NSFW AI Facial Expressions: Write the Look, Not Just the Face

Notes from the director’s chair. You have already set the light, chosen the fabric and placed the camera. Then the picture comes back and something is missing: a beautiful face with nobody home behind it. That is the expression, and in character generation it is the line most people forget to write. Here is how to direct it.

Note one: a face is three parts

Do not write “sexy look”. Write the parts. Eyes: soft and steady, heavy-lidded, narrowed with amusement, crinkled at the corners. Mouth: a closed-lip smile, a half-smile on one side, lips slightly parted. Brows: relaxed, one raised, drawn slightly together in concentration. One word for each part is enough, and the model now has something to build instead of something to guess.

Note two: give her a reason, not a label

“Happy” gets you a stock-photo grin. “Amused, as if she just heard a private joke” gets you a person. Emotion labels point at the average of every face the model has seen with that label; a reason points at a moment. A few that work:

• “smiling at someone just out of frame”
• “the look of someone who knows a secret and will not tell”
• “lost in thought, a moment before she notices the camera”
• “quietly confident, as if she has already won the argument”

Note three: the camera has to be close enough

An expression lives in a few square centimetres of the frame. In a full-figure shot those centimetres become a handful of pixels, and the model smooths them into a neutral face. If the look is the point, ask for a close-up or a medium shot; if the room is the point, accept that the face will be quieter.

Note four: light changes the mood before the face does

The same half-smile reads warm in soft window light and secretive in low-key light with one side of the face in shadow. If an expression feels wrong, try changing the light line before you rewrite the face. The expression may already be there.

Note five: what still breaks

• Wide laughter. Open mouths and rows of teeth are hard shapes; teeth multiply or blur. A laugh caught just after, with the mouth closing, renders better.
• The wink. One closed eye often becomes two. Re-roll, or choose a sideways glance that says the same thing.
• Tears.They tend to come out as glassy streaks. “Eyes bright, on the edge of tears” is more believable than tears themselves.
• Stacked emotions.“Shy, bold, sad and playful” averages into nothing. One feeling per picture.

Note six: keep the face, change the mood

When one picture is right, attach it as a reference and rewrite only the expression line. The reference holds the face; a big change of expression can loosen the likeness a little, so plan for a few takes. The full routine is in keeping one character consistent. One more practical note: if Enhance prompt quality is on, your words go through a longer rewrite before the model sees them, and a subtle expression can get diluted. For an expression study, try one take without it, as the enhancer post discusses.

And an expression can change in motion: Fellowi Video animates a still, so a smile that slowly widens is one sentence in a video prompt. The image-to-video guide covers how.

The lines that do not move

Every face in the frame belongs to a clearly adult person of legal age, and no expression is used to make an adult read as a child: leave out words like “innocent” or “childlike”. Fear or distress in an intimate scene is non-consent, and the terms refuse it everywhere, just as they refuse anything involving minors or age-play; the clearest such prompts are stopped before a single coin moves. No real person’s face belongs in the picture. The full version is in the legal guide.

What expression words cannot do

They steer a likelihood; they do not set a dial. You cannot ask for a smile at exactly thirty percent, the same prompt gives a slightly different face on every take, and the generator does not promise anatomically correct output, so mouths and eyes are where you will re-roll most. Prices and quality levels are on the product page.

A good first exercise in Fellowi Images: one prompt, three reasons behind the same smile. Put the results side by side and you will see which sentence the model actually listened to.