Face Restore AI that rebuilds a face without swapping it for a nicer one.

What is actually happening to the face

A photograph that has passed through a messaging app is not the photograph you took. Each app resizes it, compresses it, and often does both again when it is forwarded. The result is a file that still looks acceptable at thumbnail size and falls apart the moment you look at a face in it.

The damage is specific. Compression works by discarding fine variation and keeping broad shapes, which is exactly the wrong trade for a face. Eyelashes, the edge of an iris, the texture of skin, the line where lips meet — these are small, low-contrast details, and they are the first thing thrown away. What survives is a smooth mask with the right general shape and none of the specifics that make it a particular person.

This is why sharpening does not fix it. Sharpening increases the contrast of edges that are still in the file. When the detail has been discarded rather than blurred, there is no edge left to raise. You get a crisper version of the same smooth mask.

Rebuilt, not recovered — and the difference matters

Face Restore does not find the missing detail. It cannot; the detail is gone. What it does is generate detail that is consistent with what remains, using a model that has seen a very large number of faces and knows how they are put together.

That is a real distinction and it is worth being clear-eyed about. The result is a plausible face that matches the shape, proportions, expression, and lighting of the one you uploaded. It is not a photograph of what the person's face looked like at that moment. It is the model's best account of it.

In practice, on a photo that is merely small or compressed, that account is very close. Enough structure survives to constrain the model tightly, and the output reads as the same person because it is being led by the same underlying shape. The further the source degrades, the more the model is filling in rather than following, and the less you should trust the specifics.

Why this runs cautiously by default

The model behind this tool has a dial that trades likeness against smoothness. Turned one way it produces a cleaner, more attractive face that drifts from the person. Turned the other it stays closer to the original and accepts a rougher, less impressive result.

It is set well towards likeness here, and deliberately. The most common failure in this category is a face that looks great and belongs to someone else, and it is a failure people do not notice, because a smoother face is flattering and there is nothing in the image announcing that it is wrong.

The visible cost of that choice is that results look less dramatic than the marketing images you may have seen elsewhere. Some texture stays. Some softness stays. That is the setting doing its job.

The photo comes back the same size

This tool changes faces and nothing else. The background is left exactly as it was, and the image comes back at the resolution you uploaded rather than enlarged.

There is one exception, and it is a property of the model rather than a choice: an image smaller than 512 pixels on its long edge is normalized up to 512. Anything at or above that size is returned at its original dimensions.

The reason to hold it there is that a tool which quietly enlarged your photo would be two tools wearing one name, and you would have no way to tell which part of the change came from which. If you want a bigger file as well, run Upscale afterwards, which enhances faces as part of its own pass.

Why black-and-white photos are sent elsewhere

This is the limitation worth knowing before you upload anything, and it is the reason this tool is described as being for color photographs.

The model was trained on color faces, and it reconstructs in color. Given a monochrome print, it does not stay monochrome — it assigns skin tone and, more troublingly, eye color. Eye color cannot be derived from a black-and-white photograph. The information is not in the negative. Testing this tool on a sepia portrait produced a technically better image of a person with blue-grey eyes that nobody can verify.

For a monochrome print, Restore repairs the damage and handles color as an acknowledged inference rather than a silent one, and Colorize exists for the color question on its own. Sending old prints there is not a smaller version of this tool; it is a different and more honest treatment of the problem.

Reading the result before you keep it

Compare it against another photograph of the same person, ideally one taken around the same time. This is the only check that means anything, and it takes a few seconds.

Look at the things a model is most likely to standardise: the shape and thickness of eyebrows, the exact line of the nose, the set of the mouth at rest, and any asymmetry. Real faces are asymmetric, and a face that has become tidier is a face that has moved away from the person.

Skin is the second thing to check. Marks, lines, and pores that were visible in the source should still be visible. If the skin has become uniform, the result has been smoothed rather than restored, and it will look wrong to anyone who knew them.

If you cannot verify the face against anything, treat the output as an illustration rather than a record, and say so if you pass it on.

Getting the most out of a bad source

Start from the largest copy you can find. People reach for this tool on a photo pulled from a chat thread when the original is still on a phone, in an email, or on a memory card. The original will always give a better result, because the model has more structure to follow.

Do not sharpen, denoise, or adjust the photo first. Those steps alter the evidence the model reads, and a pre-sharpened face gives it exaggerated edges to build on. Run this first, then apply other tools if the result still needs them.

Crop as little as possible. The model locates faces in the frame, and a very tight crop can leave it with too little context to work from.

If several people are in the shot, all of their faces are processed in the same pass. There is no way to select one, so check every face in the result, not only the one you came for.

Cost, and what you get to keep

Face Restore costs 2 credits per image, the same as Scratch Repair and Background Remove.

The first result is free and needs no account. That is deliberate for this tool in particular, because whether the output is convincing depends heavily on your specific photo, and no description of the trade-offs substitutes for looking at your own.

Results download at full resolution with no watermark. Uploads are processed to produce your result and are not used for training.