← All guides · 2026-11-09 · PhotoUnfade team

How Sharpening Actually Works in Photo Restoration (And When It Goes Too Far)

If you have ever run an "enhance" filter on an old family photo and thought "that looks sharp — but not quite like my grandmother," you have found the central problem of sharpening in photo restoration. The honest answer up front: sharpening does not recover detail that was lost. It increases contrast along edges that already exist, which makes a soft photo look crisper without adding a single new piece of information. Done conservatively it is the finishing touch of a faithful restoration; done aggressively it manufactures detail the camera never recorded. This guide explains how sharpening actually works, how to measure it honestly, the exact settings our pipeline ships, and the warning signs that a tool has crossed from sharpening into invention.

What does sharpening actually do to a photo?

Sharpening makes existing edges look stronger — nothing more. Nearly all digital sharpening is some form of the unsharp mask (USM), a technique older than digital photography itself: make a blurred copy of the image, subtract it from the original to isolate the edges, then add a scaled amount of that edge layer back. Pixels on the bright side of an edge get brighter, pixels on the dark side get darker, and the eye reads the increased local contrast as "sharper."

Two consequences follow from that definition. First, sharpening can only work with edge information the photo already contains — a face recorded out of focus has no fine edge data to amplify, so no amount of unsharp masking will bring back eyelashes the lens never captured. Second, the technique has a built-in side effect: push the amount too far and the bright-and-dark overshoot at each edge becomes visible as a glowing outline, called a halo.

That is why "how sharp can this photo get?" is really two questions: how much real edge contrast is recoverable, and how much amplification can be applied before the cure becomes visible. Everything below is about answering those two questions with measurements instead of eyeballs.

How is sharpening measured honestly?

With two numbers: acutance and overshoot. Acutance — in plain terms, the measured contrast across edges — is the "sharpness" figure. Overshoot is the share of edge pixels pushed past their natural value, which is what produces halos. Good sharpening raises acutance while keeping overshoot low; bad sharpening raises both, and generative "enhancement" sidesteps the trade entirely by inventing edges outright.

There is a methodological trap here, and we fell into it before catching it. AI restoration models are stochastic — run the same photo twice and the output shifts measurably (we have recorded the same image coming back with a warm-cast correction of +19.2 on one run and −1.7 on the next). If you compare "before sharpening" from one run against "after sharpening" from a different run, the AI's randomness drowns out the sharpening signal. We once concluded our sharpening layer "was not working" for exactly this reason — the two numbers came from two different AI outputs. The rule we now enforce: sharpening is measured as a paired test, before versus after on the same AI output, never across runs.

With that rule in place, calibration becomes cheap and honest. We apply a grid of candidate unsharp-mask settings to one restored image — no new AI calls, so the exercise costs nothing — and score each setting for acutance gain against overshoot. The shipped setting sits where sharpness is meaningfully up while halos stay below the level already present in the aged original scan.

How much sharpening does PhotoUnfade apply?

Both restoration modes use the same conservative unsharp mask — sigma 1.2 with edge-amount parameters 0.30 and 0.60 — applied after color correction, never before. The measured results, paired before-and-after on identical outputs: our subtle mode raises acutance by 29% over the original scan, and our full mode by 77%. Overshoot comes in at 8.6% of edge pixels in subtle and 7.4% in full — versus 26.1% in the unprocessed original scan itself.

That last comparison surprises people, so it is worth stating plainly: old photos already carry edge artifacts from the original lens, film grain and the scan. Our sharpened outputs show fewer halos than the aged original because the restoration first removes the damage — scratches, haze, color casts — and only then enhances the edges that remain. The order matters for a specific reason: sharpen first and you amplify the edges of the warm yellow cast along with everything else, locking the cast into the picture.

One safety detail matters more than any parameter: if the sharpening step fails for any reason, our pipeline falls back to the color-corrected image rather than failing the job. A paid restoration must never fail because of a finishing step.

Why does "sharper" sometimes mean "invented"?

