Khronos Group, in collaboration with OGC, Niantic Spatial, Cesium (Bentley), and Esri, has formally added 3D Gaussian Splats (3DGS) to the glTF ecosystem. Two experimental extensions, KHR_gaussian_splatting and KHR_gaussian_splatting_compression_spz, lay the foundation for interoperable splat workflows across DCCs, render engines, geospatial platforms, and the web.
This development places Gaussian Splats, until now a fast-moving research and tool-integration phenomenon, onto the same structural footing as meshes, point clouds, and textures. In other words: splats just got standardized.
From Research Trick to File Format
Gaussian splatting isn’t new. The math dates back to the 1990s, but the 2023 Inria paper “3D Gaussian Splatting for Real-Time Radiance Field Rendering” gave the technique a second life. By rendering 3D scenes as clouds of anisotropic Gaussian ellipsoids, each storing radiance and opacity, researchers achieved real-time neural rendering that rivaled neural radiance fields (NeRFs) without the training overhead.
What makes splats compelling is their efficiency: no polygons, no voxel grids, no heavy neural nets. Just point-based primitives that can capture detail in thin structures, soft translucency, or reflective materials.
Within a year, Houdini gained a native GSOPs plugin, V-Ray added splat rendering in Chaos Vantage, and studios began experimenting with splats in volumetric video pipelines such as HOLOSYS+. Now, the Khronos Group has moved to lock this fast-growing format into glTF, which already brands itself as the “JPEG of 3D.”
KHR_gaussian_splatting
The first of the new extensions, KHR_gaussian_splatting, defines how splats are represented inside glTF assets. Each splat is treated as a point primitive with the following attributes: Position (XYZ), Rotation (quaternion), Scale (anisotropic ellipsoid dimensions), Opacity and Spherical Harmonics coefficients for both diffuse and specular lighting
The extension allows fallback to sparse point clouds if a consuming renderer doesn’t support splats—ensuring that files remain viewable, even without specialized splat engines. For shader developers, the inclusion of spherical harmonics (SH) is key. SH enables compact representation of lighting across a hemisphere, which allows splats not only to look correct but also to respond naturally to lighting in real-time environments. Up to 3rd-degree SH coefficients are supported.
KHR_gaussian_splatting_compression_spz
The second extension, KHR_gaussian_splatting_compression_spz, tackles the elephant in the room: storage size. Traditional PLY splat datasets are enormous, with billions of points often required to reconstruct complex scenes. To address this, Niantic Spatial contributed the SPZ compression format, MIT-licensed for free use. SPZ achieves up to 90% smaller file sizes compared to equivalent PLY datasets while maintaining visual fidelity.
SPZ blobs can be embedded directly into glTF buffers. At runtime, they can either be streamed as compressed chunks or decompressed into raw primitives. This dual mode makes SPZ especially suited for geospatial and AR contexts where bandwidth is limited.
SPZ v2.0: Better Rotations
Version 2.0.0 of SPZ adds improved encoding of splat orientations. Rotations are stored as quaternions, but instead of four floats, SPZ encodes only the three smallest components as 10-bit signed integers. The missing largest component is reconstructed at runtime, with a 2-bit index flagging which component was omitted.
This compact scheme increases precision for long, thin geometry—think power lines, fences, antennas—and improves rendering stability in massive geospatial datasets. It’s the difference between “the wire looks straight” and “the wire jitters like a bad motion-capture solve.”
Why This Matters
- Standardization
Until now, Gaussian Splatting workflows were fragmented: PLY exports here, custom Houdini nodes there, and bespoke converters everywhere. With glTF extensions, splats can flow through established pipelines with predictable interoperability. - Production Viability
Splats are great in research demos, but production requires stability. glTF’s adoption across web, AR, and DCC ecosystems ensures that Gaussian Splats won’t remain academic curiosities. - Performance & Compression
SPZ addresses the practical bottleneck: dataset size. Geospatial datasets and volumetric captures can hit terabytes in raw splats; with SPZ, they become streamable assets. - Lighting Flexibility
By embedding spherical harmonics, splats can sit naturally in lighting pipelines—whether inside Houdini’s Karma XPU, V-Ray, or CesiumJS streaming environments.
glTF as the “JPEG of 3D”
glTF’s trajectory has always been about unification and accessibility. From PBR textures to Draco compression, glTF brought complex 3D assets into a web- and engine-friendly format. Adding Gaussian Splats continues that trend, ensuring that cutting-edge formats don’t remain vendor-locked.
Roadmap and Next Steps
The current extensions are experimental—Khronos invites developers to test, validate, and propose refinements. Early validation across diverse geospatial datasets confirmed SPZ’s effectiveness, but the long-term plan is to extend the extensions with more attributes, compression modes, and streaming optimizations.
