Why AI Changes Facial Features and How to Get the Likeness Back

"Beautiful, but that isn't me" is the most common complaint about generative portraits. And the face in the output is usually not random: it resembles you, but like a relative would. This is not a malfunction — it follows from how the model works, and understanding the mechanics tells you what to do.

The model works with statistics, not with a copy

A generative model never transfers the pixels of your face into a new frame. It receives a description of appearance — a set of features extracted from the source — and synthesises a face that matches those features. There are many features, but always fewer than the details of a real face.

Everything that did not make it into the description is filled in from the statistics of the training data. That is where the slightly more symmetrical eyes, the slightly more regular nose and the mildly averaged jawline come from. The model is not flattering you on purpose — it fills gaps with the most probable option, and the most probable face is an average one.

What is lost first

Features do not disappear evenly. The first to go are exactly the details that make a face recognisable:

• Asymmetry — brows at different heights, a nose tip slightly off-centre, uneven corners of the mouth. To a person these are identifying marks; to the model they are noise around an average.
• Skin micro-texture: fine lines, pores, scars, moles. They carry age and character and are easily lost when the description is compressed.
• Mid-face proportions — eye spacing relative to face width. The smallest shift here changes the sense of "same person or not".
• The shape of the hairline, which perception latches onto even before the features.

Why a larger photo helps

The more information in the source, the more detailed the description and the less has to be invented. A face filling a quarter of the frame gives the model readable geometry; a face filling a twentieth gives only a general type.

This explains an effect that seems odd at first: two photos of the same person yield different likeness although both look "fine". The difference is not camera quality but how many pixels the face received.

Why beauty retouching hurts

Skin smoothing, enlarged eyes, a narrowed jawline — that is averaging already performed before the model. You submit a face from which individual markers have been partly removed, and the model averages again. Double averaging is what produces the "looks like a relative" effect.

What actually restores likeness

1. A large, sharp, recent frame, front-facing or three-quarter.
2. Soft diffused light: it shows structure without burning the mid-tones or flooding the face with shadow.
3. Beauty mode off and no filters.
4. An unobstructed face: no glare on glasses, no hand on the chin, no hair across the features.
5. A style whose light is close to the source: a contrasty scene over a flatly lit face redraws the structure from scratch.

What does not help

Asking the prompt to "preserve exact likeness" is pointless: the model is already trying, and the phrase adds no information about your face. Nor does running the same poor source many times — you get many variants of the same averaging.

Upscaling the source barely helps either: an upscaler invents detail rather than recovering it, and the model treats the invention as real.

When likeness is not the point

Sometimes averaging serves the task. For a stylised avatar, a decorative cover or an illustration, what matters is a recognisable image, not passport accuracy. There you can safely take a more expressive style and stop chasing exact features.

The boundary is wherever the frame will be seen by people who know you in person. For them the discrepancy reads instantly — and it is precisely those scenarios that demand the best possible source.