← All guides · 2026-09-28 · PhotoUnfade team

Do AI Photo Restorers Change Faces? We Tested 3 Tools Pixel-by-Pixel

Yes — some AI photo restoration tools change faces, and it is the single most common complaint in their reviews: "it doesn't look like my mother anymore." But "changes faces" is a vague accusation, so we decided to measure it. We ran the same photographs through three restoration pipelines — the two modes of our own engine and one popular all-in-one photo editor — and compared the outputs pixel-by-pixel against the inputs. This article reports what we found, including two real failures in our own product, the numbers behind each fix, and how you can run the same checks on any tool in about five minutes.

Why would a restoration tool change someone's face at all?

Because modern restoration is generative: the model does not just clean pixels, it predicts what the undamaged photo looked like, and when information is missing it fills the gap from patterns learned across millions of training images. That prediction is usually right — and when it is wrong, it is wrong in characteristic directions. We measured three of them:

• Skin tone drift. Models trained to make photos look "warm and nostalgic" push every face slightly tanner, ruddier or more saturated. One photo looks fine; a whole album slowly turns into strangers.
• Invented color. The same training bias says "old photo = sepia or colorized." Feed a tool a black-and-white portrait and it may hand back guessed color — a uniform, eyes and lips in shades nobody ever chose.
• Fabricated detail. Aggressive enhancement can manufacture eyelashes, iris texture and teeth detail that the original blur never contained. Sharper is not the same as truer.

None of these show up in a marketing demo. They only show up when you measure.

How do you actually measure whether a face changed?

You compare the restored image against the original on numbers that describe identity, not beauty. Our test extracts the skin-tone pixels (those where red > green > blue and the red-minus-blue gap falls between 15 and 120 — the standard range for human skin) and measures four properties of those pixels before and after restoration:

For black-and-white photos we add two more checks: the percentage of pixels with meaningful color (saturation above 0.12 — a true black-and-white scan sits near zero), and the gray neutrality of the remaining pixels (how far the gray tones lean warm).

One methodology rule turned out to matter more than any threshold: always resize both images to identical dimensions before comparing. We initially compared a 531×384 input against a 1024×736 output and "measured" a 48% sharpness loss that did not exist — it was pure resolution artifact. Re-run at matched size, the real number was −4%. Any tool comparison that skips this step is measuring the wrong thing.

What did we find in our own engine? Two failures, honestly reported.

We ran these tests on PhotoUnfade before writing a single marketing claim, and the first results were not flattering.

Failure 1: skin tone drift. Our first prompt version restored a 1940s portrait of a fair-skinned serviceman with a yellow bias of +9.8 and a saturation increase of +0.077 — both over our thresholds. His hue and lightness were stable (he was still visibly the same person, the same ethnicity), but the image came back warmer and ruddier, exactly the slow "tanning" drift described above. The fix was explicit skin-tone constraints in both restoration prompts (preserve each person's natural skin tone exactly; remove paper yellowing, never alter the underlying complexion) plus eight skin-drift terms in the negative prompt. After the fix, measured against the natural, undamaged original: hue deviation 3.4°, yellow bias 4.4, saturation change 0.025 — all passing, and the restored skin is statistically indistinguishable from the original.

Failure 2: our tool colorized a black-and-white photo nobody asked it to. A user-submitted black-and-white portrait (color pixels in the input: 2.5%) came back with 12.8% colored pixels in Subtle mode and 50.4% in Full mode — the AI had invented color for the uniform, eyes and lips. The same test showed neutral grays being pushed warm (gray neutrality drifted from +2.6 to +12.0/+13.3 on the red-minus-blue scale). This directly contradicted our own promise — black-and-white in, black-and-white out, unless you explicitly ask for colorization — so we shipped a hard correction layer: every input is analyzed first, and if it is monochrome the output is forced back to true grayscale; if the output came back yellower than the input, the gray balance is pulled back to where it started. Re-tested on the same photo: color pixels 0.0%, gray neutrality 0.0. Re-tested on color photos to make sure we had not broken normal restoration: no correction triggered, yellow-cast removal still working. We are telling you about this because a vendor that hides its failures has no incentive to fix them.

