Prompt the Details AI Gets Wrong: Hands, Eyes, Skin, Teeth

You generate a portrait and it is almost perfect. The lighting is right, the pose is natural, the outfit lands. Then you look for one more second and something is off. It is the eyes, or the hand resting on the table, or skin so flawless it looks like plastic. Ninety-five percent perfect, five percent uncanny, and the five percent is all anyone notices.

Realism does not fail in the big things. Modern models nail composition and light. It fails in a short list of small things the eye is ruthlessly tuned to catch. Get those right and a portrait crosses from impressive to believable. Get them wrong and no amount of polish elsewhere saves it.

Quick Answer: AI portraits give themselves away in four places, hands, eyes, skin, and teeth. Fix hands by keeping them simple and out of complex positions, or cropping them out entirely. Fix eyes by prompting for natural asymmetry, a real catchlight, and a steady gaze instead of a dead stare. Fix skin by asking for texture, pores, and slight imperfection instead of smoothness. Fix teeth by keeping smiles closed or soft rather than wide and toothy. When wording is not enough, reframe the shot so the hard detail is not the subject.

Why These Four Details Break Realism

The human eye is a specialist in exactly these features, which is why they are the ones that betray a fake.

We spend our whole lives reading faces and hands. We know without thinking what a real eye reflects, how skin catches light, how a natural hand rests. So when a model gets those subtly wrong, we feel it instantly, even if we cannot name what is off. Meanwhile we are far more forgiving of a slightly odd background or an imperfect fold in a shirt, because we are not experts in those.

That asymmetry is the whole game. You get enormous returns from prompting the four features the viewer scrutinizes most, and almost no return from over-specifying the things they ignore. Spend your prompt budget where the eye actually looks.

There is also a base-model factor. Different models handle these details differently, and the gap between them is exactly in the fine stuff, as I covered in Flux versus SDXL for influencer faces. Starting from a model that handles skin and eyes well means you fight these battles less often.

How Do You Prompt Hands That Look Real?

Hands are the hardest problem in AI imagery, full stop. Fingers multiply, bend wrong, and fuse. Do not expect a prompt to fully solve them, and plan your shot around that reality.

The most reliable move is avoidance by composition. If a shot does not need hands doing something intricate, keep them out of it or keep them simple. Hands at rest, in pockets, holding a single simple object, or naturally out of frame will give you a far higher hit rate than hands mid-gesture. A portrait cropped at the chest sidesteps the entire problem.

When hands must be visible, keep the position simple and unambiguous. One hand resting flat, fingers relaxed and slightly apart, holding one clear object. Complexity is what breaks them, so wording that reduces complexity helps more than wording that describes perfect anatomy. Asking for anatomically correct hands is weaker than just not asking for a complicated hand pose in the first place.

And accept the reroll. Even with good prompting, hands need more attempts than any other feature. If a shot is otherwise great and only the hand is wrong, reroll that shot rather than rewriting your whole approach. A bad hand in one frame is a seed problem, not a strategy problem.

Getting Eyes and Skin to Read as Human

Eyes and skin are the two features that most often push a portrait into uncanny territory, and both fail in the same direction, too perfect.

Eyes go dead when they are too symmetrical and lack a real reflection. A living eye has a catchlight, the little bright reflection of the light source, and it has subtle asymmetry, because no two real eyes are identical. Prompt for a natural catchlight, a soft and present gaze, and slight natural asymmetry. Avoid anything that pushes toward a flat, staring, perfectly matched pair, which is the classic doll-eye tell. The gaze should feel like it is looking at something, not through you.

Skin fails when it is flawless. Real skin has pores, faint texture, minor imperfections, and the way it catches light is uneven. Models love to smooth all of that into a plastic sheen, so you have to prompt against it. Ask for natural skin texture, visible pores, subtle imperfections, and realistic skin detail. Counterintuitively, adding a hint of imperfection is what makes skin read as real. Perfect skin looks synthetic because real skin never is.

Lighting quietly decides how well eyes and skin land, because both are about how light behaves on a surface. Good light gives skin its texture and eyes their catchlight, which is why lighting prompts that hold across a set do double duty as realism prompts.

