Facial expression and gaze prompts bring an adult AI character to life. Name an emotion (soft smile, parted lips, sultry look), set gaze direction (looking at viewer or looking away), and fix lifeless faces with detailed eyes plus ADetailer. Keep a baseline safety negative on every prompt. All subjects are adult, fictional, AI-generated characters.
A technically perfect render with a dead face fools no one. The lighting can be beautiful and the pose can be right, but if the eyes are glassy and the expression is blank, the image reads as artificial. Expression and gaze are what make a viewer believe there is a person behind the face. This guide is a practical vocabulary for emotion and eye direction in adult AI work, plus the workflow for cleaning up faces with ADetailer so the expression you prompted actually survives to the final image.
Every example here describes an adult (18 and over), fictional, AI-generated character, and every prompt carries the baseline safety negative child, minor, underage, loli, shota. Expression never changes who the subject is: adults only, no real identifiable person’s likeness, no minor or minor-appearing subjects.
Why faces come out lifeless
There are two failure modes. The first is the empty default: with no expression term, the model produces a neutral, slightly vacant face because that is the average of its training data. The second is the small-face problem: when a face occupies only a small fraction of the canvas, the model has very few pixels to work with, so eyes, mouth, and skin all degrade into mush.
The cure for the first is a clear expression term. The cure for the second is face restoration, which is where ADetailer comes in. Most lifeless adult AI faces are some combination of these two issues, and both are easy to fix once you know what to reach for. If you are setting up the whole prompt first, our NSFW AI prompt formula shows where expression sits in the overall structure.
Emotion and expression vocabulary
Here are the expression terms that reliably register, grouped by register. Pick one primary expression so the model is not pulled in two directions.
Warm and inviting
soft smile, gentle smile, warm expression, playful smile, slight smirk, relaxed expression. These read approachable and natural, and they are the safest choices for a believable face.
Sultry and confident
sultry look, confident expression, seductive gaze, half-lidded eyes, parted lips, smoldering look. These set a more charged adult mood while staying tasteful, since they describe expression rather than explicit content.
Calm and contemplative
serene expression, thoughtful look, dreamy expression, relaxed face, eyes closed, peaceful. These read intimate and quiet, good for soft, atmospheric scenes.
Energetic
bright smile, laughing, joyful expression, surprised look, flushed cheeks. These add liveliness and movement to a face.
Gaze direction vocabulary
Gaze is separate from expression and just as important. Where the eyes point changes the entire relationship between the character and the viewer.
looking at viewercreates direct connection and intimacy. It is the single most engaging gaze term.looking away,looking to the side,looking out of frameread candid and contemplative.looking up,looking downadd mood and can read as shy, demure, or dominant depending on context.eye contact,direct gazereinforce connection.eyes closedremoves gaze entirely for a soft, internal feel.
Gaze and expression combine to produce a specific read. A soft smile with looking at viewer is warm and inviting. A sultry look with looking away is mysterious. Mixing the two axes is how you fine-tune the emotional tone.
Expression and gaze comparison table
| Expression | Gaze | Combined read | Best for |
|---|---|---|---|
| Soft smile | Looking at viewer | Warm, inviting | Friendly portraits |
| Sultry look | Looking at viewer | Direct, charged | Confident adult scenes |
| Sultry look | Looking away | Mysterious, candid | Editorial mood |
| Serene | Eyes closed | Intimate, quiet | Soft atmospheric work |
| Playful smirk | Looking to the side | Flirty, casual | Lifestyle scenes |
| Confident | Direct gaze | Strong, assured | Bold figure work |
Copy-paste expression recipes per model
Every example is an adult (18 and over), fictional, AI-generated character. Each pairs one expression with one gaze and carries the baseline safety negative. Swap the expression or gaze to retarget.
