Gaussian Splatting

Tap it: random splats learn the photo. Pick another View in the Splash tab.

Gaussian splatting draws a scene with many soft, colored blobs called splats. Each one has a place, a size in three directions, a turn, a color and an opacity. A scan's splats are trained the way a network learns: start from a rough cloud of points (worked out from the photos, or random), compare the picture it makes with the photos, and nudge every splat a little, again and again, until they match.

Tap it: a cloud of random splats slides, stretches and recolors into the strawberry photo, a real fit of 2,400 splats, while its error falls on the chart. Other views show one splat with its three axes, which a tap trains: it starts from a wrong guess and gradient steps move its place, turn, sizes, color and opacity toward the target outline, big steps first and then smaller ones, as in a real fit; a duck whose splats shrink to dots, and the back-to-front order splats are drawn in.

The photo view uses a real optimizer on 2,400 flat Gaussians, with checked compositing and loss derivatives. It is a single-image 2D fit, not a multi-view 3D reconstruction. The one-splat motion is prepared, and the sorting view has a reproducible distance-versus-depth error.

1. Correct

   A photo is fit using 2,400 optimized Gaussians.¹²

   The production fit count is 2,400. Independent pixel compositing agrees with the loss; finite differences agree with all 18 tested derivatives, and a small fit reduces its loss.

2. Simplified

   The training picture is a single-image 2D version of Gaussian fitting.¹²

   It optimizes position, two sizes, angle, color, and opacity in a flat photo plane, using Adam and cached frames; it does not optimize full 3D covariance from multiple views.

3. Simplified

   The one-splat view illustrates optimization with a prescribed path.³¹

   A shared factor (1−rate) to the step count reduces every target-parameter error for 24 steps, then forces zero. There is no rendered-image loss calculation in this view.

4. Being fixed

   The sorting view shows the renderer’s back-to-front order.⁴

   The bead order is sorted by Euclidean distance to a fixed camera rather than view-axis depth. The test finds pairs whose order disagrees with depth.

1. 13D Gaussian Splatting for Real-Time Radiance Field Rendering, Kerbl, Kopanas, Leimkühler, and Drettakis, Inria
2. 2Adam: A Method for Stochastic Optimization, Kingma and Ba
3. 3Gradient descent, Dive into Deep Learning authors
4. 4Gaussian rasterizer depth, Kerbl and coauthors, official implementation

Simplified

The photo fit uses a reduced-resolution image and 2D parameters, caches its optimization, and replays keyframes. It does not infer a 3D scene from several cameras or adapt the splat count. The one-splat path follows prescribed parameter-error decay and snaps to the target at the end.

This link opens it in the Playground. To put it on your own page, add the element and one script tag. Every option.