Because there are two completely different ways to make a photo look sharper, and only one of them is restoration. The unsharp mask amplifies real edges. Generative AI enhancement — the kind behind the dramatic before-and-after demos of the big consumer apps — repaints the image at high frequency, synthesizing skin texture, hair strands and fabric weave from statistical averages of millions of other photos.

We measured the difference directly. On the same sharp original, the popular editor we tested registered +101% measured sharpness and eight times as many strong-edge pixels as the original contained. Those edges did not come from the photo; they came from the model's idea of what photos generally look like. On a face, that is precisely the mechanism that makes a restored grandmother look like a stranger — the eye reads the new texture as detail, and only later notices the person has changed. The face-level measurements, including skin-tone drift, are in our pixel-by-pixel accuracy test, and the "why does my mother look different" question is covered in our explainer on generative face enhancement.

Our design decision follows from that measurement: the AI model is never asked to invent sharpness. The generative step repairs damage; the unsharp mask — a deterministic, parameter-fixed, non-AI operation — is the only thing allowed to touch edge contrast. That separation is what makes the numbers in the table above stable and auditable, and it is why our sharpening figures are modest compared with a +101% marketing demo.

When does sharpening go too far?

Sharpening has gone too far when you can see it — and we know what that looks like because our first attempt did it. The initial parameter guess for our sharpening layer, with edge-amount values roughly three times the final settings, measured +180% and +460% acutance in calibration. On screen, skin turned gritty, hairline halos glowed, and the photo had the "crunchy" look of over-processed images. We caught it in offline grid calibration before it ever shipped — which is exactly what the calibration step is for.

You can run the same inspection on any tool's output at 100% zoom:

1. White or dark outlines tracing high-contrast edges — jawlines, glasses, rooflines. This is the halo signature of overshoot.
2. Gritty or sandy skin on faces that were smooth in the original — amplified noise being mistaken for detail.
3. Doubled edges, where a single contour renders as two parallel lines.
4. Suspiciously uniform texture in hair, fabric or foliage — the signature of generative synthesis rather than sharpening, and the hardest one to unsee once you notice it.

Damage type matters as well. Fine surface scratches and haze respond beautifully to the restore-then-sharpen sequence; deep emulsion scratches and thick overlaid damage are treated by the AI as picture content and will survive any amount of sharpening. The scratch-specific workflow is in our guide to fixing scratched photos.

Can sharpening rescue a blurry photo?

Only partly, and only for some kinds of blur. Soft focus and mild motion blur leave real edge information spread across extra pixels, and careful sharpening can pull some of it back — that is the case where our measured +29% to +77% gains apply. True missed focus and severe motion blur contain no edge information to amplify, and sharpening them only sharpens the blur. The full four-type blur diagnosis, including when a blurry photo should honestly stay blurry, is in our blur repair guide.

This is also where we owe readers our own correction, because it illustrates the measurement trap better than any theory. In our first accuracy test we compared a 531×384 input against a 1024×736 restored output and "measured" a 48% sharpness loss — alarming, and wrong. Comparing different resolutions measures resampling, not sharpness. Re-tested at matched dimensions, the true change was between −4% and +10% depending on mode, before the sharpening layer was added. We keep that mistake in the published test because it is the most common way sharpening numbers deceive: whenever a tool claims dramatic sharpness changes, the first question to ask is "compared at what resolutions?"

The bottom line

Good restoration sharpening is boring by design: a fixed, conservative unsharp mask that raises measured edge contrast by 29% in subtle mode and 77% in full mode, with fewer halo artifacts than the aged original scan already carries — plus a hard rule that the AI is never allowed to invent sharpness by repainting texture. If a tool's output looks dramatically sharper than the original, check the edges at 100% zoom before you celebrate; some of that detail may never have been in your family's photo at all.

The surest test is your own photo. PhotoUnfade offers three free full-resolution restorations — no account, no credit card — so you can zoom to 100% and check our edges against this guide yourself. If you are planning a larger archive project, the per-photo arithmetic across all routes is in our restoration cost guide.