Diamond Match Factory — A Real Building, Rebuilt From One Drone Video.
A single DJI drone, one continuous helix orbit of the Diamond Match Factory in Chico, California — and from that one video, a fully walkable 3D Gaussian-splat reconstruction of the real building. No LiDAR, no rig, no studio capture: just photogrammetry and a real flight.
The source is the orbit footage above — a helix flight that rises while it circles, so the camera sees every façade from multiple elevations. That coverage is the whole trick: a flat circle leaves the vertical surfaces ambiguous, but a helix gives the solver the parallax it needs to resolve window lancets, cornices and brickwork. We pull sharp frames from the flight, then run them through a classical Structure-from-Motion pass.
COLMAP solved all 407 of 407 cameras — a clean 100% registration, every frame placed in 3D space with the sparse point cloud to match. That posed dataset feeds Brush, a WGPU Gaussian-splat trainer, which optimises millions of splats into a continuous radiance field you can fly through. The result is a walkable world reconstructed from a real place, rendered in a Three.js viewer — the third local splat rail in the OMNI Labs creative-tech stack.
Frames captured live inside the walkable Gaussian-splat viewer — the same scene you can fly through, posed at varied angles from hero three-quarter to near top-down.
The Helix Orbit
One continuous DJI flight, rising as it circles the building — a helix, not a flat ring. The ascent gives the camera multiple elevations on every façade, which is exactly the parallax a single-orbit capture lacks. Architectural detail — window heads, cornices, buttresses — only resolves cleanly when the surfaces are seen from varied angles.
COLMAP — 407/407 Cameras
Sharp frames are extracted from the orbit and run through COLMAP Structure-from-Motion. Every frame is matched, the cameras are triangulated, and a sparse point cloud is reconstructed. The solve landed at a full 407 of 407 cameras registered — a 100% clean alignment with no stray two-image fragments, the sign of well-conditioned footage.
Brush — Splat Training
The posed dataset feeds Brush, a WGPU Gaussian-splat trainer that runs entirely on local GPU — no CUDA-toolkit compile, no cloud. It optimises millions of 3D Gaussians into a continuous radiance field that captures the brick, glass and street context of the real building.
The Walkable World
The trained splat loads into a Three.js viewer where you can fly the camera through the reconstruction — a real place, navigable in the browser. It's the video-to-world counterpart to our generative single-image splat rail: that one dreams a world, this one faithfully rebuilds one that exists.