It's really cool, but I can't shake off an "uncanny valley" feeling, with all of the small quirks in the geometry. What I think I'd be interested in is a post-processing step where this splat is automatically converted to a 3d model that approximates each component, only falling back to the point cloud if there's no simple shape that fits the observation at a particular location.
This is close to the idea of convex splatting (recent paper) in which convex shapes are used to approximate these real 3d objects as they are better suited than gaussians
