3D Gaussian Splatting extension lineage

The glTF extensions for 3D Gaussian splatting went through several incarnations in 2025–2026. Understanding the lineage is what pinned down the SPZ-version bug above, so it is documented here in full.

Timeline

• ~June 2025 — a single experimental extension KHR_spz_gaussian_splats_compression shipped in CesiumJS. It bundled the SPZ blob and the splat semantics together.
• Mid 2025 — Khronos split it into two:
  • KHR_gaussian_splatting — renderer-agnostic base: defines the splat attributes (POSITION, COLOR_0, and namespaced KHR_gaussian_splatting:SCALE / :ROTATION / :SH_DEGREE_*), the optional shape (ellipsoid/…) and hints (sorting method, projection), and allows uncompressed or meshopt-compressed data.
  • KHR_gaussian_splatting_compression_spz_2 — a child extension nested inside the base on the primitive, carrying only the SPZ-compressed blob (a bufferView).
• Sep 2025 (CesiumJS) — the original single extension was deprecated; the tiler in Cesium ion switched to outputting the base + _spz_2 pair.
• Nov 2025 (CesiumJS v1.135) — the original extension was removed; only the base + _spz_2 pair is supported. Existing tilesets must be re-tiled.
• Jan 2026 — KHR_gaussian_splatting (base) merged into the glTF repo as a release candidate.

Why the _2 matters (the bug). The _2 in KHR_gaussian_splatting_compression_spz_2 is the SPZ format version (v2). As lexaknyazev notes in the PR, a KHR extension cannot float across SPZ versions — for IP and portability reasons it is locked to a specific one. CesiumJS decodes via the Niantic spz library against the v2 layout. SPZ versions differ in the rotation block:

rad-to-3dtiles had been emitting v3 (because SparkJS writes v3 by default). Under the _spz_2 extension, CesiumJS read that 4-byte rotation block at a 3-byte stride → corrupted rotations that looked like global scale/falloff weirdness. Fix: emit SPZ v2 (see blocker #1 above). SparkJS's reader version-dispatches, so v2 also renders in the 3DTRJS 3DGS plugin.

A separate, related issue: glTF #2509 reports splats rendering upside-down in Cesium — that is a coordinate-system matter (SPZ's Left-Up-Front convention / SH handedness baked into the Niantic CoordinateConverter), distinct from the rotation-stride bug. If, after re-tiling to v2, orientation (not scale) still looks off, that conversion is the next thing to check.

Primary sources

• glTF base extension — KhronosGroup/glTF#2490
• glTF SPZ compression child — KhronosGroup/glTF#2531
• CesiumJS migration issue — CesiumGS/cesium#12837
• CesiumJS implementation PR — CesiumGS/cesium#12843
• "upside-down in Cesium" coordinate issue — KhronosGroup/glTF#2509
• Niantic SPZ reference codec — nianticlabs/spz (src/cc/load-spz.cc)
• Java reference writer — javagl/JSplat (GltfSpzSplatWriter)

---