Pony Diffusion: warm and inviting
score_9, score_8_up, score_7_up, 1woman, adult woman, 26 years old, mature,
fictional character, soft smile, looking at viewer, detailed eyes, warm expression,
soft lighting, photorealistic, sharp focus on face
Negative: child, minor, underage, loli, shota, blank expression, dead eyes,
lowres, bad anatomy, deformed face, watermark, blurryOur Pony Diffusion guide covers how the scoring tags interact with face detail.
Illustrious / anime: sultry and mysterious
masterpiece, best quality, 1girl, adult woman, mature, fictional character,
sultry look, half-lidded eyes, looking away, parted lips, detailed eyes,
blush, cinematic lighting
Negative: child, minor, underage, loli, shota, blank expression, lifeless eyes,
worst quality, lowres, bad hands, extra digits, watermarkThe Danbooru tags reference and the Illustrious model guide list the anime-specific expression and eye tags.
SDXL realistic: confident direct gaze
photo of an adult woman, 29 years old, fictional AI character, confident expression,
direct gaze, looking at viewer, detailed natural eyes, slight smile,
soft studio lighting, natural skin texture, sharp focus
Negative: child, minor, underage, loli, shota, blank stare, glassy eyes,
plastic skin, lowres, deformed, watermark, textFlux: serene and intimate
Close portrait of an adult fictional woman, early thirties, with a serene
expression and eyes gently closed, soft natural lighting, calm and intimate
mood, shot on film, detailed realistic skin, tasteful.
Negative cue: child, minor, underage, loli, shota, clearly an adult,
no deformed face, no extra eyes, no watermark.Flux reads natural language, so describe the expression and gaze in sentences. Our Flux NSFW guide explains its prompting style.
Try these in our free generator and switch only the gaze term to feel how much the connection with the viewer changes.
How to fix lifeless faces
When a face still looks dead after you have added an expression term, work through this checklist in order.
Add eye detail tokens. detailed eyes, detailed face, sharp focus on face, and on anime models beautiful detailed eyes push the model to spend more attention on the face. Glassy or asymmetric eyes are the number one tell of an AI face.
Negate the dead look. Put blank expression, dead eyes, glassy eyes, lifeless in your negative. Steering away from the failure mode is often as effective as steering toward the goal.
Frame the face larger. A portrait or upper-body framing gives the model more pixels for the face. If you need a full-body shot, the small face will need restoration, which is the next step. Our camera angle prompts explain framing choices.
Run ADetailer. This is the single most effective fix and it deserves its own section below.
| Symptom | Cause | Fix |
|---|---|---|
| Blank, vacant face | No expression term | Add one clear expression |
| Glassy or asymmetric eyes | Too few face pixels | Add detailed eyes, run ADetailer |
| Melted face in full body | Small face on canvas | ADetailer face restoration |
| Wrong emotion | Conflicting terms | One primary expression only |
ADetailer for face cleanup
ADetailer is an extension for AUTOMATIC1111, Forge, and other interfaces that automatically detects the face, crops to it, regenerates that crop at higher effective resolution, and blends it back in. It is the standard tool for fixing the small-face problem, and it transforms full-body adult renders where the face would otherwise be mush.
The basic workflow: enable ADetailer, select a face detection model such as face_yolov8n, and optionally give it its own prompt. A common trick is to put the expression and eye detail tokens in the ADetailer prompt specifically, so the face regeneration focuses on exactly the expression you want. For example, an ADetailer prompt of detailed face, soft smile, detailed eyes, looking at viewer with the baseline safety negative cleans up the face while honoring your intended expression.
Keep the ADetailer denoising strength moderate, around 0.3 to 0.45. Too high and the face drifts away from the rest of the image. Too low and it does not fix anything. ADetailer fits naturally into a larger editing pass, which our NSFW photo editing workflow covers end to end, and it runs inside ComfyUI as well through equivalent face-detailer nodes.
Always carry the baseline safety negative into the ADetailer prompt too. Because ADetailer regenerates the face, it must respect the same adult anchoring and child, minor, underage, loli, shota negative as the main prompt, so the restored face stays unmistakably adult.
