对应工作流（因哩布节点更新滞后，建议下载使用）： https://www.liblib.art/modelinfo/e5eb9b1761fe493e860ffbde4fbfd81e

模型：Wan2.1-I2V-14B-480P

推荐搭配Teacache加速、Wan Video SLG以及Wan Video Enhance-a-video节点使用

建议lora强度：0.8-1

建议步数：30

建议分辨率：480*832   832*480 其他尺寸亦可

建议生成帧数：41

触发提示词：

1.“某个视频主体” first shakes violently, accelerating from slow to fast, then rapidly shrinks and bursts open, splitting into many clones.

2.“某个视频主体“ first shakes violently, accelerating from slow to fast, then rapidly shrinks and bursts open, splitting into many “其他物体“.

这两个区别在于最后分裂成的物体，1是分裂成和自己一样的克隆体。 2可以分列成其他物体

示例：

1.生成结果参见手机GIF：

A smartphone is on a white desk.The smartphone first shakes violently, accelerating from slow to fast, then rapidly shrinks and bursts open, splitting into many smaller-sized small ones.

2.生成结果参见柠檬GIF：

A lemon on a wooden desk surface.The lemon first shakes violently, accelerating from slow to fast, then rapidly shrinks and bursts open, splitting into many cute dogs.

训练相关：

训练集：

10个渲染的抖动后分裂复制视频，包括不同主体背景，长度均为3秒，每秒16帧

训练参数：

重复repeat: 10  

训练轮数epoch：40

训练尺寸：288*288*41F

在云平台上使用2个4090进行训练，总训练时长为3小时左右