Why not use coarse rendering loss?

[bring728]

In the existing Mip-NeRF paper, loss is calculated for both coarse rendering and fine rendering. EG3D performs importance sampling like Mip-NeRF, but does not calculate loss for coarse rendering. But it seems to work well nonetheless.

I think that in order for importance sampling to work well, we need to calculate loss for coarse rendering as well. Why does EG3D train well even without this loss?