What AI photo restoration can actually fix — and what it can't

There's a version of AI photo restoration that gets marketed like magic: feed in a blurry, damaged old photo, get back something crisp and clear. It sounds good. It's also not quite true.

AI restoration tools are genuinely useful, but they work on a specific category of problems. Understanding which problems those are will save you frustration and help you get better results when the tool is actually a good fit.

This article covers what types of photo damage AI handles well, where it runs into walls, and what you can do before you upload to improve your chances of a good result.

What kind of damage does AI actually fix well?

Most old photos suffer from a predictable set of problems. Film degrades over time. Paper yellows. Physical prints get scratched or folded. Scanners pick up dust. Exposure metering from older cameras was hit or miss. AI restoration was essentially built around these exact problems, and it handles most of them reasonably well.

Fading and yellowing. This is arguably where AI restoration does its best work. A photo that's lost contrast over decades, or one that's taken on that orange-yellow cast from acid in the paper, can often be brought back to something close to its original tonal balance. The AI looks at the remaining color information and infers what the original values should have been.

Film grain and digital noise. Older film stocks, especially high-ISO or pushed film, produce visible grain. Early digital cameras produced noisy images at anything above ISO 400. AI smooths this out while trying to preserve edges and detail. Results are usually good on faces and medium-detail areas. Very fine textures (fabric weave, individual hairs) can sometimes get over-smoothed.

Light scratches and dust. Small surface scratches on a print or negative, and dust artifacts from scanning, fall within the AI's damage-repair capabilities. It looks at surrounding pixels to fill in the damaged area. For small, thin scratches against a relatively uniform background, this works well.

Slight blur from camera movement or focus error. A photo that's slightly out of focus, or blurred by a small amount of camera shake, can be sharpened to some degree. The AI identifies edges and brings them into higher contrast.

Low resolution. An old print scanned at low DPI, or a small print that gets cropped heavily, produces a photo with very few pixels. AI upscaling adds pixels by generating plausible detail. It doesn't recover what was lost. But the result often looks significantly better than a simple bicubic upscale.

Where AI restoration hits a wall

Here's the honest part. AI photo restoration can't create information that was never captured in the first place.

Heavy blur doesn't become sharp. If a photo is very blurry, whether from significant camera shake, a badly out-of-focus lens, or heavy subject motion, the AI doesn't have enough edge information to work with. It can improve things marginally. It cannot reconstruct a sharp image from one that has almost no sharp information in it.

A face that's 12 pixels wide doesn't gain recognizable features. This is one of the most common disappointments. You have a group photo from the 1950s, and there are people in the background whose faces are tiny. You restore the photo, and those faces still aren't recognizable. The AI generates a plausible face from those pixels, but the features it generates are invented, not recovered. The person looking back at you may not resemble the original person at all.

Large missing areas can't be accurately filled. A photo that's been torn, or one where mold or water damage has destroyed large sections of the image, is beyond what restoration can reliably fix. The AI will attempt to fill in the missing area by guessing from context, but the results on large gaps tend to look artificial.

Severe physical damage compounds the problem. A photo that's been folded many times, heavily water-damaged, or stored in poor conditions for a very long time often has multiple types of damage overlapping. The AI handles one or two types of damage gracefully. When you stack heavy yellowing, severe grain, multiple deep scratches, and partial tears in the same photo, the result gets less predictable.

Color information that was never there can't be added. The tool restores degraded color. It doesn't add color to black-and-white photos. If you have a black-and-white photo that you want colorized, that's a different process entirely.

How to get better results before you upload

The quality of what comes out is meaningfully affected by the quality of what goes in. Here are a few things worth doing before you upload a photo.

Scan at higher resolution than you think you need. If you're digitizing physical prints, scan them at 600 DPI minimum. 1200 DPI for smaller prints or negatives. More pixels means the AI has more information to work with. A photo scanned at 150 DPI and upscaled gives the AI very little to work with. A photo scanned at 600 DPI gives it much more.

Use PNG for your uploads, not JPEG. JPEG compression introduces artifacts that can interfere with restoration. If you're scanning a print, save it as a TIFF or PNG from your scanner, then upload the PNG. If you only have a JPEG, it will still work, but you might see the AI trying to fix JPEG artifacts rather than the actual photo damage.

Straighten and crop before uploading. The AI focuses on repairing the photo content. If the photo is skewed or surrounded by a lot of scanner border, crop that out first. You want the AI spending its effort on the actual image.

Clean the scan if possible. If you're scanning physical prints, wipe the scanner glass before scanning. Dust on the glass shows up in the scan as uniform noise across the whole image. It's easier to clean the glass than to have the AI deal with widespread dust artifacts.

Don't over-expect on very small photos. A wallet-size print from the 1960s, scanned even at 1200 DPI, doesn't have that many pixels. The AI can improve it, but there are real limits to what can be done with a small amount of source data.

How the processing works

When you upload a photo, it gets sent to a server for processing. Nothing happens in your browser. The AI model runs on remote infrastructure, which is why processing takes a moment rather than being instant. Once it's done, the restored image comes back and you can download it.

This also means the processing quality is consistent regardless of your device. The same restoration model runs every time, on the same hardware. You aren't limited by your phone's processor or your laptop's memory.

The tool keeps a history of your restored photos, so you can come back and download them again if needed without re-processing.

Practical expectations

Here's an opinionated take: most people are better off thinking of AI photo restoration as triage rather than transformation. You have a box of old family photos, many of them damaged. You run them all through the tool for five cents each. Some will come back looking genuinely better. A few will show only marginal improvement. One or two might not improve much at all. That's fine. At five cents per image, even a modest improvement on most of the batch is worth it.

What you shouldn't expect is that a severely damaged photo will emerge looking like it was taken yesterday. That's not what the technology does. What it does do, in the right circumstances, is take a photo that's difficult to look at or nearly unusable and turn it into something you can actually show people. For family history and personal archives, that's often enough.

The format support is worth noting: JPG, PNG, and WebP files work. No SVG, no raw camera files, no TIFF directly. If your photos are in a different format, convert them first.

AI 老照片修复

使用 AI 智能修复老旧、破损的照片

→

Z.Tools·z.tools

AI Old Photo Restoration · Z.Tools

Restore old and damaged photos with AI