Gaussian splats

3D Gaussian splats from a .splat or 3DGS .ply file, drawn on the host's GPU with EWA splatting and returned as a layer, the same pixels for the same camera every time.

1import { createSplatRenderer, loadSplats } from './render/splats.ts'23// in open4const splats = createSplatRenderer(host.gpu!, { width: W, height: H })   // format 'rgba16float'5splats.upload(await loadSplats(host, 'assets/garden.ply'))             // or .splat67// in render(t, options): the host's camera when it passes one, else the film's own8const { view, projection } = options.camera ?? myCamera(t)   // column-major Mat4s9return [{ color: splats.render(view, projection), alpha: 'premultiplied' }]

The layer is an rgba16float texture (or rgba8unorm with format) at the frame size, premultiplied alpha, transparent where there are no splats. Colours are as stored in the file (display values, no conversion, no grading). The texture is reused: composite or read it before the next render. Pass any view and projection in: column-major, the projection in WebGPU clip space (z 0–1). It needs no other plugin; camera-keyframes' matricesOf(shot, W / H) gives both.

• Formats: .splat (32 bytes a splat) and binary little-endian .ply from 3DGS training (parseSplat, parsePly, or parseSplats, which reads the header). Splats are plain typed arrays: build or edit them in code before upload.
• Spherical harmonics degree 0 only: f_dc is read, f_rest is ignored, so colours do not change with the view angle.
• The depth sort runs on the CPU every render: a stable radix sort of Float32 view depths, ties in splat order, O(N), about 20 ms a frame for 1M splats on a desktop CPU. Splats behind the camera are left out.
• 64 bytes a splat in one storage buffer, up to the device's maxStorageBufferBindingSize: the Shaderera hosts ask for the adapter's maximum (2 GB, about 33M splats, on an NVIDIA Blackwell), WebGPU's default is 128 MiB (about 2.1M); upload says so when a file is larger.
• No depth output for the host's depth test: the layer composites over.