Why AI Portraits Stop Looking Like You

“It’s lovely, but it isn’t me.” That’s rarely the model being bad at faces. It ran short of evidence somewhere and filled the gap itself, with a face that’s attractive in general and not yours in particular.

“Not me” is usually one of three things

You got beautified. Skin too smooth, jaw narrowed, eyes enlarged. That’s the model’s default taste deciding for you.
The proportions drifted. Each feature looks right on its own, but together they look like a cousin. It happens when strength is too high or the face is too small in the frame.
The expression went missing. The structure is right, but it isn’t you on that day. Common when the original expression was blurred or half hidden.

What the two settings control

Style strength: how hard the style presses

Strength sets how heavy the treatment is. It doesn’t act on the face directly. Soft is a light touch, and the original is still obvious at a glance. Balanced is a confident middle. Bold commits to the style completely.

Turn it up and more of the texture, colour and mark-making gets rewritten, and the face goes along with the rest of the picture. That’s how strength affects likeness indirectly. For a first portrait, stop at balanced.

Subject fidelity: how closely the subject follows the source

This is the setting that acts on likeness directly, and it has two positions. High keeps detail and proportions strictly faithful. Natural allows a little artistic licence in texture, as long as the subject stays clearly recognisable.

Use high for profile pictures, family portraits and anything people who know you will see. Save natural for when you want charcoal, watercolour or woven texture to really take hold.

Composition: keep or reframe

Keep holds the original framing. Reframe allows a slightly different crop while the subject stays dominant. Keep is safer for portraits, because reframing changes the size of the face, and a smaller face has fewer pixels to work with.

One rule sits underneath your settings. Every style spec preserves facial structure, expression, pose, hairstyle, skin tone and accessories like glasses, with no prettifying, no swapping and no change in the number of people. The settings work inside that rule.

The source photo sets the ceiling

Settings can only work with what the photo gives them. If the information isn’t there, no setting brings the likeness back. Three things matter, in this order.

Light: the face needs readable shape

Soft light with a direction is best, from a window, an overcast sky or open shade. Avoid the extremes. Too dark, and the structure of the face sinks into noise. Too bright, and the nose, cheeks and chin burn out to flat white.

The app checks brightness and sharpness when it reads the image, and warns about dark, blown or soft frames. Those warnings are good predictors, so don’t ignore them. Hard top light digs black holes under the eyes and strong backlight turns the face into a silhouette; either way, the model has to guess.

Resolution: the face needs enough pixels

The size of the whole file matters much less than the number of pixels on the face. The app flags images with a short edge under 768 pixels as low resolution, and in practice likeness starts to slip once the face is narrower than about a sixth of the frame.

Sources are scaled to around 1024 pixels before generation. A 40-megapixel shot of someone across the street can end up with a face a few dozen pixels wide, so crop to half length or chest up before uploading. One person, half length or head and shoulders, is the steadiest framing, and a frame can hold at most six people.

Obstruction and angle: the evidence has to be complete

Sunglasses, masks, a hand near the face, hair across the eyes, a steep turn of the head: each one removes evidence. Ordinary glasses are fine and get kept as an accessory, but a hard reflection on the lenses hides the eyes, which are the most important part.

Before you shoot

Face a soft light, so there’s direction on the face but no blown-out patches.
Frame at half length or head and shoulders, so the face is properly large.
Keep the eyes visible, with no glare or anything in the way.
Focus on the eyes, not the background or the tip of the nose.
Wear the expression you want to keep. The model keeps it; it won’t improve it.

When it drifts, fix it in this order

Hold the image to see the original and drag the slider across it, so you know exactly what moved. Then set fidelity to high, take strength down one step, and put composition back to keep.

If it still misses, the photo is the problem. Use one with a larger, better-lit face instead of trying more styles. Switching exhibition comes last; soft-light directions like Quiet Portraits keep likeness more easily than the poster styles do.