An image generator has to be right once, a video generator has to be right thousands of times in a row
Generating a single image only requires the model to produce one coherent, plausible frame. Video multiplies that difficulty enormously, because a model now has to produce dozens of frames per second that are each individually coherent and, critically, consistent with each other over time. This second requirement is called temporal consistency, and it is the central hard problem in generative video.
Without careful handling, a naive approach of generating each frame independently produces flickering, morphing objects, backgrounds that subtly shift, and faces that drift into different people from one frame to the next, because nothing forces frame 47 to agree with frame 46 about what a character's shirt or a building's shape actually looks like. The model isn't just generating images anymore, it's implicitly modeling physics, object permanence, and motion, none of which it was ever directly taught, only inferred from patterns in training video.