AI Face Restore

AI Face Restore works on the part of a photo people actually look at. There are two paths through it, they produce very different results, and which one you get is a decision you make rather than one made for you.

The two paths

• GFPGAN v1.4 is a face restoration model. It reconstructs eyes, teeth and hair in a face that is only a few dozen pixels across. It is a 340 MB one-time download, cached afterwards, and it only runs once you have allowed it.
• The classical pipeline — an edge-preserving bilateral denoise, a local contrast boost (CLAHE-lite over 8×8 tiles), then an unsharp mask — needs no download and runs instantly. It evens out and sharpens what is already in the file. It invents nothing, which is sometimes exactly what you want.

The caption under the result names the path that ran and how long it took, so you never have to guess which one you are looking at.

Nothing downloads until you allow it

Before any bytes move, the tool shows what the model costs and what your device will do with it, then asks. Allow the download fetches GFPGAN once and caches it; Not now keeps the classical pipeline. The answer is remembered on this device and you can flip it later from the same line of text. This is the same consent flow described in why AI tools download a model once.

Allowing the download is necessary but not sufficient. GFPGAN is only selected on a device the readiness check rates as capable of running it — WebGPU and enough reported memory. A phone or a low-memory laptop runs the classical pipeline instead, even with consent granted, rather than downloading 340 MB to disappoint you. And if the model fails to load or fails mid-run for any reason, the classical pipeline finishes the job instead of the tool dead-ending.

How the model sees your photo

GFPGAN v1.4 is fixed at 512×512 — that is the only input size the network accepts. Everything else follows from that constraint:

1. A landmark detector looks for a face and returns a bounding box, padded outwards to take in hair and chin.
2. That crop is scaled to 512×512 and run through the model.
3. The result is scaled back to the crop's original size and feather-pasted into the source photo, so the restored region blends rather than sitting in a visible rectangle.

One face per run. The detector is configured to find up to four faces, but the restore pass takes the first one it returns. In a group shot the other faces come back exactly as they went in. If no face is found at all, the whole image is treated as one crop and processed that way — still an improvement, just not a face-specific one.

On the CPU path, restoring a face was measured at about 6.7 seconds, on top of the 340 MB download the first time you run it.

Strength

Four settings — Subtle, Moderate (the default), Aggressive, and Vintage repair — control how much of the restored crop is blended back over the original when it is pasted in. Subtle keeps more of the real skin texture; Aggressive commits fully to the reconstruction and can look synthetic on a close-up. Vintage repair is a strong recovery composited a little softer, for scans and prints. The setting applies to both paths.

What "restore" honestly means

GFPGAN does not recover the original pixels. They are gone, and nothing recovers them. It generates a plausible face consistent with what survived, using a prior learned from a very large number of other faces. On a soft or small photo of someone, enough structure usually survives to constrain the result and it looks like them. On a severely degraded photo the model will still return a crisp, confident face — and that face can differ from the person in ways the output gives you no way to detect.

So it is a good tool for a print you want to hang on a wall, and the wrong tool for identification. Keep the original file.

Limits

• One face per run, and only the first face found. Group photos need one pass per face if you crop them yourself.
• The classical path sharpens and evens; it does not rebuild detail that was never captured.
• JPEG, PNG or WebP in, up to 32 MB. PNG out.
• Everything runs in this tab. The photo is never uploaded, which for family photographs is not a small detail.

In the image editor

The tool is also in the image editor's launcher, under AI Tools, as AI Face Restore. It is deliberately not offered in the video editor, where it would be a control that cannot do anything. A run started from the editor uses whatever download answer you have already given; if you have not answered, it takes the classical path rather than starting a download from a surface that never asked.

Related

• AI Deblur — for a photo that is soft overall rather than just in the face
• AI Describe — write alt text for the finished image
• Colorizing black-and-white photos — restore first, colorize second
• Why AI tools download a model once
• System requirements