Prompting Facial Expression and Emotion

Writing "happy" or "sad" as your expression descriptor produces the most statistically average version of that emotion the model has seen in its training data — which, for "happy," is almost always an open-smile, teeth-showing, creased-eye expression that reads as performance rather than genuine feeling. Real human facial expressions are vastly more specific. Happiness can be a barely perceptible smile in someone who is holding back emotion; or a wide, laughing, head-thrown-back expression; or a quiet, warm look of contentment. Each of these is emotionally distinct, and the differences are expressed through specific micro-muscle movements — the Duchenne smile marker in the outer eye corner, the degree of mouth opening, the relaxation or tension in the brow. To produce portraits that feel psychologically real rather than stock-photo generic, you need to describe expression at this micro-component level. AI models have seen enough photographic annotation to respond to these micro-descriptions, but you have to supply them explicitly rather than relying on a single emotion label to do all the work.

Describe facial expression by decomposing the face into its key expressive zones and naming the state of each one. Eyes: describe pupil direction (direct camera gaze, looking slightly down, eyes cast to the side), eyelid state (slightly narrowed, fully open, hooded, wide with tension), and the presence or absence of emotional engagement in the eye itself — "soft eyes with genuine warmth visible," "flat, distant eyes — blank expression," or "eyes slightly wet with held-back emotion." Brow: furrowed (concentration, concern, anger), relaxed (neutral, calm), slightly raised (surprise, openness), asymmetrically raised one side (skepticism, curiosity). Mouth: corners turned up slightly, corners slightly downward, lips pressed together (restraint), slightly parted (anticipation, openness), full smile with teeth, soft closed-mouth smile, tight compressed smile (masking emotion). Overall muscle tension: relaxed face vs. jaw tension vs. high facial muscle engagement. A composed complex expression description: "soft closed-mouth smile, slightly upturned corners only, eyes warmly engaged with a hint of amusement in the outer eye corners, relaxed brow, no jaw tension — an expression of quiet, private pleasure." This is specific enough that a model can approximate the intended emotional register without defaulting to a generic smile.

Expression alone does not communicate emotion fully — context shapes how an expression is read. The same slight smile in a professional headshot reads as "confident and approachable"; in a dimly lit intimate setting it reads as "tender or nostalgic"; in bright overhead light with the camera below eye level it reads as "threatening or sarcastic." Add emotional context layers around your expression description to steer interpretation. Narrative context: add a phrase that implies the emotional situation — "a woman who has just heard unexpectedly good news, expression catching between surprise and relief" or "a man looking at something he lost a long time ago, expression of bittersweet recognition." These narrative clauses tell the model the emotional story rather than just the muscle state, and models trained on captioned photography have learned to associate these narrative cues with specific facial combinations. Environmental context: the setting amplifies expression — a contemplative expression in a rain-soaked window scene reads differently than the same expression in a sun-lit garden. Name the environment as emotionally supportive of the expression you want. Lighting: direction and quality of light affects how expression reads — soft frontal fill light shows expression clearly and benevolently; hard sidelight casts dramatic shadows that intensify any expression; backlight reduces expression readability and increases mystery. Choose your light to complement the emotional register.

Several recurring expression problems appear in AI portrait generation, each with a specific prompt fix. Generic stock-smile: the model defaults to a wide, performative smile. Fix: replace "smiling" with "subtle closed-mouth smile, soft eyes, relaxed facial muscles — a genuine rather than posed expression." Dead eyes: the face is technically smiling but the eyes are flat and emotionless — the most common tell of an AI-generated face. Fix: explicitly describe the eye state — "genuine Duchenne smile reaching the outer eye corners, slight crow's feet visible, eyes engaged and warm" — or add "natural, authentic expression, not posed." Symmetrical brow uniformity: real faces are subtly asymmetric; AI faces are often perfectly symmetric in ways that read as uncanny. Fix: specify slight asymmetry — "left brow fractionally higher than right, thoughtful expression" — or describe "natural facial asymmetry, slightly imperfect symmetry." Frozen action expression: open-mouth expressions for laughing or speaking often look frozen mid-action. Fix: describe the surrounding context as a live moment — "caught mid-laugh, eyes almost closed, genuine unposed laughter" — or "candid moment, expression not posed for camera, natural candid energy." Adding "candid photography" or "documentary portrait" to your style description also shifts the model away from posed studio convention toward more naturalistic expression capture.

When generating multiple images of the same character across different contexts — as in an avatar system or a character-consistent content series — expression management becomes a consistency challenge as well as an accuracy challenge. The same character needs to convey different emotions across images while remaining recognizably themselves. The Floniks /ai-avatar tools and character-consistent /editor workflows address this at the identity level, but the expression layer still needs your prompt guidance on top of the character reference. In character-consistent workflows, keep the identity description in one fixed node and the expression description in a variable node that you update per image. Do not try to encode both identity and expression in a single combined prompt — the model may trade off between the two constraints, producing a face with the right expression but drifted identity. Separate the concerns: Node 1 establishes the character with their physical identity descriptors and a neutral reference expression; Node 2 takes that output as an image reference and applies the target expression description as a new prompt layer. This two-pass approach is significantly more reliable for expression variation in character-consistent work than single-pass generation, because the model has a stable visual identity to modify rather than constructing both identity and expression from text simultaneously.