What an old photo gives up when AI restores it

The restored copy has just come back. You open it next to the original, and the first thing you notice is how much cleaner it is: the scratch across one corner is gone, the yellow cast has faded and everything looks clearer. You should be pleased. Instead you keep looking, with a feeling that's hard to name, as if the photo had gained something and lost something at the same time.

That feeling isn't you being fussy. It has a mechanical cause, and the cause is built into how a restoration run works.

Three fixes in one press, none of them optional

A run of Restore old photos doesn't do one thing. It does three: it sharpens the photo and brings out detail, corrects color that has shifted with age and removes scratches and creases. Today the tool runs all three on every photo, and there's no switch to leave one out. You upload the photo, and all three happen together.

That keeps things simple, and for many old photos the first result is already a better copy. It also has a consequence that rarely gets mentioned: every photo gets all three fixes, even a photo that only needed one. And each fix, alongside what it repairs, takes something away.

The principle worth remembering for the rest of this article fits in one line: what a photo gives up runs opposite to how damaged it is. On a print that's torn, stained and faded beyond recognition, the gain is so large that the loss barely registers. On a print that's nearly intact and just looks old, the gain is small and the loss shows. Same run, same three fixes, and the value of the result can be completely different from one photo to the next.

What restoration can repair, and how to scan a print well before you start, is covered in what AI photo restoration can and can't fix. This article is about the other side of the ledger.

Faces that come back smoother than the person was

The loss that hurts most tends to land exactly where you care most.

Open both versions and look closely at the face, and specifically at what makes it that person's face rather than a face in general: the fine lines beside the eyes, the creases beside the mouth, a small scar on the forehead, the texture of the skin on the cheeks, a beard that grew in unevenly.

Cleanup and sharpening steps tend to treat fine texture as noise, and skin is full of texture that looks a lot like noise: pores, spots, film grain sitting on a cheek. Clear those away and the skin can come back smoother than the person's skin ever was. That's the "plastic" look people complain about.

How far can a machine go when a face has lost most of its information? The researchers who built PULSE, a model for upsampling tiny, pixelated photos of faces (its code accompanies a CVPR 2020 paper), answered that directly in their project notes, after people worried it could reveal faces that had been blurred out: "PULSE makes imaginary faces of people who do not exist, which should not be confused for real people." The tool, they add, "will not help identify or reconstruct the original image." PULSE is an extreme research case, built for faces only a few pixels wide, and it isn't part of Restore old photos. But it shows the principle at its limit: the less information a face has left, the more of what you see afterward is a guess.

That explains a sentence many people say and then feel silly about: it looks like her, but it isn't Grandma. The feeling rarely comes from imagination. It comes from the small personal marks being cleaned away, leaving a face with the right proportions and the right features but without the details that let you recognize someone at a glance.

If the new version has lost what made a face that person's face, that's a good reason to keep both versions, not a reason to run it again. Another run applies the same three fixes in the same way.

Detail that was never there

The second loss is harder to spot, because it leaves no gap for you to notice.

Every old print has areas so faded that nothing is left inside them: a sleeve that has turned into a gray patch, the pattern of a curtain that has dissolved, lettering on a sign behind someone reduced to a few marks. In those areas there's nothing left to recover, so whatever detail appears in the restored copy was produced by the tool. It may well look plausible. It isn't a record of what was there.

Professional conservators hold themselves to the opposite standard. The American Institute for Conservation's Code of Ethics and Guidelines for Practice puts it this way in Guideline 23 (Compensation for Loss): "Any intervention to compensate for loss should be documented in treatment records and reports and should be detectable by common examination methods. Such compensation should be reversible and should not falsely modify the known aesthetic, conceptual, and physical characteristics of the cultural property, especially by removing or obscuring original material." Detail filled in by software meets none of those three tests: it isn't documented, nothing marks where it is and it can't be undone on that file. That's why the untouched scan matters. It is your record.

This is also different from editing a photo with a text prompt. There, you ask for a change and the whole picture is redrawn to match it. Here, the tool fills gaps you never asked about.

Check the blurriest part of the old print first

From this comes a habit that sounds backward but holds up: the areas that were blurriest in the old print are the areas to trust least in the new one. People usually do the opposite and trust whatever looks crisp, when the crispest areas may be exactly where the tool had to invent the most.

To be fair, suspicious doesn't mean wrong. Often the result matches what was really there, especially for things that follow a pattern, like the edge of a collar or the folds of fabric. The problem is that you can't tell whether this particular guess was right, so don't treat those areas as evidence of anything that matters, such as who stood where or what a sign said.

Grain and age: the look that gets "fixed"

This loss is the one people miss most and find hardest to name.

An old film photo isn't clean, and that's part of what it is. There's fine grain spread across the image. Edges are soft, with none of the hard precision of a new phone photo. Time has laid a warm tone over the paper. Together they make up the look that tells you at a glance the photo is old.

People who work with film take that grain seriously. In a paper on film grain for the AV1 video codec, Andrey Norkin of Netflix and Neil Birkbeck of Google write: "Film grain is abundant in TV and movie content. It is often part of the creative intent and needs to be preserved while encoding." They also explain where it comes from: "The grain is inherent in analog motion picture film due to the process of exposure and development of silver-halide crystals dispersed in photographic emulsion". Their paper describes a way to model the grain and synthesize it again, rather than let compression wipe it out.

The paper is about movies, not family albums. The connection is our own: the same kind of grain is in your old prints, and a sharpening step can read it as noise and clear it away. The color correction can do something similar to the warm tone of age, pulling it back toward neutral. The run has no way of knowing which grain you'd keep or which warm tone is simply the color of the years, and nothing on the screen lets you tell it.