The Teeth Problem, and When to Just Reframe

Teeth are the detail people forget until a smiling portrait looks wrong for no obvious reason. The reason is usually the teeth.

Models tend to render teeth too numerous, too even, and too white, an unsettling perfect grid. The fix that works most reliably is not describing perfect teeth, it is avoiding the wide toothy smile that exposes the problem. A closed-mouth smile, a soft smile, or a relaxed neutral expression sidesteps it entirely. When you do want visible teeth, prompt for a natural, slightly imperfect smile rather than a bright even one, and accept that this is a reroll-heavy detail.

This points at the single most useful realism principle. When a detail resists prompting, reframe so the detail is not the subject. This applies to all four features. Hands break, so crop them out. Teeth break, so close the smile. A wide grin with both hands gesturing is asking the model to nail its two weakest features at full size in one frame. A calm three-quarter portrait with a soft closed smile and hands at rest is asking for almost none of that.

Composition is a realism tool, not just an aesthetic one. The same camera and lens choices that set a mood also decide how much of each hard detail you are forcing the model to render. Frame to hide weaknesses and you will reroll far less.

Building These Into Your Standard Prompt

You do not want to reinvent these clauses every time. The move is to fold them into a reusable base you attach to portrait work, then vary only the scene.

Keep a small standing block that always asks for natural skin texture and a real catchlight in the eyes, and always leans toward simpler hands and softer smiles by default. Then your per-image prompt only has to describe the scene, the outfit, and the mood, while the realism floor is handled underneath. This is the same layering logic behind photoreal product prompts built from subject, surface, light, and lens, just pointed at people instead of objects.

The payoff compounds across a batch. Bake the four fixes into your default and every image starts one notch more believable, so your reroll rate on the uncanny details drops for the whole set, not just one lucky frame. Realism stops being something you chase per image and becomes a property of your standard setup.

FAQ

Why Are Hands So Hard for AI to Get Right?

Hands are complex, highly variable, and appear in countless positions, so models struggle to render them consistently. Fingers duplicate or fuse easily. The most reliable fix is not better wording but composition, keeping hands simple, at rest, or out of frame, and rerolling when a shot needs them visible.

What Makes AI Eyes Look Dead?

A missing catchlight and too much symmetry. Real eyes reflect the light source and are never perfectly matched, so a flat, symmetrical, reflectionless pair reads as artificial. Prompt for a natural catchlight, a soft steady gaze, and slight asymmetry to bring them to life.

Should I Prompt for Perfect Skin?

No, the opposite. Perfect smooth skin looks plastic and synthetic. Real skin has pores, texture, and minor imperfections, so prompt toward those. A little imperfection is what makes skin believable, while flawlessness is a dead giveaway.

How Do I Fix Bad Teeth in a Smile?

Usually by avoiding the wide toothy smile that exposes them. A closed or soft smile removes the problem entirely, and a relaxed neutral expression works too. When you want visible teeth, prompt for a natural, slightly uneven smile rather than a bright, perfectly even one.

Is It Better to Fix Details in the Prompt or Reframe the Shot?

Both, but reframing is the stronger lever when a prompt keeps missing. If a detail resists wording, compose so that detail is not the subject. Crop out hands, close the smile, choose an angle that hides the weakness. Composition solves what prompting cannot.

Do Some Models Handle These Details Better Than Others?

Yes. The gap between base models shows up exactly in fine details like skin and eyes. Starting from a model known to handle those well means fewer battles per image. The Flux versus SDXL comparison covers how that choice affects faces specifically.

Wrapping Up

A portrait does not fail in the obvious places. It fails in four small ones the eye is built to catch, hands, eyes, skin, and teeth. Prompt hands toward simplicity or crop them out. Give eyes a catchlight and honest asymmetry. Push skin toward texture, never smoothness. Keep smiles soft to dodge the teeth problem.

And when wording is not enough, reframe so the hard detail is not the star of the shot. Fold these fixes into a reusable base and every portrait you make starts closer to believable. The last five percent is where realism lives, so that is where your prompt should spend.

Generate photoreal portraits in Apatero, first generation free →