整理数据集提示词时，把 pan、tilt、tracking、close-up 等镜头词与主体动作分列。这样可以重组出同一动作的多个镜头版本，用于测试模型是否真的理解运镜。 When organizing captions, put camera terms such as pan, tilt, tracking, and close-up in a separate field from subject actions. You can then generate multiple camera versions of the same action and test whether the model understands movement.