How to Restore Old Photos with AI: An Honest Guide

Somewhere in most households there is a shoebox, a photo album with sticky plastic sheets, or a folder of scans nobody has opened in years. The photos in it are the ones that matter most and look the worst — small, faded, soft, and taken with equipment that was never meant to survive a large screen.

"AI photo restoration" is a phrase that gets used loosely, so this guide separates it into parts: what genuinely works today, what does not yet, and how to get the most out of the part that does. If you are here to be sold a miracle, this will be a disappointing read. If you want your grandmother's face to be legible again, keep going.

"Restoration" is really four different problems

The word covers several distinct kinds of damage, and they are not solved by the same technology. Sorting your photo into the right bucket saves a lot of frustration.

1. Not enough detail. The print is small, the negative was grainy, or the scan was made on a cheap flatbed years ago. Faces turn to mush when you enlarge them. This is a resolution problem.
2. Physical damage. Creases, scratches, tape marks, torn corners, water stains, mould spots. This is an inpainting problem — something has to invent what was underneath.
3. Colour shift and fading. The magenta cast that old prints drift towards, or the general washed-out grey of a photo that lived on a sunny wall. This is a colour problem.
4. No colour at all. A black-and-white photo you would like to see in colour. This is a colourization problem, and it is closer to informed invention than to repair.

Most real photos are a mix. The good news is that problem one — not enough detail — is both the most common and the one AI handles best today.

What AI enhancement genuinely does for an old photo today

ImageEnhance runs an AI super-resolution pipeline. Instead of stretching the pixels you already have, the model predicts what a sharp version of each soft edge should look like and draws it in. On old photos, that produces four real, verifiable improvements:

• Faces come back into focus. Eyes, hairlines and the edges of glasses are the first things to return. This is the single biggest emotional difference in a restored family photo, and it is exactly what super-resolution is good at.
• Small scans become usable at size. Enhancing at up to 4x the original resolution turns a photo that only worked as a thumbnail into one you can print or put on a wall.
• Light grain and softness ease off. Film grain and the gentle blur of an old lens are partly cleaned up as a side effect of rebuilding detail. It is not what a dedicated denoiser would do, but the improvement on lightly noisy scans is real.
• Compression damage from previous digital lives is reduced. Many "old photos" have been through email attachments, chat apps and Facebook. The plastic, blocky look those leave behind eases considerably.

That is a genuinely meaningful set of results for a photo you had written off, and it costs nothing to see it on your own picture: without an account you get 8 free credits every 24 hours, enough for 2 photos at 2K.

What is not live yet, said plainly

Our homepage marks three tools "coming soon", and we would rather you read it here than discover it after uploading:

• Dedicated denoise — for heavy night-shot grain and severe sensor noise.
• Old photo restoration — the specialist tool that removes scratches, tape marks and stains, and repairs torn areas.
• AI colourization — adding believable colour to black-and-white photographs.

Those are not built yet. What is live today is enhancement and upscaling, which run on the same super-resolution pipeline. So if your photo's main problem is that it is soft and small, you are in the right place now. If its main problem is a scratch straight across someone's face, enhancing will make the scratch sharper too — and honestly, that is worth knowing before you start.

Step by step: getting the best result from an old photo

The quality of what you feed the model matters more than any setting. Most disappointing results are decided before the upload.

1. Find the best available source. A print beats a photo of a print. A negative or slide, if you have one, beats both. Take the extra detour to the original rather than working from a picture someone messaged you.
2. Scan it, or shoot it properly. A flatbed scanner at a high setting is ideal. If you are using a phone, put the print on a flat surface in soft, even daylight — near a window but not in direct sun — hold the phone parallel to the print, and check the preview for glare and for your own shadow. Wipe the print with a dry, soft cloth first; dust scans as detail.
3. Save the scan properly. Export as PNG if your scanner offers it. If it only gives you JPG, save it once at the highest quality and never re-save that file.
4. Crop before you enhance. Trim away album pages, table edges, fingers and white borders. Everything you crop is resolution the model no longer has to spend on things you do not care about.
5. Upload to ImageEnhance. JPG, PNG and WebP are supported, up to 4MB per photo. If your scan is larger than that, export it again at a slightly lower quality setting rather than shrinking its dimensions — dimensions are what you are trying to keep.
6. Choose your output size. 2K is free and takes a few seconds. 4K takes around ten seconds and is the one to choose if you plan to print or frame the result; it requires a subscription, starting at $8 a month on the Starter plan. The pricing page has the details.
7. Judge it at full size, on the details that matter. Drag the before-and-after slider onto a face, not onto the whole frame. That is where you will see whether it worked.

Realistic expectations, photo by photo

A faded colour print from decades ago. Expect a clear win on sharpness and a photo that finally looks respectable at a decent size. The colour cast will still be there — colour correction is a separate job, and one you can do in any basic photo editor afterwards.

A grainy black-and-white portrait. Often the most satisfying case. Detail in the face improves noticeably and light grain settles down. It stays black and white; colourization is not live.

A phone snapshot from an early camera phone. Very good candidate. These are small, soft and compressed — exactly the combination super-resolution was built for.

A print with a crease or a tear across it. Manage your expectations. The undamaged parts improve; the damage stays, and gets sharper along with everything else. This is what the coming restoration tool is for.

A photo that is severely out of focus or almost black. There is very little real information to rebuild from. You will get a cleaner version of a difficult photo rather than a good one. It is free to try, so try it — but do not plan a gift around it.

Once you have a good version, protect it

Digitising a photo is only half the job. The other half is not losing it again.

• Keep the untouched scan as your master file and treat the enhanced version as an export, not a replacement.
• Store both in more than one place — a computer plus a cloud folder or an external drive. One copy is not a backup.
• Name the files with what you know: names, place, roughly when. A folder of numbered scans is a burden for whoever inherits it.
• Do not archive the version that came out of a chat app. Compression is cumulative, and it is easier to keep a good file than to rescue a bad one later.

Frequently asked questions

Can AI really restore an old photo?

It depends what you need restoring. If the problem is that the photo is soft, small or grainy, AI super-resolution helps a great deal and is available today. If the problem is a scratch, a stain or a torn corner, that needs a dedicated restoration tool, which we have marked "coming soon" because it is not built yet. Being clear about which problem you have saves disappointment.

Do I have to sign up to try it on a family photo?

No. Upload a photo and get the result back without an account or a card. Anonymous visitors get 8 free credits every 24 hours, which covers 2 photos at 2K. Signing up adds 30 credits on the spot and raises the daily allowance to 10 credits.

My scan is bigger than the upload limit. What should I do?

The limit is 4MB per photo, and format matters more than you would expect. Re-export the same scan as a JPG at a high — but not maximum — quality setting, or as a WebP; both usually come in comfortably under the limit at the same dimensions. Resist the urge to shrink the image itself, because resolution is precisely what you want the model to work with.

Should I choose 2K or 4K for an old family photo?

2K is free, finishes in a few seconds, and is more than enough for sharing with relatives, phone screens and social posts. Choose 4K if you intend to print the photo, frame it, or display it on a large screen, where the extra resolution genuinely shows. 4K output requires a subscription — see pricing.

Dig out the one photo you always wished was sharper and put it through — the first one is free, and it takes seconds to find out whether it can be saved.