Applies to: All users generating video or images with people or animals.
Use these seven fixes when generated people have distorted faces, extra fingers, warped limbs, or unnatural poses — the most common anatomy complaints with AI video models.
Limit the number of people in the shot. Each additional person multiplies anatomy errors. One clearly described subject is far more reliable than a crowd.
Describe the pose explicitly in the video prompt. "Standing with arms at her sides" gives the video model less room to improvise than leaving the pose unstated. Ambiguity invites anatomical chaos.
Keep hands busy or out of frame. Hands holding a defined object ("gripping a coffee mug") render better than hands doing nothing. Framing choices like "medium shot, hands out of frame" also work.
Use anatomy-targeted negative prompts where supported. A negative prompt is a separate input field, available on some video models, where you list what you do not want to see. For anatomy, use: "bad anatomy, extra limbs, missing limbs, mutated hands, extra fingers, fused fingers, deformed, long neck, bad proportions."
Avoid vague, loaded descriptors. Generic adjectives like "beautiful woman" are a common error — describe concrete features instead (age, build, hair, clothing).
Faces distort more in wide shots. If a face matters, use a closer shot type (medium close-up or close-up). Faces in the background of wide shots are the most likely to break.
If the anatomy is structurally wrong every time, don't endlessly tweak adjectives — simplify the scene or try a different video model, since different video models interpret anatomy through different training biases.