A 3D Gaussian splat of the house, trained from the same drone photos as the mesh.
1.45M Gaussians trained for 30k steps with Brush on the 33 drone photos, reusing the COLMAP camera solve from the mesh pipeline. View-dependent spherical-harmonic color was dropped and the cloud was pruned of faint and oversized splats to get the download to a web-friendly size; it still loads the whole scene up front, so the first load takes a moment.
The same 33 photos were also used to train a NeRF — which came out much worse, and is kept up as a comparison.