Why does AI give people six fingers?

A hand has no fixed appearance to learn. It articulates into thousands of configurations, occludes itself constantly, and appears small in most training images, so the model learns a strong sense of hand texture and a weak sense of hand structure. Finger count is structure, and structure is the part it did not learn well.

Why do AI faces look almost right but slightly wrong?

Because you are an expert at faces and a novice at hands. Humans read faces with dedicated neural machinery tuned to millimetre asymmetries, so a face can be anatomically correct and still read as wrong, while a badly built elbow passes unnoticed. The model is not failing worse on faces; you are simply much better at catching it.

Does writing perfect hands in the prompt help?

No. The failure is structural rather than stylistic, so there is no quality dial for the prompt to turn up. What helps is giving the hand more pixels and less to do: frame it larger, reduce how many hands are in shot, and avoid interlocking fingers or hands gripping complex objects.

Guide

Why hands and faces come out wrong

They are the two things everyone notices, and they fail for completely different reasons — one is a gap in what the model learned, the other is a gap between the model and you. Only the first has a technical fix.

Hands

A hand has no fixed appearance to learn

Almost every other body part is close to rigid. A nose points one way, an ear has one silhouette, a shoulder bends at one joint. A hand has roughly twenty degrees of freedom and folds into effectively unlimited configurations, most of which hide part of itself behind another part.

So a model trained on photographs sees the word hand attached to millions of images that share almost no geometry. What it learns well is the texture of a hand — skin, nail, knuckle, tendon. What it learns badly is the structure: that there are exactly four fingers and one thumb, that they attach in a particular order, and that they bend only one way.

Finger count is structure. That is why the classic failure is not a blurry smudge but a confident, well-lit, correctly shaded hand with six fingers on it. The part the model knew, it rendered beautifully.

Scale compounds it. In most training photographs a hand occupies a tiny fraction of the frame, so there were few pixels to learn from and there are few pixels to render into. A detail that small is being reconstructed from a weak signal.

Faces

The model is not doing worse — you are doing better

Faces are the opposite problem. Structurally they are among the easiest things in the picture: one configuration, bilaterally symmetric, always oriented roughly the same way relative to the head. Models render them very well.

The trouble is that you have dedicated machinery for reading them. A substantial region of human visual cortex does little else, tuned over a lifetime to differences of a millimetre or two in the spacing of features, because telling people apart and reading their mood has always mattered.

Nothing comparable exists for elbows. A badly built elbow passes unnoticed in a picture you would reject instantly for a face that is a fraction too wide between the eyes. The failure rate is not higher on faces; your detection rate is.

This is also why a face can be flawless and still feel wrong. There is no error to point at, which is the maddening part — the geometry is plausible, and the sense of wrongness arrives from a system that does not report its reasons.

What helps

Pixels and simplicity, not adjectives

Give the hand more pixels. Generate larger, or frame tighter. A hand occupying a meaningful share of the image has far more capacity spent on it than one that is forty pixels across.

Give the hand less to do. Interlocked fingers, a hand gripping a complex object, and two people holding hands are the three reliable disasters, because each demands exactly the structure the model learned least. A hand at rest, in a pocket, or out of frame cannot fail at all.

Reduce how many hands are in shot. Each is an independent chance to go wrong, and errors are noticed per hand rather than per image.

Fix afterwards, not upstream. Regenerating only the hand region at full resolution, with the rest of the image held, is the technique that actually works. It is a second pass rather than a better prompt.

What does not help is asking. Phrases like perfect hands or five fingers are descriptions competing for the same finite attention as everything else in your prompt. There is no correctness dial for them to turn, for the reason at the top of this page: the model did not learn hand structure and then choose to ignore it.

For faces there is one genuine fix, and it is a different kind of answer: stop asking the model to invent a face each time. A face conditioned on a fixed reference is being copied rather than imagined, which is both how consistency works and why it tends to look better.

Asked often

Short answers

Why does AI give people six fingers?: A hand has no fixed appearance to learn. It articulates into thousands of configurations, occludes itself constantly, and appears small in most training images, so the model learns a strong sense of hand texture and a weak sense of hand structure. Finger count is structure, and structure is the part it did not learn well.
Why do AI faces look almost right but slightly wrong?: Because you are an expert at faces and a novice at hands. Humans read faces with dedicated neural machinery tuned to millimetre asymmetries, so a face can be anatomically correct and still read as wrong, while a badly built elbow passes unnoticed. The model is not failing worse on faces; you are simply much better at catching it.
Does writing perfect hands in the prompt help?: No. The failure is structural rather than stylistic, so there is no quality dial for the prompt to turn up. What helps is giving the hand more pixels and less to do: frame it larger, reduce how many hands are in shot, and avoid interlocking fingers or hands gripping complex objects.

Keep reading

The rest of the series

Where this comes from

Unlimited Studio is built around this one problem

Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.

How the pipeline works What it costs