Keeping expression consistent across a character
If you are building a recurring adult character, expression and the face itself should stay recognizable across images. Reference based tools help here: our IPAdapter guide carries a face across renders, and the broader character consistency techniques article compares LoRAs, embeddings, and reference methods. Combine a consistency tool with a fixed expression term to keep both the identity and the mood stable across a set.
Building an expression and gaze library
The fastest way to get reliable faces is to stop improvising and build a small library of expression-and-gaze pairs that you know work on your checkpoint. Generate a test batch where you hold everything constant except the expression, then save the pairs that produce a living, believable adult face. Over a few sessions you will end up with a personal shortlist, for example soft smile plus looking at viewer for warm portraits, sultry look plus looking away for editorial mood, and serene plus eyes closed for quiet intimate scenes.
A library also makes consistency easier. When you reuse the exact same expression and gaze tokens across a set, the mood stays coherent even as you change pose, outfit, or setting. This pairs naturally with the build and outfit work in our body type prompts and outfit prompts, since a consistent face on a consistent character is what makes a series feel like one person rather than ten different ones.
Keep the safety baseline baked into every saved recipe so you never accidentally reuse one without it. Each entry in your library should carry the adult anchor and the child, minor, underage, loli, shota negative as part of the saved prompt, so that adults only, fictional AI characters, and no minor or minor-appearing subjects is the default rather than something you remember to add.
Expression in profile and three-quarter views
Gaze and expression read differently depending on the angle of the face. A front-on face shows the full expression and both eyes, so looking at viewer lands with maximum impact. A three-quarter view, the most flattering angle in portraiture, softens the expression and adds dimension, and it pairs beautifully with looking to the side or looking away. A pure profile hides one eye, so gaze terms matter less and the silhouette of the expression, such as parted lips or a lifted chin, carries the mood instead.
When you change the face angle, adjust the gaze term to match. Asking for a strict profile plus looking at viewer fights itself, because a profile face cannot easily meet the camera, and the model may produce a twisted neck or crossed eyes. Match the gaze to the angle: front-on for direct connection, three-quarter for flattering candor, profile for graphic silhouette. Our camera angle prompts cover how face angle fits the larger composition.
Subtlety beats exaggeration
The most common expression mistake is overdoing it. Stacking seductive, sultry, smoldering, parted lips, bedroom eyes all at once produces a cartoonish, uncanny face. Pick one primary expression term and let it breathe. A single well-chosen expression with good eye detail reads far more convincingly than a pile of intense ones fighting for control of the same face.
The same restraint applies to gaze. One gaze direction per image. You cannot have a character looking at the viewer and looking away at once, and asking for both produces crossed or doubled eyes. Choose the relationship you want with the viewer, name it once, and let the expression carry the rest.
If you want to combine subtle expression with strong art direction, our art style prompts and lighting prompts layer cleanly on top. Soft window light plus a soft smile plus looking at viewer is a reliably beautiful, lifelike combination.
Expression by checkpoint and style
Different base models render the same expression differently. Photoreal SDXL checkpoints reward subtle, naturalistic expressions: a slight smile reads as a real micro-expression, while an exaggerated one looks staged. Anime and Illustrious models love graphic, readable expressions: blush, half-lidded eyes, and parted lips are part of the visual language and register strongly. Flux handles nuanced descriptive expressions well when you write full sentences, and it tends to produce calmer, more photographic faces by default.
A practical rule: on photoreal checkpoints keep expressions subtle and lean on eye detail. On anime checkpoints use the established expression tags freely because the style expects them. On Flux, describe the feeling rather than tagging it. Our best checkpoints roundup helps you pick the base model, and from there the expression vocabulary follows the style.
Blush, lips, and micro-detail
Beyond the broad expression, a few small tokens add a lot of life. blush and flushed cheeks add warmth and a sense of feeling. parted lips and slightly open mouth add softness and movement. glistening eyes and wet eyes add catchlights that make the gaze feel alive. Use these as supporting detail behind one primary expression, not as replacements for it.
