Gaussian splatting
A photoreal 3D capture made from millions of soft colored blobs instead of a polygon mesh. It looks solid from the camera, but has no surfaces.
See it
What it is
Gaussian splatting represents a captured scene as a cloud of soft, colored 3D ellipsoids called splats. Each splat stores a position, size, orientation, opacity, and view-dependent color. The renderer projects those ellipsoids onto the screen and blends them front to back, producing a solid-looking image without triangles or surface textures. The modern reference is the 2023 3D Gaussian Splatting paper by Bernhard Kerbl and collaborators.
Reach for it when you want a photoreal, navigable capture of a real room, object, or outdoor location from many overlapping photos. Phone apps such as Luma AI and Polycam run the capture and training end to end. It can render trained captures in real time and preserves small lighting and surface details that are painful to rebuild as a mesh. It is less useful when the result must be rigged, edited face by face, given collision, or fitted with ordinary UV maps.
Gotcha: the reconstruction only knows what the cameras saw. Thin objects, reflections, moving people, and unseen angles can turn into floaters, holes, or cloudy streaks. A splat scene can also contain a large number of primitives, so file size, sorting, and overdraw still matter even though there is no polygon count.
Ask AI for it
Turn this calibrated photo set into a 3D Gaussian Splatting scene. Use COLMAP to solve camera poses, train Nerfstudio's splatfacto model with spherical harmonics for view-dependent color, and prune low-opacity floaters after densification. Export the trained splats as PLY, then build a three.js viewer with the GaussianSplats3D renderer that loads the capture progressively, sorts splats by camera depth, and provides orbit controls. Include a quality preset for mobile that reduces the rendered splat count.