Why does the skin controller have INV_BIND_MATRICES?

Can these not be computed directly by traversing the joint graph and inverting the world transforms in the visual_scene pass? Or is there additional information that these matrices encode that I lose by computing the bind pose from the graph? It just seems like an odd place to put the inverse bind matrices, considering that we have the information at the joint level; unless I’m missing something.

Thanks!
Zach.