A Portrait Prompt Taken Apart: A Worked Example

Prompting theory becomes clearer through a worked example. Let us take a typical vague request and bring it to working condition, explaining each step.

Step zero: the original request

"Portrait of a man, beautiful, professional"

What is here: a subject. What is missing: shot size, light, setting, expression, optics. "Beautiful" and "professional" are judgements — they carry no visual information and the model effectively ignores them.

The result: an unpredictable portrait in an arbitrary style. Every run gives something different.

Step one: shot size and setting

"Portrait of a man of forty, waist-up, grey studio background"

The two most influential parameters after light. Shot size determines what enters the frame; the background removes the randomness of the setting. Even at this step the spread between runs narrows noticeably.

Step two: light

"…soft single-source light from upper left, gentle fill on the right"

The key addition. A physically sensible lighting scheme appears: one direction, readable volume, shadows with comprehensible logic. This is the step that removes the characteristic AI look of even illumination from nowhere.

The fill is mentioned deliberately: without it the shadow side of the face goes too deep.

Step three: expression and gaze

"…looking into the lens, a calm focused expression, a slight turn of the torso"

Without this the model picks a neutral catalogue expression. The torso turn adds naturalness: a strictly frontal pose reads as static.

Note the phrasing "calm focused" rather than "not smiling" — positive description instead of negation.

Step four: optics

"…portrait focal length, shallow depth of field, background in soft defocus"

Optics determine facial proportions and separation from the background. A portrait focal length compresses perspective and gives familiar proportions; a wide lens up close distorts them.

Step five: texture and credibility

"…visible skin texture, natural pores, light grain"

The final touch responsible for realism. Without it you get a smooth surface that reads as a drawing.

The finished request

"Portrait of a man of forty, waist-up, grey studio background, soft single-source light from upper left, gentle fill on the right, looking into the lens, a calm focused expression, a slight turn of the torso, portrait focal length, shallow depth of field, visible skin texture, light grain"

Eleven meaningful elements — at the top of the working range. Adding more is unwise: further refinements would start crowding out the existing ones.

What to change while refining

If the result disappoints, change one element at a time. Too dark — adjust the fill. Face distorted — adjust the optics. Looks like a drawing — strengthen the texture description. Wrong character — adjust the expression.

The temptation to rewrite the whole request is strong, but then you lose the information about what was working.

The photo-mode variant

If the portrait is built on your own shot, the description of appearance drops out (it comes from the photo) while everything else stays. Plus a compatibility consideration appears: the scene's light should not diverge too far from the source's.