Hi Kyle,

I'm implementing a iterative reconstruction with a large learnable voxel grid (906×906×320), which results in slow reconstruction. However, most of the voxels are empty, and I’ve managed to obtain a volume mask (as shown in the demo you provided).

I’m wondering if it’s possible to accelerate the computation by using the volume mask—specifically, by skipping the forward and backward projections (i.e., loss gradient computation) in the empty regions where the mask is zero.

I noticed in the set_volume_mask function comments that the volume mask doesn't improve speed, but I’m curious if there’s any way to modify the pipeline to make this feasible. Any advice or insights would be greatly appreciated.

Thanks in advance!