Gaussian-Splat-Lite v1.7.0 cuts splat memory and changes rotation encoding

Version v1.7.0 of the Gaussian-Splat-Lite renderer has been released. In the changelog, the author describes moving the RAD and SOG models to a shared SH codebook instead of storing SH coefficients per splat, including during streaming loading. CPU and GPU memory for splat attributes drops to 25%, 43.75% and 62.5% for SH1, SH2 and SH3 respectively; the actual savings depend on the codebook size and do not account for other rendering and loading overhead. Splat rotation is now stored in 32-bit smallest-three quaternions. Existing model files remain compatible, but live viewers need to be reloaded after the update.

What it means

Gaussian-Splat-Lite is an open-source 3D generator under the Apache-2.0 license, a 3D Gaussian Splatting renderer for Three.js with WebGPU and WebGL2 support. It works with the PLY, SPZ, SOG and RAD formats, including streaming LOD for RAD, SOG and LCC2. Release v1.7.0 affects the internal representation of splats: a shared SH codebook instead of per-coefficient storage reduces memory consumption, and smallest-three quaternions improve rotation precision. The author removed alpha, color and sh from the postDecode inputs and outputs — now only position, scale, quaternion and opacity are available on load, and transparency and color are meant to be changed via opacity and mesh properties or SDF edits. For reference users, this means RAD and SOG scenes fit more detail into the same memory budget, and files created earlier will continue to open. The encodeQuatOctXy1010R12 and decodeQuatOctXy1010R12 utilities have been removed; utils.encodeQuatSmallestThree32() and utils.decodeQuatSmallestThree32() have been added in their place. If you use postDecode in your integrations, you will need to adapt your code to the new set of fields after the update.