How to Fix Blurry Faces in Photos (and Keep the Person)

Q: Can AI fix a blurry face in a photo?

It can make almost any blurry face look sharp. Face restoration models do this by generating detail from what faces generally look like, so the sharper result is not guaranteed to be the same person. Mild softness usually comes back faithfully; a tiny or heavily smeared face is mostly reconstruction.

Q: Why does the fixed face look like a different person?

The software filled in detail the photo never recorded. The researchers behind CodeFormer describe a trade-off between quality and fidelity: settings that give a cleaner face also stay less close to the input. Use the setting that stays closest to the original and accept a softer result.

Q: Should I fix blurry faces in a group photo?

Yes, but check every face separately. In a group shot each face covers far fewer pixels than in a portrait, so each one is reconstructed more heavily, and one person can come back well while the person beside them does not.

Q: Can I use an AI-enhanced face as evidence or for identification?

No. In Washington v. Puloka a court refused to admit AI-enhanced video; a forensic video analyst had testified that the tool created false image detail. Keep the original file for anything that has to show what someone actually looked like.

Q: Does 43frames have a face restoration mode?

No dedicated one. Photo restoration repairs the whole image and returns a sharper, high-resolution copy, which often makes soft faces clearer, and photo upscaling has Advanced sliders for staying true to the original and adding detail. Restoration costs 8 credits and upscaling 2 credits per image.

Almost any blurry face can be made to look sharp now. The harder question is whether it still looks like the person. Face restoration tools don't recover the face your camera missed; they generate a plausible one from what faces in general look like. So the job isn't finding the tool that "unblurs" best. It's choosing how much you let the software invent, then checking the result against people who knew the face.

If your whole photo is soft, not just a face, start with how to fix blurry photos, which covers the three kinds of blur. This post is about the face, because that's where invention matters most.

Why faces get special treatment

Generic sharpening raises edge contrast in the pixels you have. Face restoration models do something different: they were trained on large numbers of faces and use that as a prior, a built-in expectation of what an eye, a nostril or a hairline should look like.

The authors of CodeFormer, an open face-restoration model built into many online tools, are unusually candid about what that means. Their paper opens with the line "Blind face restoration is a highly ill-posed problem" — many sharp faces are consistent with the same blurry input. Their stated goal is "the discovery of natural faces that closely approximate the target faces." Approximate is the honest word.

That's why these tools beat a sharpen slider on faces, and also why they can hand you a stranger.

The one control that matters: fidelity vs. quality

Face tools handle the same trade-off under different names, or hide it.

CodeFormer calls it the fidelity weight, w, from 0 to 1. Its README puts it plainly: "smaller w tends to produce a higher-quality result, while larger w yields a higher-fidelity result." Quality here means a clean, natural-looking face. Fidelity means staying close to the input. You can't max both. The project page shows the same face sliding between the two.

Topaz sells the same dial as "dual mode face recover technology (creative and realistic modes)" on its unblur page. Many one-click tools don't show it at all and pick a setting for you.

For a face you care about, push toward fidelity. A result that's slightly soft but unmistakably your grandmother beats a crisp portrait of someone who resembles her. Start at the most faithful setting a tool offers, and only move toward "quality" or "creative" if the result is still unusable.

Sharp is not the same as right

A sharp result tells you the model was confident. It doesn't tell you it was correct. The more the input was blurred, the more of that sharpness came from the model rather than your photo.

Check how much face you actually have

Before running anything, look at how many pixels the face covers. This predicts the outcome better than the kind of blur.

CodeFormer's own instructions work on faces cropped and aligned to 512×512 pixels. A face that's 60 pixels wide in a group photo has to be enlarged more than eight times to reach that working size. Most of what comes back is reconstruction, not recovery.

Three practical consequences:

Portraits restore best. A face that fills the frame has real detail for the model to hold on to.
Group photos restore unevenly. The person in front and the person in the back row are different problems. Check every face on its own.
Rescanning beats restoring. If the source is a print, a higher-resolution scan gives the model more real pixels to work with. Our scanning guide covers resolution and handling.