How did an all-in-one editor do on the same test?

We ran the same black-and-white scan through the "restore/enhance" mode of a popular all-in-one photo editor — the kind that bundles restoration, beautification and colorization into one button. Result: 77.5% of output pixels carried meaningful invented color, against 0.0% from our corrected pipeline. The output is, subjectively, a prettier picture. It is also no longer a document of what the photograph actually showed. If your goal is a family archive, a genealogy upload or a memorial print, invented color is a defect no matter how attractive it looks. (If you genuinely want colorization, that is a legitimate choice — but it should be a mode you select knowingly, not a side effect of clicking "restore." It is on our roadmap as an explicit, opt-in option.)

Does sharpening change how a person looks?

It can, and this is where "sharper" quietly becomes "fabricated." Honest sharpening — unsharp masking, which amplifies contrast along edges that already exist — makes a restored photo crisper without inventing content. We calibrated ours against a grid of settings and measured with acutance (edge energy, the standard objective sharpness metric) and overshoot (the bright halo artifacts that mark over-sharpening). Our final numbers: acutance +29% in Subtle mode and +77% in Full mode versus the input, with overshoot artifacts lower than the input (8.6% and 7.4% versus 26.1% in the damaged original) — sharper, without the white-edge glow. The all-in-one editor we tested scored higher on raw sharpness (roughly +101%), but produced about 8× more strong-edge points than the image could physically support: much of that extra "clarity" was manufactured edge, i.e., detail that was never in the photograph. On a landscape that is style; on your grandfather's face it is fiction.

How can you check any tool yourself in five minutes?

You do not need our scripts. Five checks catch almost every face-changing tool:

1. Demand a true side-by-side. If a tool will not show your original next to the result at 100% zoom, assume it has something to hide. Thumbnails conceal identity drift.
2. Test it with a black-and-white photo. If color comes back, the tool is inventing, not restoring.
3. Compare skin, not scenery. Zoom to 100% on a face: is the complexion the same family (pale, olive, deep) or did everyone get a tan and a flush?
4. Look at eyes and teeth first. That is where fabricated detail appears — an uncanny fixed stare, or enamel texture the original blur never contained.
5. Ask what the tool refuses to do. Vendors who list their limits (we do not colorize by default; large missing areas are reconstruction, not recovery) are more trustworthy than those claiming to fix everything.

What can't any AI do — including ours?

Keep expectations honest, because this is where AI restoration stops being restoration. When roughly 30% or more of a face is missing — torn away, dissolved by water, eaten by mold — no model can recover it, because there is no data left to recover. Anything it prints into that gap is an average face from its training set: plausible, smooth, and not your relative. That is invention, not recovery, and it matters enormously for genealogy and memorial work, where a believable fake is worse than an honest blank. For those few irreplaceable photos, a professional retoucher working from other real pictures of the same person remains the right answer — and starting them from an AI pass cuts their cost substantially. Likewise: input quality is the ceiling. A 600 dpi scan gives the model real information; a glare-covered phone photo caps the result no matter how good the algorithm is (our digitizing guide covers the exact settings). And color cast removal can only recover tone that is still encoded in the image — a print faded to uniform beige has less signal to work with than one yellowed but still contrasty.

The bottom line

Does AI photo restoration change faces? Measured honestly: a well-constrained pipeline keeps skin hue within a few degrees of the original, leaves black-and-white photos black-and-white, and sharpens real edges without inventing halos — ours now passes all of those thresholds, and we publish the numbers including the failures. An unconstrained "enhance everything" pipeline will tan your relatives, colorize their photographs and paint detail into their eyes, and call it restoration. The difference is not model size or marketing; it is whether the vendor measures, and whether they tell you what they found. If you want to judge with your own photograph instead of our test charts: PhotoUnfade gives you three free restorations — full resolution, no account, no credit card, no watermark — with the original always sitting next to the result so you can drag the slider and see exactly what changed. How we handle your photos is written down in our privacy policy, and what we refuse to process is in our acceptable use policy. For the complete workflow, see our step-by-step restoration guide.