Gaussian Splat Playground

Scanned, trained, and dissected in the browser: editing the raw Gaussians live. A private side project, July 2026

Marcel Padilla
ETH Zürich

Hi! I was tinkering with GS because I had asked myself how small simple manipulations affect the rendering. To my surprise, the quality of the rendering remained really good. I decided to post this here in case anyone else asks themselves the same question. This started as something I had to figure out for a research project.

GitHub Code on GitHub

Controls

Everything runs in the browser on the raw Gaussians. Pick a scene, or open your own .splat or .ply (up to 250 MB, read on your machine and never uploaded).

Adjustments

  • Spin: turntable rate.
  • Hue: rotates the dominant colour about the gray axis.
  • Amount: wave amplitude. Off also disables Period and Freq.
  • Period: seconds per oscillation. Off freezes the bend.
  • Freq: waves across the object. Off bobs it as one.
  • Melt: curl noise advection.

Histogram

  • Object: fly back to the scanned object.
  • Size: RMS extent, from the covariance trace.
  • Brightness: luminance of the splat colour.
  • Opacity: the stored alpha.
  • Anisotropy: how far from a round blob.

Each splat is one sample: the object is rearranged into the chart of its own statistics, then flies home.

Acknowledgments and credits

This page is glue on top of other people's open-source work. The viewer is built on GaussianSplats3D by Mark Kellogg (splat sorting and rendering) and three.js (scene graph and camera), both MIT.

The tarot-card scene was scanned here from a phone video with FFmpeg, COLMAP for camera poses, and Brush by Arthur Brussee to train the Gaussians. All are open source and free to use. The technique is 3D Gaussian Splatting (Kerbl et al., SIGGRAPH 2023). The other scenes are Creative Commons captures from SuperSplat; each author and license shows in the viewer on selection.