@luma.gl/splats

Experimental Gaussian splat rendering utilities for luma.gl, without dependencies on Apache Arrow, loaders.gl, or deck.gl.

makeGPUSplatData(...) prepares caller-owned GPU data. SplatRenderer supports WebGPU and WebGL2; GPUSplatGraphRenderer progressively streams preserved batches through reusable WebGPU command graphs, global GPU sorting, and one indirect draw. GPUPagedSplatRenderer projects sparse rows from independently owned source pages into bounded GPU segments while preserving one global cross-page depth order.

Prepared batches support degree-one through degree-three spherical harmonics, semantic class IDs, and in-place row updates. Both rendering paths evaluate view-dependent radiance and filter semantic classes. SplatPicker and GPUSplatGraphPicker resolve original source rows; their corresponding mixed-scene helpers compose splats with caller-owned depth-tested meshes.

SplatHierarchyManager traverses frustum-culled, foveated source hierarchies while preserving coarse parent fallback and bounded asynchronous loading. SplatRADHierarchyManager follows Spark's authored per-row global child links, retaining mixed source-page leaves and parent fallback until the complete child frontier is resident. SplatResidencyManager bounds intact streamed batches by GPU bytes, rows, or chunks. Structural glTF adapters accept decoded KHR_gaussian_splatting attributes, mesh feature IDs, and caller-owned SPZ v2 decoder handoffs.

Use optional expectedSplatCount and expectedBatchCount hints to reserve graph capacity. Renderers borrow source batches, preserve HDR colors, and must be destroyed before their data.

Install the experimental package with yarn add @luma.gl/splats. Its APIs may evolve without a 9.4 semver compatibility promise.

See luma.gl for documentation.