

Blurry, soft faces
Eyes, teeth and hair come back into focus, and Grandma still looks like Grandma. Most people upload for this.
Upload an old or blurry photo, and GFPGAN restores the faces and brings back natural color. It's free, there's no watermark, and you don't need an account.


Illustrative AI restorations made with an image-generation tool; details may be reconstructed.


Torn military portrait — AI reconstruction example


Eyes, teeth and hair come back into focus, and Grandma still looks like Grandma. Most people upload for this.


Yellowed prints get their original tones back. You can also colorize black-and-white shots.


Dust spots, creases, surface scratches and small rips get cleaned up at the same time as the faces. Big tears are harder. See the section on what works below.


Tiny scans, early phone pictures and images that were saved and re-shared until they got blocky.
Download your restored photo free, with no watermark. For large prints, you can download in 2K or 4K.
Scan the print, or take a straight-on picture of it with your phone. Then drop the file into the box above.
The AI finds every face in the picture and rebuilds the detail that blur and age wiped out.
Drag the slider to check the result against the original. Download it free, or choose 2K or 4K.
GFPGAN is an open-source face restoration model released by Tencent ARC Lab in 2021. It learned from a huge number of faces, so it has a good sense of what eyes, skin and hair should look like, even when a picture is badly blurred.
Running it yourself takes work, though. You need Python, the model files and a decent graphics card, and the install doesn't always go smoothly. gfpgan.io runs it online for you. Open the page, upload a photo and get the result in your browser, on a computer or a phone.
We're an independent site. We aren't part of Tencent, and Tencent doesn't endorse us.

We'd rather tell you now than after you upload.
If a face is mostly gone, no AI photo restoration tool can get it back. It can only invent a new face. Be careful with apps that promise otherwise.


Surface scratches can be reduced, but the small faces still have limited detail.

Damage and low detail obscure several faces. This scan does not provide enough evidence to verify a faithful reconstruction. We show the original rather than presenting an invented face as a successful repair. A better scan or another photo of the people may help.
Tips for a better result: Scan prints at 600 DPI if you can. If you use a phone, shoot in daylight and hold it straight above the photo. Take the picture out of its frame first, since the glass causes glare. Then crop out the table and the album page.



Clean up faces from old albums before adding them to a family tree or sending them around the family group chat.


Sometimes the only decent photo of someone is small, old or out of focus. A quick restore can make it clear enough for a service program or a framed print on the table.

A restored wedding photo printed in 4K is a good anniversary gift for parents or grandparents.


Paid plans include commercial use, so you can restore photos for clients. Batch processing and an API are on the way.
CodeFormer is another popular face restoration model. Both fix blurry faces. If you run CodeFormer yourself, you have to install it and adjust a "fidelity" setting by hand. Set it too low and you can get a sharp face that doesn't look like the person anymore.
Here, there's nothing to adjust. You upload, and the result aims to keep the person's real features.
| gfpgan.io | Running CodeFormer yourself | |
|---|---|---|
| Setup | None, works in your browser | Python, model files, a GPU |
| Settings to tune | None | Fidelity slider |
| Colorize photos | Yes | No |
| 2K / 4K output | Yes | Depends on your hardware |
| Works on a phone | Yes | No |
Yes. You get 4 free credits a day, which covers 2 photos, and downloads are free with no watermark. Paid plans add 2K and 4K downloads, higher-quality output and commercial use.
No. You can upload and restore a photo without creating an account.
No. Restored photos download clean, including free ones.
It shouldn't. The face restoration model rebuilds detail from what's already in the photo, so the face stays true to the original. The exception is a face that's torn away or faded almost completely. Then there isn't enough left to work from.
Light and moderate ones, yes. Heavier damage can often be repaired too, but only if enough of the face is still visible.
Yes. It adds natural color to black-and-white photos and brings faded color prints back toward how they looked when they were taken.
Nobody else, unless you share a link. Your photos and results are never published or listed anywhere, and you can delete them at any time. Free photos are deleted after 30 days; on a paid plan they stay until you delete them.
Yes, on a paid plan. Free restorations are for personal use.
JPG, PNG and other common image formats, up to 30 MB each.
Yes. Open this site in your phone's browser, pick a picture from your camera roll and download the result.
No. GFPGAN is an open-source model from Tencent ARC Lab. We're an independent site that runs it online, so you don't have to install anything.
Not yet. Batch processing and an API are coming for studios. For now, photos are restored one at a time.
Start with the one you keep meaning to fix. It's free, with no watermark and no account needed.