← All guides · 2026-11-11 · PhotoUnfade team

Which Photo Damage Can AI Actually Fix? A Measured Answer

The honest answer up front: AI photo restoration is excellent at reversing decay — fading, yellowing, haze, collapsed contrast — noticeably weaker against added damage like heavy stains and deep scratches, and unable to bring back content that is simply gone, such as a face that is mostly missing. Those three categories sound like marketing language, so let us be specific: we know where the boundary sits because we measured it, using controlled damage tests on known photos. This article is the plain-language version of those measurements, so you can predict what will happen to your photo before you upload it.

A note on method, because it matters for how much you should trust the numbers. Our calibration tests work like this: we start with a clean photo, apply controlled synthetic damage in known amounts — haze, yellowing, scratches, stains, blotches — then run the damaged version through the same restoration pipeline that serves the website, and measure the difference pixel by pixel. These are not customer photos and not cherry-picked demos; they are engineering tests we run whenever the pipeline changes. Where we cite real production behavior, it comes from timing and resolution data logged by the service itself.

Which damage does AI restoration handle best?

Decay — the slow, even, whole-photo changes that age brings. Yellowing, fading, a milky haze over everything, blacks that have turned to mush: this is the category where AI restoration performs closest to magic, because the original information is still there, just buried under a uniform veil. Removing a veil is exactly what these models do well.

Our controlled test for this category applies a haze-dominated aging pattern — a blend toward a warm gray fog color across the whole frame, the signature of a photo that sat in a shoebox for forty years. Restored through our full mode, the result showed a mean pixel shift of 21.4 on the 0-255 brightness scale (our reference benchmark for a visibly successful repair is about 20.6), contrast up by 9.9, and skin-tone hue drift of just 0.6 degrees — meaning the repair removed the veil without touching who the people are. In plain terms: the photo came back, and it still looked like the same family.

If your photo's main problem is that it has faded, yellowed, or gone flat and hazy, you are holding the best-case scenario. The chemistry behind why prints yellow and fade — and what the correction actually reverses — is covered in our guide to yellowed and faded photos, and the related case of photos that shifted pink, blue or brown is diagnosed in our color-shift article.

Which damage does AI struggle with?

Added damage — things sitting on top of the picture: heavy stains, dark blotches, thick scratches, sticky residue, ink. The reason is fundamental. An AI model has no way to know that the dark blotch over your grandfather's shoulder is coffee and not a shadow, a hat, or a tree. It reads the image as the image, and anything that looks like content gets treated as content — preserved, not removed.

We measured how hard this category is. In one calibration round we applied heavy local damage — the opposite of the even haze above — and ran it through restoration. The mean pixel shift came back at 2.8, against the 20.6 benchmark for a successful visible repair: almost nothing changed. Measured directly, 52.4% of the low-frequency stain structure survived the restoration, and 83.2% of the dark blotch damage was still there afterward. Compare that with the haze test above — same pipeline, same model, opposite outcome. The difference is entirely the damage type.

There is a middle zone. Fine, light surface scratches and ordinary dust do get removed well, because they are thin enough that the model can bridge across them from both sides. The scratch-specific guidance — including which scratch types survive — is in our scratched photos guide, and the related case of blur, which is a loss of edge information rather than added damage, is covered in our blur guide. For heavy stains, the honest expectation is improvement, not erasure.

What about torn or missing pieces?

Missing content is the third category, and the rule is: rebuild yes, invent no. When a tear removes a patch of repeating texture — grass, wallpaper, sky, a plain jacket — AI inpainting can continue the surrounding pattern across the gap convincingly. When the missing piece contains something unique — the lettering on a sign, the pattern of a specific lace collar, part of a face — any replacement the AI produces is fabricated. It may look plausible. It is not recovered.

We apply a hard internal rule here: if more than about 30% of a person's face is missing, restoring it means inventing a face, and invented faces are not restoration. The full reasoning, including why we consider fabricating a face worse than leaving the damage visible, is in our torn photo guide. The same honesty applies to mold: mold that has physically eaten the emulsion has removed the image itself, and no software can rebuild what is gone — see our mold guide for what can and cannot be saved there.

Does more pixel change mean a better repair?

No — and this is the measurement mistake almost everyone makes, including vendors. When we evaluated successive versions of the image model we license, the numbers looked backwards at first glance. One earlier version changed 13.1% of pixels on a damaged test photo; the version we ship changed only 11.6%. The 13.1% version was worse. Measured by direction rather than amount, it was actively harmful: it pushed the yellow cast up by 16.8, dragged gray neutrality down by 13.3, and raised brightness by 21 into washed-out territory. The version we ship changed fewer pixels but in the right directions: yellow cast down 2.5, contrast up 12.5, brightness essentially stable at −2.1. A third version changed just 1.4% of pixels, which turned out to mean it barely repaired anything at all.

The lesson generalizes: count what changed and in which direction, never just how much changed. A tool that rewrites a quarter of your photo has not restored a quarter of it — it has replaced a quarter of it. We hold ourselves to the same standard, which is why our published tests always report direction-aware measurements: yellow-cast shift, contrast, brightness, skin-tone hue — each compared at matched resolutions, after we once "measured" a 48% sharpness loss that turned out to be nothing but a resolution mismatch (the true figure, re-tested correctly, was between −4% and +10%). The face-level version of this discipline is in our pixel-by-pixel accuracy test, and the sharpening-specific numbers are in our sharpening guide.

How do you tell which category your photo is in?

Try the squint test. Step back, half-close your eyes, and look at the photo blurred. If the whole image looks uniformly wrong — too yellow, too faint, too foggy — you are looking at decay, the best case. If specific blobs and lines jump out against an otherwise readable picture, that is added damage, the harder case. If something is simply not there — a corner torn off, a face half gone — that is missing content, with the rebuild-versus-invent boundary described above. Many old photos carry all three at once; expect the decay to clean up beautifully, the light scratches to mostly vanish, and the heavy stains to fade rather than disappear.

What should you do before you upload?

A few preparations measurably change which category your photo lands in and how well the repair goes:

1. Scan at 600 dpi in TIFF or high-quality JPEG, with all automatic enhancement turned off. Fine scratches are far easier to remove when the scan actually resolves them, and scanner "auto-fix" bakes in changes nobody can undo. The full settings walkthrough is in our digitizing guide.
2. Handle physical emergencies first. Wet, moldy, or stuck-together photos need physical triage before any scanner touches them — the order of operations is in our water damage guide.
3. Do not pre-edit. Resist the urge to crop tight, brighten, or run a phone filter first. Every edit you make before restoration is one more thing the AI can mistake for content.
4. Set size expectations. Our pipeline outputs up to 2048 pixels on the long edge — enough for a true 300 ppi 4×6 print and a 5×7 just short of it, as the arithmetic in our cost guide shows. Production timing across 30 measured restorations averaged 22.7 seconds per photo.

The bottom line

AI restoration in 2026 is a measured capability, not a magic wand. Decay — fading, yellowing, haze — comes back almost fully; our controlled haze test shifted pixels by 21.4 against a 20.6 success benchmark with skin tones held to 0.6 degrees of drift. Heavy added damage resists: in our stain-heavy test, 83.2% of blotch damage survived restoration. And missing faces past roughly 30% are not restorable by anyone honest, AI or otherwise. Knowing which of the three you are holding turns "will this work?" into a question you can answer yourself in ten seconds — and if your photo sits in the first category, you are going to be pleased.

The fastest way to find out which category your photo is in is to run it. PhotoUnfade gives you three free full-resolution restorations — no account, no credit card — and we would rather you test our honesty on your own photos than take this article's word for it.