Imagery
Hands in AI images: five fingers on the first try
Hands are the most common flaw in AI images, and they fail for a reason. Say what the hand is doing, what it holds and which fingers show, and the generator stops guessing.
By: Vadász MárkPrompting video models (Seedance, Kling, Veo)
Six fingers, fingers fused together, a thumb growing out of the wrong side. If you've made an AI film, you've seen it. Hands are the most common flaw in generated images, and because the video is built from the start frame, a bad hand travels through the whole shot.
The short answer: hands go wrong when the prompt doesn't tell them what to do. Give them an object, describe the grip, or keep them out of frame. You rarely need more than that.
Why it's always the hands
A hand takes up little of the frame yet carries a lot of detail. Fingers hide one another, bend differently with every grip, and the generator has seen hands in countless positions in its training images. If you don't say which one you want, it assembles an average, and an average easily ends up with six fingers.
A face is different: two eyes, one nose, one mouth, roughly in the same place every time. Hands have no fixed layout like that, so a precise description pays off here more than anywhere else in the prompt.
Give the hand a job
Our planner's rule is simple: a hand either holds an object or stays out of frame. A mug, a steering wheel, a knife, a phone, a coat pocket all give the hand a shape. Around an object the fingers have only a few possible positions, so the generator has far fewer chances to get it wrong.
If the scene doesn't need the hands, don't force them in. In a medium close-up they can sit in a pocket, below the frame or behind the character's back. That isn't cheating. It's composition.
Medium close-up, eye level, 50 mm lens: a man in his forties in a charcoal wool coat waits at a bus stop in light rain, framed from the chest up, both hands pushed deep into his coat pockets, raindrops on his shoulders, overcast evening light, photorealistic, fabric with real weave and natural creases
What to avoid
A few hand positions fail almost every time. Cut them at the script stage, before you write a prompt for them.
- Counting on fingers: the generator doesn't know how many fingers it should show, and it won't get it right.
- Interlaced fingers: ten fingers woven together is where most fused stumps come from.
- Empty gestures in the air: waving or pointing with nothing to hold, so the hand has nothing to anchor it.
- Two people's hands together: in a handshake or a hand-in-hand walk the generator mixes up whose finger is whose.
Three things to write into the prompt
If the hand is in the frame, answer three questions. That's what makes the start frame deliberate rather than the generator's idea.
- What the hand doesA verb, not a state: wraps, grips, turns, presses. Saying the hand is next to the mug isn't enough.
- What it holdsName the object and its material, because the grip depends on it: you hold a heavy ceramic mug differently from a paper cup.
- Which fingers showSay what hides the fingers: thumb on the rim, the rest behind the mug. What doesn't have to be drawn can't be drawn wrong.
Close-up, eye level, 85 mm lens: a woman in her late thirties wraps both hands around a heavy white ceramic mug on a wooden kitchen table, both thumbs resting on the rim, the other fingers curled around the far side of the mug and hidden behind it, morning window light from the left, anatomically correct hands with five fingers each, natural knuckle creases, short unpainted nails, photorealistic, pore-level skin detail, no beauty retouch. avoid: extra or fused fingers, waxy or plastic skin, airbrushed look
Whatever the object hides, the generator doesn't have to invent.
Which model, and where the negatives go
For a start frame with a person and their hands, our planner picks Nano Banana Pro: it makes the fewest hand errors. An image prompt can end with a short avoid list that includes "extra or fused fingers". Image models handle that well.
Video is different. If the hands are working (kneading, cutting, knitting), Seedance moves them most naturally, but you don't give it a negative list. In Kling the negatives go into a separate field, never into the prompt. With both, the positive description does the work: what the hands do, and that they stay visible throughout.
A baker's hands knead a ball of dough on a floured wooden counter. The camera holds a static top-down close-up for the full five seconds. Her palms push the dough forward, fold it back towards her and press again in a continuous rhythm, flour dusting her knuckles. Both hands stay fully in frame, five fingers each, the dough always between them.
If it still goes wrong
Don't regenerate the same prompt and hope. Rewrite only the affected prompt, name the problem in it concretely ("anatomically correct hands with five fingers each"), and leave the character description and the scene world untouched. If it broke in the video, simplify the movement too.
An example. Before: "a woman waving at the camera, cinematic". After: "a woman raises her right hand to shoulder height holding a set of house keys, palm towards the camera, thumb and index finger pinching the key ring, the other three fingers folded into her palm, anatomically correct hand with five fingers". Same gesture, but now it has something to hold on to.
Before your next prompt, go through the shot list: in every frame with a hand in it, is there an object and a verb beside it? Where there isn't, either add them or frame the hand out.
Sources
- Lu, W. et al. (2023): HandRefiner: Refining Malformed Hands in Generated Images by Diffusion-based Conditional Inpainting, arXiv
- Google DeepMind: Gemini 3 Pro Image (Nano Banana Pro)
- fal.ai: Kling 2.1 Master Text to Video API, the negative_prompt field
Common questions
Why do AI images get hands wrong, even adding extra fingers?
A hand is small in the frame but made of a great many details: fingers cover one another and bend differently with every grip. If you do not say which pose you want, the generator builds an average, and an average can easily end up with six fingers. A hand has no fixed layout like a face, so it needs a precise description.
How can I get hands right on the first try?
Give the hand an object: either it holds something, or it is not in the frame. A mug, a steering wheel or a coat pocket gives the fingers a shape, so there are fewer ways for them to go wrong. If the scene does not need the hand, hide it in a pocket, below the bottom edge or behind the back. That is composition, not cheating.
Which hand poses should I avoid in AI generation?
Four poses almost always go wrong: counting on the fingers, clasped fingers, a gesture in mid-air with no object, such as waving, and two people's hands together. In a handshake the generator mixes up which finger belongs to whom. It is worth striking these from the script before you write any prompt.
What should I do when the generator ruins the hands?
Do not regenerate the same prompt and hope for the best. Rewrite only the affected prompt and name the fault specifically, for example that the hand has five fingers, while leaving the character and scene description unchanged. In video, simplify the movement too. State what the hand holds and which fingers are visible.