图生视频先把参考帧里不能变化的内容写清楚，例如主体身份、服装和构图；提示词只增加时间上的动作。这样更容易判断是运动不稳定，还是参考帧约束不足。 For image-to-video, state what must remain fixed in the reference frame, such as identity, clothing, and composition. Use the prompt mainly to add temporal motion so instability can be diagnosed separately from weak conditioning.