Generating Realistic Faces With AI (Without the Uncanny Valley)

What makes a face believable

Three things human eyes lock onto instantly:

1. Eye symmetry and gaze direction — asymmetric eyes read as wrong immediately.
2. Skin texture — slightly rough is real; plastic-smooth is uncanny.
3. Micro-expressions — a subtle asymmetry in the smile makes a face read as alive.

Nail those three and 80 % of the "uncanny" problem is solved.

The prompt formula

A [DESCRIPTOR] portrait of a [PERSON DESCRIPTION], [AGE], natural expression, direct eye contact, subtle skin texture, [LIGHTING], shot on Sony A7 IV, 85mm f/1.4 portrait lens, 3:4 aspect

Real example:

A cinematic portrait of an elderly Korean man with silver hair and reading glasses, 70s, warm gentle expression, direct eye contact, subtle skin texture visible pores, warm window light, shot on Sony A7 IV, 85mm f/1.4 portrait lens, 3:4 aspect

Negative prompt

plastic skin, waxy skin, asymmetric eyes, cross-eyed, distorted, blurry, low quality, bad anatomy, extra fingers, mangled hands, oversaturated, cartoon, doll-like

Aspect ratio

3:4 (896×1152) always for a single-person portrait. This is the ratio SDXL was trained on for the majority of portrait images — the model's internal composition sense expects it.

The eye-fix trick

If the eyes look wrong on an otherwise-good generation:

1. Inpaint just the eye area
2. Prompt: "detailed eyes, symmetrical, natural iris color, subtle reflections"
3. Denoise 0.4

Fixes 80 % of eye artifacts in one pass.

Ethics reminder

Do not generate photorealistic images of real people you don't have permission to depict. Regardless of platform rules, it's a fast path to legal trouble.