To be fair, if they first vocalize it they could just use Sparse Voxels which is my favorite differentiable rendering technique https://github.com/NVlabs/svraster
To be fair, if they first vocalize it they could just use Sparse Voxels which is my favorite differentiable rendering technique https://github.com/NVlabs/svraster
I find the workflow of fitting Gaussian splats and then immediately turning them into voxels to be rather surprising, too. I can imagine some performance benefits at large scale, or maybe a reduced tendency to generate certain types of artifacts, but going straight to voxels seems more straightforward.