← All guides · 2026-10-23 · PhotoUnfade team

Does AI Photo Colorization Change the Photo? We Measured Our Own Colorize Mode

Yes — colorization changes a photo, by definition: it adds color the camera never captured. The question that actually matters is whether it changes anything else. Done well, AI colorization is paint laid over the photograph's existing structure: same faces, same expressions, same fabric folds, same lighting, with color inferred on top. Done badly, it quietly repaints the photo underneath, and you get back a prettier picture of someone who is not quite your relative. We built an opt-in Colorize mode for black-and-white photos at PhotoUnfade, and before shipping it we measured it the same way we measured our restoration engine: pixel-by-pixel, against acceptance thresholds we wrote down in advance. Here is what we found — including the places where AI colorization cannot be trusted at all.

What does AI photo colorization actually do?

AI colorization infers plausible colors for a grayscale image from two sources of evidence: the brightness values in the photo itself, and statistical patterns the model learned from millions of color images. Skin at a given brightness usually sits within a narrow hue band; foliage, sky, denim and brick each have their own likely ranges; and the model combines those priors with what it recognizes in the scene — a face, a uniform, a lawn — to pick specific colors.

Two implications follow. First, the result is a historically plausible reconstruction, not a historical recovery: the color of a dress, a car, or a medal ribbon is an educated guess, and it can be wrong even when it looks convincing. Second, because the model works from the existing tonal structure, a well-constrained colorizer should change only color — the lightness and structure of every pixel should stay put. That second property is measurable, and it is the heart of our test below.

Is AI-colorized color historically accurate?

No — it is plausible, not factual, and the distinction matters most for exactly the details family historians care about. Take military uniforms. The common failure in consumer tools is a 1940s US Army uniform coming back bright blue; the real thing was olive drab, a muted brownish-green. In our own September 2026 evaluation, our colorize prompt returned olive-drab, brown and khaki tones for a 1940s serviceman's uniform — plausible. But there is no guarantee for your photo: insignia, medal ribbons, regiment-specific piping and heraldic colors are exactly the kind of detail the model cannot know, because they are not encoded in a grayscale image. The same goes for the dye colors of a specific 1930s dress or the paint on a specific 1955 Chevrolet.

Skin is the hardest and the most important case. Push skin saturation too far and everyone comes back with the candy-orange "AI plastic" look. Measured on the same style of portrait, a popular all-in-one photo editor's colorized output ran a mean saturation of 0.264 (on a 0-1 scale) — nearly double the 0.134 we measured on our own Colorize output, which sits in the range portrait skin reads as natural. And where the photo itself is missing information — a torn corner, a face half destroyed — the colorizer has nothing to infer from; color laid over a reconstructed area is inference stacked on inference.

How do you measure whether colorization changed the photo?

With three numbers, all computable with ordinary image tools. The first checks that only color changed; the second checks that color actually arrived; the third checks that people still look like people.

1. Luma structure delta. Convert the before and after images to grayscale, resize them to identical dimensions, and take the mean absolute pixel difference on a 0-255 scale. ("Luma" is the brightness channel of an image, stripped of color.) This measures whether the underlying photograph was repainted. Our rule of thumb from calibration: under 8 means the structure is essentially untouched; over 15 means visible repainting.
2. Colorful pixel share. The percentage of pixels whose saturation exceeds 0.12. A genuine black-and-white original measures only a few percent, so this confirms colorization actually happened rather than returning a slightly warm grayscale.
3. Skin-region hue and saturation. Extract the skin pixels (where red is greater than green, green greater than blue, and the red-minus-blue gap falls between 15 and 120 — the standard range for human skin), then compute median hue in degrees. Healthy human skin sits in a roughly 15-50° orange-red band. Pair that with mean saturation: portrait skin looks natural near 0.13-0.17 and plastic once pushed toward 0.26.

What did we measure in our own Colorize mode?