A workflow that keeps the person

Start from the best copy. The original file or a fresh scan, never a screenshot or a messaging-app download. Keep that file untouched.
Repair damage first. Scratches and dust across a face confuse face models into odd shapes. The order of operations is in how to restore old photos.
Run one pass at the most faithful setting. Repeated passes compound the model's guesses.
Compare against other photos of the same person. Put the result beside two or three photos from around the same age, not just the blurry original.
Show it to someone who knew them. Family will spot a wrong smile in a second. You probably won't.
Save the result beside the original, not over it. Label it as enhanced, especially if you share it with relatives or in a family archive.

What to check on the face

Because the model fills gaps from what faces generally look like, check the features that make someone look like themselves rather than like a typical face:

Eyes: shape, the distance between them, which one sits higher, and eye color if the photo shows it.
Teeth and smile: a real crooked tooth or an asymmetric grin is identity, not damage.
Marks: moles, scars, freckles and wrinkles that were in the original should still be there.
Glasses: frame shape and thickness, and whether the lenses now look empty or doubled.
Hairline and ears: CodeFormer's README warns that whole-image processing involves "face-background fusion that may damage hair texture on the boundary." Look at where hair meets face and background.
Age: a 70-year-old shouldn't come back with 40-year-old skin.

If two or more of these are wrong, go back to a more faithful setting or keep the original.

When not to fix a face at all

Some photos need to show what a person actually looked like, not what a model thinks they looked like.

In Washington v. Puloka, a court refused to admit AI-enhanced video. The ABA's summary of the case reports that a certified forensic video analyst testified the enhancement "created false image detail," changing the shape and color of objects. The court held that the original video was the best evidence of what it showed.

The same logic applies at home. For anything evidentiary, historical or identifying — insurance claims, genealogy records where a face is the evidence, anything official — keep and share the original. Use an enhanced version only as a clearly labeled companion.

Fixing a blurry face in 43frames

43frames doesn't have a dedicated face-restoration mode, and we won't pretend otherwise. What it has are two whole-image tools, and one of them exposes the fidelity dial directly.

Photo upscaling enlarges the image 2× or 4×. Under Advanced there are two sliders:

Stay true to original (default 60%): "Higher values preserve the original more closely."
Add detail (default 35%): "Higher values may invent details that were not in the original."

That's the same trade-off CodeFormer describes, in plain words. For a face you care about, raise Stay true to original, lower Add detail, choose 2× before 4×, and compare. Our upscaling guide explains how to work out the size you actually need.

Photo restoration repairs damage across the whole image and returns a sharper, high-resolution copy, keeping the original colors (with an optional Add color setting). Use it first if the face is damaged as well as soft, and compare the face with your scan.

Restoration costs 8 credits and upscaling 2 credits per image, and new accounts get 10 starter credits once. Both accept JPEG, PNG or WebP up to 10 MB. If you're weighing a phone app built around face enhancement, our Remini alternative comparison covers the trade-offs.

FAQ

Can AI fix a blurry face in a photo? It can make almost any blurry face look sharp, by generating detail from what faces generally look like. Mild softness usually comes back faithfully; a tiny or heavily smeared face is mostly reconstruction.

Why does the fixed face look like a different person? The software filled in detail the photo never recorded. Settings that give a cleaner face stay less close to the input, so use the most faithful setting and accept a softer result.

Should I fix blurry faces in a group photo? Yes, but check every face separately. Small faces are reconstructed more heavily, so results vary from person to person in the same image.

Can I use an AI-enhanced face as evidence or for identification? No. A Washington court refused to admit AI-enhanced video, and a forensic analyst testified the tool created false image detail. Keep the original.

Does 43frames have a face restoration mode? No dedicated one. Restoration and upscaling work on the whole image; restoration sharpens the whole image as it repairs it, and upscaling's Advanced sliders let you favor the original over added detail. Restoration costs 8 credits and upscaling 2 credits per image.