Why does AI get hands and faces wrong, and how do I fix it?

It's the defining AI-image tell, and the usual explanation — 'the model doesn't understand anatomy' — is close but not quite it. The model doesn't understand anything. What matters is how much reliable signal a hand provides during training compared to how much variation it has to explain.

Why hands specifically

• —They're tiny in frame. In a typical full-body shot a hand occupies well under one percent of the pixels — a handful of pixels per finger at generation resolution. Faces get orders of magnitude more.
• —Their appearance varies enormously. A face is one topology from a limited set of angles. A hand is five articulated digits that can form thousands of configurations, and looks completely different from each one.
• —They occlude themselves constantly. Fingers overlap fingers. In training data the model repeatedly sees four visible digits, or three, or seven across two overlapping hands — so 'a plausible number of fingers' is a genuinely fuzzy target.
• —They're usually not the subject. Photographers focus on faces, so hands are disproportionately soft, motion-blurred, or partially out of frame in the training set.

Faces fail differently and less often. Frontal, well-lit, medium-distance faces are near-solved. Faces still break at small scale (background crowds), at unusual angles (extreme three-quarter, looking up), and in the uncanny mid-zone where skin texture is nearly right — which reads worse than an obviously stylised face, because your visual system holds real faces to a much finer standard than it holds hands.

Fixes that actually work

Ordered by how reliably they hold up, best first:

1. 1.Give the hands a defined job. 'Hands resting on a ceramic mug', 'holding a folded newspaper', 'one hand in a coat pocket'. A hand doing a specific known thing has a strong prior behind it; a hand floating in space has none. This is the highest-yield change by a wide margin.
2. 2.Simplify the pose. Relaxed, partly closed, or gently curled hands are far more forgiving than splayed fingers or complex gestures. Never ask for pointing, counting, or a peace sign and expect it to survive.
3. 3.Crop them out. If hands aren't the point, frame at head-and-shoulders or three-quarter length and end the composition above the wrists. There is no rule that a portrait must include hands.
4. 4.Occlude them naturally. Pockets, long sleeves, behind the back, gloves, held objects, a hand out of frame at the edge — all of these are ordinary photographic reality, not evasion.
5. 5.Increase their size in frame. If hands must be visible and correct, make them the subject: a close-up of hands on piano keys is far more likely to succeed than the same hands in a wide shot, because the pixel budget finally matches the complexity.
6. 6.Change models. Genuinely different architectures fail differently on the same prompt — the roll is not the same roll.

What doesn't work

Adding 'perfect hands, five fingers, detailed hands, anatomically correct' to your prompt. These read as emphasis on hands, which frequently makes the model render them larger and more prominently — a bigger, more visible failure. 'Bad hands, extra fingers, mutated hands' in a negative prompt is somewhat more defensible but still weak: you're steering away from a defect class the model has no clean internal representation of. Composition beats incantation here, consistently.

A note on the current generation of models

This is genuinely better than it was two years ago. gpt-image-2 and Nano Banana 2 both produce correct hands in ordinary poses most of the time, and the catastrophic seven-fingered failures that defined 2023 are now the exception. The remaining failure zone is narrower but real: small hands in wide shots, multiple people's hands interacting, and anything with an unusual gesture. Worth checking against a second model before you conclude a prompt is unusable — see how the same prompt lands across engines on the model pages.