Every acceptance check passed. On October 9, 2026 we ran two black-and-white originals through the live production pipeline — a demo photograph and a genuine archival-era portrait — using the same servers, model and prompts our users get. The thresholds were fixed before we saw the results.

The number we care about most is the structure delta: 8.0 and 10.5 out of 255 means the colorization moved color without moving the photograph — faces, fabric folds and lighting are pixel-for-pixel the same scene, just in color. The skin measurements back that up: a median hue of 29° with every skin pixel inside the healthy band, at a natural saturation. Our earlier standalone evaluation on a 1940s soldier portrait told the same story — structure delta 9.7, uniform tones olive-drab and khaki with no blue-uniform error, saturation 0.167.

Two honest caveats. First, two or three photos is a quality gate, not a population study — your photo is the test that matters, which is why the mode shows the original beside the result. Second, generative models are stochastic: in our calibration runs the same input restored twice produced noticeably different tones (one pair measured +19.2 and -1.7 on our warmth scale). We validate with paired before-and-after measurements rather than trusting any single output, and you should eyeball your own result the same way.

Why are colorization and restoration separate modes?

Because merging them measurably damages fidelity. During evaluation we also tested the single-pass approach most one-click tools use — one instruction that restores damage and colorizes together. Its luma structure delta came back at 12.5, well into repaint territory, and the output had begun redrawing background regions, including invented blue blotches with no source in the original. That is the same failure we measured in our face-accuracy test, where an all-in-one editor put meaningful invented color on 77.5% of pixels. So PhotoUnfade splits the two jobs. Subtle and Full modes repair damage and are forced to leave black-and-white photos black-and-white (a deterministic correction layer strips any invented color; measured 0.0% after correction). Colorize only adds color and never repairs anything.

The practical consequences: Colorize is strictly opt-in — the button appears only when you upload a photo our checks detect as black-and-white, because colorizing on purpose is a feature and colorizing by accident is a bug. Each run costs one credit, the same as a restoration (see pricing), takes roughly eleven seconds, and costs us about $0.025 of compute. If a photo is both black-and-white and badly damaged, run Full first and Colorize second — two passes, in that order — because colorizing torn or mold-eaten areas just tints the damage. Our AI vs. professional comparison covers when damage is severe enough to hand the print to a human retoucher instead.

Which photos are good candidates for colorization — and which are not?

Good candidates share one property: the photograph is intact, and color is being added for sharing and display, not for evidence. Not every photo qualifies.

Colorize is a good fit for:

• Intact black-and-white portraits, weddings, family gatherings and street scenes where faces are clearly visible
• Photos headed for a memorial display, a family photo book, or a gift print
• Well-scanned originals — scan at 600 dpi so the model has real detail to work from (settings in our digitizing guide)

Colorize is the wrong tool when:

• The photo is genealogical or archival evidence. A guessed color can be mistaken for a fact by the next researcher; if you colorize it, label the file as colorized and keep the untouched scan alongside it.
• The specific colors are the research subject — uniform identification, vehicle restoration, dating by clothing dye. Plausible is not evidence.
• The photo is heavily damaged. Colorization does not repair tears, mold or water damage — restore first, colorize second.
• You do not have a preserved original scan. Never let a colorized version become the only surviving copy.

The bottom line

Does AI colorization change the photo? It should change exactly one thing — color — and our measurements say ours does: structure deltas of 8.0 and 10.5 on a 255 scale, skin at a natural 29° hue and 0.134 saturation, with every acceptance threshold set before we ran the test. What it can never give you is certainty about the colors themselves, so treat every colorized photo as a well-labeled interpretation, keep the original scan, and be skeptical of any tool — including ours — that will not show its numbers. You can run the experiment yourself for free: upload a black-and-white photo at PhotoUnfade, choose Colorize when it appears, and drag the slider between original and result. Three free restorations, no account, no credit card. For the full path from shoebox to restored and colorized archive, start with our step-by-step restoration guide.