But isn't this what the author is doing already? That's what I got from the video. SDF is sampled on a sparse grid (only cells that cross the level set 0) and then values are sampled by interpolating on the grid rather than full reevaluation.