A hand is a small object with a complicated shape
An image generator can make a convincing face, jacket, and background, then give the subject a thumb that grows from the wrong side of a palm. The mistake is noticeable because people have a strong sense of how hands connect and move. The model is not building a miniature skeleton and photographing it. It is producing image patterns that its training makes likely under a prompt.
Hands are particularly difficult because they change shape so much. Fingers bend, overlap, disappear behind objects, and reappear from unfamiliar angles. A fist, a hand holding a mug, and a hand partly inside a pocket share anatomy but show very different visible outlines. One photograph often leaves several joints hidden, so the model has to infer structure it cannot directly copy from the visible pixels.
Why “five fingers” is not enough
Counting fingers sounds like a simple rule. A real hand can show zero, two, or five fingertips depending on pose and occlusion. A finger behind a cup has not vanished; a visible skin-colored patch may belong to a thumb, wrist, or another person’s hand. The generator must maintain a consistent three-dimensional arrangement while drawing a two-dimensional view.
Diffusion image models refine noisy images through many steps. They can learn common arrangements from enormous image collections, but a prompt such as “exactly five fingers” does not force every local detail to satisfy anatomical constraints. The model may satisfy the broad scene—person holding a cup—while leaving one contact point ambiguous or duplicating a digit as it resolves the texture.
The difficulty extends beyond the hand itself. A hand gripping a pen must wrap around the pen in a plausible way. Fingers must meet the object at sensible places; shadows, pressure, and perspective must agree. If the hand and the object are generated as visually compatible patterns without a reliable physical model of contact, a small inconsistency can reveal the illusion.
The training image may not be a clean anatomy lesson
A training photo may show a hand partly hidden behind an object or another finger. That is an incomplete view of its anatomy. The HandRefiner study identifies deformation and occlusion as difficulties in learning hand structure from images, and tests a correction method guided by a reconstructed hand mesh. This supports a geometric explanation for some errors; it does not establish one cause for every model or image.
Research on improving generated hands has explored conditioning models on hand pose and separating hand generation from the rest of the human image. The need for such targeted methods illustrates the problem: generic image quality does not guarantee reliable geometry in a difficult region.
A crop can also change the outcome. If the wrist is outside the frame, the generator has fewer visible constraints on how the hand attaches to the arm. If two hands overlap, ownership of a finger can become ambiguous. A low-resolution image may conceal an error that becomes obvious when enlarged.
Why video raises the bar
A still image needs one plausible moment. A moving hand must remain the same hand across frames, keep a consistent finger count, grip the object continuously, and change pose along a believable path. An artifact that flashes for a single frame may be hard to notice at speed, but temporal inconsistency can make generated video feel unstable.
Targeted controls have reduced hand errors in the evaluations reported by these studies. There is no fixed rule that AI always draws hands badly; some images are anatomically convincing. The useful question is whether a particular image or sequence is trustworthy at the detail level you need. A thumbnail may only need a plausible gesture. A medical illustration, product instruction, or safety diagram needs accurate anatomy and contact, which calls for expert checking.
Read the image at two levels
First inspect the overall scene. Then follow the geometry: connect each visible fingertip to a palm and wrist, decide which parts disappear behind the object, and check whether the grip and shadows agree. Jewelry, sleeves, reflections, and overlap can make that tracing difficult. Look for impossible bends or duplicated tips rather than relying on a fingertip count alone. This is not a foolproof test for whether an image was generated; photographs can be confusing and newer models can produce correct hands. It is a way to check a detail that visual fluency often hides.
The larger lesson is about image AI generally. Producing a coherent-looking picture is different from maintaining an explicit, physically consistent model of everything inside it. Hands make that difference visible because their structure is familiar and unforgiving.
