KISS-GS Compresses 3DGS Scenes 228x, From 738 MB to 3.5 MB

For much of the last year, one of the most popular gaussian splatting compression methods has been Self-Organizing Gaussians or SOG. Now, the team that wrote the paper are back with their newest method and they are keeping it simple and small with KISS-GS. 

KISS-GS: 3D Gaussian Splatting Compression Kept Simple takes a 738 MB scene down to 3.5 MB at the same quality, and the author list is the first thing that got my attention. Wieland Morgenstern and Florian Barthel, two of the eight authors from Fraunhofer HHI, Humboldt-Universität zu Berlin and TU Berlin, wrote Self-Organizing Gaussians, the method PlayCanvas shipped in Engine 2.7.5 and the ancestor of most image-based splat compression today. Barthel also built splatviz.

Most compression methods stack several tricks at once and report one headline ratio at the end. The paper calls that the attribution gap. You can't tell which trick bought you what, and you can't lift the good part out and reuse it. KISS-GS runs as separate stages on any vanilla 3DGS .ply and measures each stage on its own.

Stage one is POPSpa, which throws away the Gaussians you weren't getting much from. It combines score-based pruning from GaussianPOP with the alternating optimize-sparsify scheme of GaussianSpa, plus an effective-rank regularizer that discourages needle-shaped Gaussians, erank near 1, in favor of disk-shaped ones, erank near 2. POPSpa alone is 15.7x, 738 MB down to 47.1 MB.

Stage two, SOG-XT, is where the Self-Organizing Gaussians lineage shows. It sorts primitives into a 2D grid so neighbors in space become neighbors in an image, then stores the attributes as ordinary pictures. Spherical harmonics get split, the DC term written out as an 8-bit image and the rest reduced to a k-means codebook whose centroids are themselves sorted into a 2D grid (new with SOG-XT) and indexed by two 8-bit UV channels. SOG-XT adds a further 6.6x, to 7.13 MB.

My favorite piece of the paper is PRAS, Parallel Representative Assignment Smoothing. A single Gaussian's covariance can be written 48 different ways, because you can permute the three scale axes, flip eigenvector directions and negate the quaternion without changing the shape. PRAS picks whichever of the 48 lands closest to a primitive's smoothed neighbors, which lets quaternions quantize at q=99 where q=255 is otherwise needed. 

SOG-XT restricts storage to 8-bit images, WebP specifically, where the original Self-Organizing Gaussians used JPEG XL. You might remember Vincent Woo investigating WebP as a medium in 2024. Metadata is 0.2% of the container, and the encoder needs neither camera parameters nor a renderer. 

An optional fine tuning pass with the codec inside the training loop takes a further massive 2.2x off, landing at 3.23 MB. That is 228x on Mip-NeRF 360 and 319x on Tanks and Temples.

Despite gaussian splatting’s large improvements across the past three years, there is still large room for enhancement across the board. To learn more about KISS-GS, you can check out the paper or you can go to their Project Page. There is no word yet about a code publication or its license, but if it follows the precedent of SOG, the community at large will be very happy.

View the Project Page here.