Explanation of the code "cldm/cldm.py": ControlledUnetModel class

Nobody has claimed this yet.

Thank you for amazing works first.

I have read your paper and have been looking at the implementation, especially the cldm.py
There's one thing I failed to understand that is (I assume) not specified on the paper.

In the ControlledUnetModel class, in the middle of the code definition, there is

if control is not None:
h += control.pop()

So, If a control tensor is provided, it adds it to the output of the middle block?
control must mean, for example, edge map feature maps.
But why is it added to the output of the middle block?

Thank you.

Start with cldm/cldm.py, focusing on the ControlledUnetModel class and the control.pop() use, then compare it with the paper referenced in the issue. Done means providing a clear explanation of the control tensor and why it is combined at that point in the model.

Written by the indexing model from the issue text.