LookLightweight
Photogrammetry & NeRF
A 3D model rebuilt from photos.
Turning photos of a real object into a 3D model: take a few dozen pictures around it, and software rebuilds the shape. The photos decide the quality; the web work is shrinking the result so it loads fast.
- Breaks when
- capture quality decides everythingmeshes need heavy decimation for web
- Also called
- scan, NeRF, Polycam, Luma
Prompts
Capture brief for a NeRF
Write the capture protocol for a hand-sized object to be reconstructed with nerfstudio and exported as splats for the web: 80–120 photos, two orbit heights (15° and 40°), overlap 70%, fixed exposure, matte turntable, no motion blur, a marker for scale. Then the exact ns-process-data and ns-train commands and the export to .splat.
Decimated scan on a page
Given a photogrammetry mesh (2M triangles, 8K texture), produce the web version: decimate to 60k triangles preserving silhouette (quadric), bake the texture to 2048 with a normal map from the high-poly, export glTF with Draco. Load in Three.js with MeshStandardMaterial and a RoomEnvironment. Target 3 MB total.
Show the cameras
Visualize a reconstruction: render the recovered mesh and the 96 camera frustums from the COLMAP output (cameras.txt/images.txt) as small pyramids in ink; on hover a frustum shows its photograph as a plane at the image plane. Animate the cameras appearing in capture order.
Related
Resources · 1
- nerfstudioTraining side, not web; exports splats for the browser.nerfstudio-projectGitHub · codeJun 18, 2026