House NeRF

A neural radiance field of the house, trained from the same drone photos as the Gaussian splat — and a good demonstration of where the technique breaks down.

This one didn't really work. A NeRF needs dense, overlapping coverage — typically a few hundred photos. Trained on just the 33 aerial shots, the ground geometry only half-resolves and the open sky fills with drifting semi-transparent “floaters.” It's kept here as an honest side-by-side: the Gaussian splat was built from the exact same 33 photos and came out clean. Radiance-field MLPs are far less forgiving of a sparse capture than splats are.
The viewer downloads about 75 MB of baked texture atlases before the first frame — give it a moment.

Drag to look · W/A/S/D to move · needs a desktop browser with WebGL2 and a discrete GPU

Trained with the torch-merf PyTorch implementation of MERF (Memory-Efficient Radiance Fields) — 20,000 iterations on an RTX 5060 Ti, reusing the COLMAP camera solve from the mesh and splat pipelines.

MERF bakes the trained network into a stack of texture atlases (a sparse voxel grid plus three feature planes) so it can be ray-marched in real time in the browser with plain WebGL2 — no network inference at view time. That baking step is what makes a NeRF hostable as a static page at all; the price is the ~75 MB of PNGs it downloads up front.