I hadn't seen that paper, looks really interesting! It definitely would improve the resolution but I think fundamentally it still runs into the issues I had with voxels where it mostly just learns low level correlations and not high level topology.