Arrival.Space has open-sourced Animated SOG, a file format and toolkit for playing 4D Gaussian splat clips in the browser.
One .asog file holds a clip of moving Gaussians with spherical harmonics and, optionally, its soundtrack. Each splat lives for a short time and moves rigidly over its life, the motion model of the TRiGS paper. Static attributes are stored the way SOG v2 stores them, with 16 bit log positions, smallest three rotations, 256 entry codebooks and an SH palette, so a SOG renderer only has to add the motion. Animated SOG stores the motion in an exact 17 value form, quantized.
The repository carries a Python encoder, reference decoder and asog command line tool, plus a streaming JavaScript decoder that works while the file downloads. Players cover PlayCanvas on WebGPU and WebGL and three.js on WebGL 2, with motion and SH evaluated on the GPU, and rely on a few PlayCanvas 2.22 engine internals.
The encoder takes TRiGS clips written by Open4DGS. A flip book pipeline converts folders of independently trained per frame 3DGS PLYs, which requires a CUDA GPU and the external Open4DGS trainer, an AGPL-3.0 tool outside the repository.
By Arrival.Space's measurements on one RTX 5080 desktop, .asog files are 11.1x to 11.5x smaller than their sources across the six public TRiGS editor sample clips. The 1.29 million splat Charlie clip, 319 MB as source, measures 45.3 dB PSNR against the original at the default 30 percent pruning. A 900k splat .asog of Charlie decodes in 0.9 seconds in the PlayCanvas player. A six second flip book of a standing person, 3.6 GB as raw PLYs and 202 MB as per frame SOG, became an 8.6 MB .asog after a 23 minute distillation, measuring 41.0 dB on held out cameras.
Animated SOG follows Splat.js, Arrival.Space's in browser splat trainer. Arrival.Space says it is integrating the format. The README points to animated-sog packages on PyPI and npm, but neither registry lists them yet, so the code installs from the repository.
The code is available on GitHub under the MIT license.





