A 3D scene can be millions of anisotropic Gaussians: each has a position, a covariance (“blob shape”), a color, and an opacity. Rendering splats those blobs onto the image plane and alpha-blends them. The pipeline is fast, differentiable, and strong at novel views.
Training starts from multi-view photos plus camera poses. The optimizer moves, reshapes, and recolors the blobs until rendered views match the photos.
| Approach | Representation | Query |
|---|---|---|
| 3D Gaussian Splatting | Explicit particle cloud | Project / splat to screen |
| NeRF | Continuous field | Shoot rays; sample along each ray |
Splats are editable as objects in space. A NeRF is a function you query; editing it usually means changing the field, not moving a particle.
A splat scene is the explanation: you walk around a persistent, renderable 3D account of what the cameras saw. The idea lives in the artifact, not only in a paragraph about the artifact.
| File | What it does |
|---|---|
two_d_splat.html | Three 2D Gaussians alpha-blended on a grid — drag blobs, change radius/opacity |
# Browser — open locally (no server needed)
# Linux: xdg-open gaussian-splatting/two_d_splat.html
# macOS: open gaussian-splatting/two_d_splat.htmlThe scene is an explicit list of blobs (position, radius, color, opacity) splatted onto a canvas. Move one blob and the render updates — you edited an object in space.
NeRF stores a continuous field: you query density/color along rays. Editing means changing the field, not grabbing a particle. 3D Gaussian Splatting keeps particles you can optimize, split, or delete. This 2D toy shows why splats feel tangible: the representation is the things you see.