Early generative video models are prone to problems such as anatomical errors, limb interpenetration, and incoherent motion trajectories when handling two semantically unrelated elements (such as a bird and a bicycle) and actions requiring physical plausibility (such as riding).