Why not add clip embedding to denoising-unet in high-resolution block?

[Naive-Bayes]

Why only add clip embedding to denoising-unet in low-resolution block?

In AnimateAnyone, as the Figure 2 shows, they add clip embedding in all blocks. And in Champ Fig 2, it seems only add clip embedding in low-resolution blocks.

Do you make ablation about this design? And any conclusion?

By the way, I have not found the command to remove the clip embedding in high-resolution blocks. Maybe it's a figure mistake? Sorry if I was wrong.

######################################################

论文的Fig 2中，clip的embedding貌似只在降噪UNet中间的block上加，但Animate Anyone在所有的block中都注入了clip embedding，这样设计的原因是什么，有什么结论表明这样做的好处吗？

还有就是在code中，我好像没有找到不注入clip embedding的开关，能告诉在哪里移除clip embedding吗？