So the practical point is this: if what you love about a photo is its age, make the restored copy a second copy, not a replacement.

What archivists do instead

US federal agencies that digitize historical collections have written this lesson down. The Federal Agencies Digital Guidelines Initiative (FADGI) publishes technical guidelines for digitizing cultural heritage materials, and its third edition, from May 2023, starts from the same place: "Digitized material should document the appearance of the original at the time of capture, not what it may once have looked like if restored to its original condition."

It says traditional adjustments such as contrast and brightness may be allowed, then lists what to avoid: "Certain digital enhancements should be avoided such as: Colorization, color restoration, digital infill, excessive cropping of the image area, over sharpening, de-saturation, blurring, removal of blemishes in the original such as stains, dust, or scratches, and other transformative image processing." Read that list next to the three fixes above. Color restoration, digital infill, over sharpening and removing scratches line up closely with what a restoration run does.

That doesn't mean anyone in government is telling you not to restore family photos. These guidelines govern archival copies of cultural heritage, not what a family does with its own pictures. What they explain is why an unedited scan is worth keeping. The guidelines' section on prints and photographs draws a line between two kinds of file. On master images: "Color, tone enhancement or restoration is not allowed on master images." On the copies made for people to use: "Adjustments to correct or enhance the image may be made to access versions, and noted as such in embedded metadata and file naming." In other words, one master with no enhancement or restoration, and adjusted copies for use, noted as adjusted. That's the same habit this article recommends for the shoebox in your closet.

Keep the unrestored scan as your reference

Once a photo has been through three fixes you can't separate, the only thing that lets you compare later is the version nobody touched. Without it, every question like was this really how it looked? or how deep were those lines around her eyes? has nothing to be checked against. The new version becomes the only version, and nobody can check a single version against anything.

Turning this into a habit takes three small steps:

• Keep the unedited scan somewhere you know you'll find it, not mixed in with all the other files you've been testing.
• Name it so that six months from now it's obvious which file was never touched. Pick names that explain themselves without you having to remember anything.
• Never overwrite. The restored copy is a new file, not an update to the old one.

The before-and-after view on the result screen is handy, but it can't replace that habit. It only works while your browser still holds the original. Come back after a page reload and the editor tells you so: "The original photo isn't kept in your browser after a page reload, so only the result is shown here." The comparison you need is the one you'll make months from now.

Don't treat the result screen as storage either. The restored copy in My images stays there only for a limited time (Vipix's Privacy Policy gives the periods), so download both versions and store them properly yourself. Whether a restored download comes back with a small vipix.vn mark depends on the tool and on whether you were signed in when you ran it, so look at the corners before you print it. For how to scan and back up the originals in the first place, see the guide linked at the top of this article.

Worth it, or better left alone?

Back to the principle from the beginning, because this is where it answers the question for you.

Badly damaged: run it

Torn, stained all over, faded until you can't tell who's in it: there's hardly anything to weigh. The gain is so large that the loss disappears. You're trading a photo you can barely look at for one you can. Run it.

In between: weigh the three losses

The photos in the middle are the ones worth thinking about, and the three sections above are the scale. How much do you care about the old look of this print? Is there a face in it that you remember in detail? Do the faded areas fall on something important?

Nearly intact: leaving it alone is a real choice

Paper in good shape, color still fine, just a little old: here the most common right answer is to leave it alone. To be clear about it: leaving a photo alone is a valid choice, not a failure. A photo doesn't have to go through software to be worth keeping, and not everything that can be done should be.

When another tool fits better

A few cases belong somewhere else:

• Torn prints with missing pieces, and photos of loved ones who have died. These are the photos most often brought in for restoration, and how far to go with them deserves its own discussion: which kinds of damage can still be repaired.
• A photo in good shape that you only want bigger, say for an 8×10-inch print. That's a job for Upscale image, not this tool.
• A black-and-white photo you want in color. Adding color is the job of a separate tool, Colorize black-and-white photos, run on its own. Before you do, it's worth reading what your black-and-white photo really is, because not every gray photo suits colorizing.

The yearbook photo, the album from a 1970s wedding, a stack of Polaroids, the photos chosen for a memorial display, the album that comes out at every family reunion: each one deserves the same question before you press the button.

Frequently asked questions

Why does my restored photo look fake?

Usually because fine texture and film grain were smoothed away and some faded areas were filled in with guessed detail. The result looks cleaner but also newer than the photo really is, and faces can come out smoother than the person's real skin.

Can AI restoration change someone's face?

It can. Sharpening and cleanup can smooth away the small marks that make a face recognizable, and where a face has lost a lot of information, part of what you see afterward is a reconstruction rather than a record. Keep the original next to the restored version.

Can I turn off the sharpening or the color correction?

Not today. The tool runs all three fixes (sharpening, color correction and scratch removal) on every photo, and there's no switch to leave one out.

Should I keep the original scan after restoring it?

Yes. For prints and photographs, federal digitization guidelines allow no color or tone enhancement or restoration on master images and let adjustments be made to access versions, and professional conservators expect any compensation for loss to be documented and reversible. Your unedited scan plays that role for your family photos.

Does restoring a black-and-white photo also colorize it?

Colorizing isn't one of its three fixes. Turning a black-and-white photo into a color one is the job of a separate tool, Colorize black-and-white photos, run as its own step. Because one of the three fixes does adjust color, look at a black-and-white result for any tint it didn't have before.

Start with the photo in the worst shape

Start with the most damaged photo in the box, because that's where restoration gives the most and takes the least, and keep the untouched scan next to the result. When a run of Restore old photos uses credits, the cost appears on the tool page before you confirm.