Recover Detail in a Small or Soft Face

How it works

Your image is scaled to about one and a half megapixels — more than the other portrait pages use, because the face may be a small part of the frame — and repainted at a moderate denoise with a fresh seed each run.

The instruction asks for definition to be rebuilt in the four places a face reads from: eyes, lashes, lips, and hairline. It also states three constraints explicitly, which are the same person, the same expression, and the same head position, and the negative prompt names the specific failure of this class of model, which is a symmetrical doll face replacing an individual one.

A sliding scale, not a switch

The useful way to think about this page is as a ratio between what was measured and what was assumed.

A head nine hundred pixels wide contains real information about the shape of an eyelid, the width of a nostril, and the line of a lip. Sharpening it up is mostly refinement: the model is deciding how to render detail whose position is already determined by the data.

A head eighty pixels wide does not contain that. An eye is perhaps four pixels across. There is enough there to say a face is present, roughly which way it is turned, and roughly how light it is, and almost nothing beyond that. Everything the output shows at feature level was supplied by the model.

Nothing in the interface marks which case you are in, and both produce a clean, sharp, confident result. Counting the pixels across the head in your original takes five seconds and is the only reliable way to know.

Symmetry is the tell

There is one artifact worth learning to spot, because it appears reliably when the prior is doing most of the work.

Real faces are not symmetric. One eye sits slightly higher, one side of the mouth lifts more, the nose leans a little, the ears are at different heights. Those asymmetries are a large part of how people recognize each other, and they are also fine detail — which means they are the first thing lost when a face gets small in the frame.

A model rebuilding from a strong prior produces a face closer to the average, and average faces are more symmetric than real ones. So when an enhanced face looks somehow cleaner and more balanced than the person you know, that is not the photograph improving. That is the prior showing through.

What it is genuinely good for

The warnings are not an argument against using it, only against trusting it in one particular way.

Recovering a family group shot for printing, cleaning up a face in an old scan that will hang on a wall, making a video still usable as a thumbnail, rescuing a portrait that was slightly soft — all of these are cases where a plausible face is exactly what is wanted, and where nobody is going to make a decision about a person based on the pixels.

Use it there without hesitation. Keep it away from anything where the question is who somebody is.

Publication gate

This page ships once the workflow has been run against a large soft portrait, a medium face at around three hundred pixels, a small face in a group shot, and a face already sharp, with eye spacing and mouth width measured against the source in every case.