Current models still have relatively low accuracy in executing motion instructions such as pan left, and keyframe control and skeleton-guided generation are important directions for improving predictability.