and layouts are pure convention	Keep the black-and-white version as the reference copy

One thread runs through all fourteen rows, and remembering it saves remembering the table: how wide is the natural range of colors for that thing? Where nature sets the color — sky, clouds, grass, water — the plausible answers cluster together, so the guess lands close. Where people set the color — clothing, paint, uniforms, flags — the guess often misses, for a simple reason. That information left the photograph the day it was taken.

Why skies, lawns and water usually look right

The sky is the clearest case, because physics leaves it so few options. NASA's explanation for young readers works well here: "Blue light is scattered more than the other colors because it travels as shorter, smaller waves." And near the edge of the frame, "Closer to the horizon, the sky fades to a lighter blue or white." A colorizer that paints the top of a daytime landscape blue and lightens it toward the horizon is following the same rule the atmosphere follows.

The rule has an exception you can see coming. At sunset, NASA explains, "Even more of the blue light is scattered, allowing the reds and yellows to pass straight through to your eyes." Nothing in a gray print records which of those two skies you had, which is why an evening snapshot can come back looking like noon. Lawns, trees and lakes behave much like the daytime sky: their colors are narrow and predictable, so the guess is usually close even when it isn't exact.

Why uniforms, flags and cars usually don't

Now take the opposite case, a kind of photo people often want colorized: a grandfather in uniform. The National WWII Museum has a piece on the Army officer's uniform of that era, the one nicknamed pinks and greens. It explains: "The original WWII Army officer's winter service uniform consisted of a dark olive-drab gabardine wool coat with a sewn-on cloth belt (greens) and light-shade drab trousers (pinks). The brim of the service cap and service shoes were Army russet brown." And for everyone else: "Enlisted personnel of the era wore a far less flashy uniform; both trousers and service coat were of a medium olive-drab shade."

None of that can be read off a gray print. A colorizer has no way to know that an officer's trousers were a lighter drab than his coat, or that his cap brim was russet brown, because those colors were set by regulation, not by anything in the light. Expect a plausible soldier's outfit in some green or brown, and expect the details to be off.

Flags are even more clearly a matter of convention. Executive Order 10834, printed with the flag law in the U.S. Code, describes the flag in words: "The flag of the United States shall have thirteen horizontal stripes, alternate red and white, and a union consisting of white stars on a field of blue." A black-and-white print of a flag only records that some stripes are darker than others. The red, the white and the blue come from the rule, not from the photo, and the same goes for school colors, team jerseys and the paint on the family car.

How to check the colors you got

The best check isn't software. It's people and old color pictures.

Ask someone who remembers. Before you decide that a grandmother's dress was blue, ask her, or ask the aunt who was there. The AI doesn't have that information. Someone in the family might, and the question often turns into a longer story than the photo itself.

Look for color film from the same years. Color snapshots go back further than many people think. Kodak's own timeline records that in 1935 "KODACHROME Film was introduced and became the first commercially successful amateur color film. It was initially offered in 16 mm format for motion pictures; 35 mm slides and 8 mm home movies followed in 1936." And in 1961, "The company introduced the first in its very successful line of KODAK CAROUSEL Projectors, which featured a round tray holding 80 slides." If there's a round slide tray in an attic or a closet, it may hold the real color of the same house, car or coat. Real color from the time beats any guess.

Write the right color down instead of painting it in. When you learn that the car was red, put that in the caption or the file's description. A note keeps the fact. Repainting the photo turns one guess into another.

Old family photos are personal, so a word on uploads. Vipix does not use your photos to train AI. They're encrypted on the way in, the AI company that runs the colorizing wipes its copy inside 24 hours and Vipix clears its own storage on the timetable spelled out in the Privacy Policy.

Share it as a colorized version, not as the original

A colorized photo is one reading of the original, not a new edition of it. A few habits keep it honest:

• Label it. "Colorized with AI" in the caption or the file name is enough.
• Send both versions. If you share a colorized portrait with relatives, attach the black-and-white original, so whoever looks at it later can tell what the camera recorded from what a computer inferred.
• Keep it out of the record. A family history book, an archive or anything used to prove a point should carry the black-and-white original. If you include the colorized version, mark it as colorized.

Before a colorized photo goes online or into a print order, give the downloaded file a look; some downloads carry a small vipix.vn mark, depending on which tool produced them and on whether you were signed in when you ran it.

Restore first, colorize second

One more order matters, and it affects the colors directly. A scratched, stained or faded photo can be colorized, but it's better to run Restore old photos first and colorize the repaired copy. The reason is the mechanism described above. The colorizer has to work out which area is skin, which is a sleeve and which is sky before it can choose colors, and a white scratch across a face or a large stain blurs those boundaries. When an area is misread, the color lands in the wrong place.

The two tools do different jobs. Restoration is built to repair damage, and one of the things it does is rebalance color; colorizing is a separate tool built to make the color version. Compare each result with your original before you use it. If you're not sure how damaged a print is or what can still be fixed, which kinds of damage AI can and can't repair goes through it case by case. And if you're not sure the photo was ever truly black and white, what kind of black-and-white photo you actually have is worth five minutes first. For what a single restoration run covers, see how one AI restoration run handles fading, scratches and soft focus.

If you want to test everything in this article yourself, there's a quick way that's also fun. Pick a photo whose colors someone in the family still remembers — a wedding dress, the family car, a front door — run it through Colorize black-and-white photos and look at the result together with that person. When a run uses credits, its cost is listed on the tool page before you begin. You'll see right away how well the table holds up, and how good the guesses are where the AI has something to go on.

Frequently asked questions

Are AI-colorized photos accurate?

They're plausible, not original. The colors are a guess based on what each part of the photo looks like, so areas whose color nature sets, such as sky, grass and water, usually come out close, while clothing, uniforms and flags are often off.

Why did my grandfather's uniform come out the wrong color?

Uniform colors were set by regulation, not by anything a gray print records. The World War II Army officer's uniform nicknamed pinks and greens, for example, paired a dark olive-drab coat with light-shade drab trousers, a difference no colorizer can see in the gray tones.

How can I find out what the real colors were?

Ask relatives who remember, and look for color slides or home movies from the same years. Kodachrome film was introduced in 1935, so many families have real color from earlier than they expect.

Should I restore a photo before colorizing it?

Yes, if it's damaged. Scratches and stains blur the boundaries the colorizer relies on, so repair the photo first and colorize the repaired copy. One of the things restoration does is rebalance color, so compare the repaired copy with your original before you colorize it.

Is it OK to post a colorized family photo?

Yes. Label it as colorized with AI, and keep the black-and-white original alongside it so nobody mistakes the guessed colors for the real ones.