Catchlights, the small bright reflections in the eyes, are what most strongly separate a living face from a dead one. Lighting drives catchlights, so an expression prompt works best alongside a directional lighting term. Our lighting prompts explain how a soft key or rim light puts a believable spark in the eyes, which is why expression and lighting are best tuned together.
Putting it together
A believable adult AI face comes from three deliberate choices: one clear expression, one gaze direction, and enough face detail to render the eyes well. Add the eye detail tokens, negate the dead-eye failure mode, frame the face large enough or restore it with ADetailer, and the same character that looked vacant becomes alive and engaging.
Keep the safety baseline on every step, including the ADetailer pass. Adults only, fictional AI characters only, never a real identifiable person, and child, minor, underage, loli, shota in the negative every time. Build a small set of expression-and-gaze pairs you trust, run them through our generator, and pick the combinations that consistently produce a face with a person behind it.
Frequently asked questions
Why do my AI faces look dead or vacant?
Two reasons. First, with no expression term the model renders a neutral, average face that looks empty. Add one clear expression like soft smile or confident expression. Second, when the face is small on the canvas there are too few pixels to render the eyes well, so it turns to mush. Frame the face larger or run ADetailer to regenerate the face at higher effective resolution.
What is the most engaging gaze direction?
Looking at viewer. A direct gaze creates immediate connection and intimacy, which is why it is the most engaging single gaze term. Pair it with a warm expression for an inviting read, or with a confident expression for a strong one. If you want a more candid, mysterious mood, looking away or looking to the side reads contemplative. Use only one gaze direction per image.
How does ADetailer fix faces?
ADetailer automatically detects the face, crops to it, regenerates that crop at a higher effective resolution, and blends it back into the image. It is the standard fix for the small-face problem in full-body renders. Give it its own prompt with detailed face plus your expression tokens, keep denoising around 0.3 to 0.45, and always carry the baseline safety negative so the restored face stays clearly adult.
Can I set expression and gaze separately?
Yes, and you should. Expression is the emotion, like soft smile or sultry look. Gaze is where the eyes point, like looking at viewer or looking away. They are independent axes, so combining them lets you fine-tune the mood precisely. A sultry look with looking away reads mysterious, while a soft smile with looking at viewer reads warm and inviting. Choose one of each.
Why does stacking many expression terms backfire?
Because the model tries to satisfy all of them at once and produces a cartoonish, uncanny face. Stacking seductive, sultry, smoldering, and parted lips together fights for control of the same features. Pick one primary expression term and let it breathe, supported by detailed eyes. A single well-chosen expression with good eye detail reads far more convincingly than a pile of intense ones.
Should the ADetailer pass include safety negatives?
Absolutely. Because ADetailer regenerates the face, it must respect the same rules as the main prompt: adults only, fictional AI characters, never a minor or minor-appearing subject. Carry the adult anchor and the child, minor, underage, loli, shota negative into the ADetailer prompt so the restored face stays unmistakably adult. Skipping this on the face pass would defeat the safety steering you applied to the main prompt.
What eye tokens improve realism most?
On realistic models, detailed eyes, detailed face, and sharp focus on face push the model to spend more attention where it matters. On anime models, beautiful detailed eyes works well. Pair these with negatives like glassy eyes, dead eyes, and blank expression. Glassy or asymmetric eyes are the number one tell of an AI face, so steering both toward detail and away from the failure mode is the most effective combination.
How do I keep the same expression across a character set?
Use a consistency tool together with a fixed expression term. IPAdapter carries a face across images, and LoRAs or embeddings lock identity even harder. Keep the same expression and gaze tokens in every prompt of the set so the mood stays stable alongside the identity. For posed continuity, combine this with pose control so the character keeps the same face, expression, and position across the whole series.
