Splats · interface

SplatSource

A capture as a reader hands it over: linear values, one array per attribute.

interface SplatSource

import type { SplatSource } from '@driftengine/splats';

In depth

Linear, and that boundary is load-bearing. A .ply from a training run stores scale as its logarithm and opacity as its logit, and undoing both is the reader's job rather than this one's. Getting the boundary wrong produces a capture that is either invisible or a solid block, and both have been reported against other viewers — because both look like a rendering fault.

Properties

Name	Type	Description
`count`readonly	`number`
`positions`readonly	`Float32Array`	Three per splat, in the capture's own space.
`scales`readonly	`Float32Array`	Three per splat: the Gaussian's standard deviation along each of its own axes, in metres.
`rotations`readonly	`Float32Array`	Four per splat, xyzw, normalised here rather than trusted. More Both file formats store wxyz and this does not, which is a deliberate seam: `xyzw` is what `gl-matrix` uses and what every other quaternion in this engine is, so a consumer building a source by hand is not asked to learn a second convention for one type. Each reader reorders on the way in, in one line, at the point where the file's order is already on screen.
`colors`readonly	`Float32Array`	Three per splat, 0 to 1, linear. More Not sRGB-decoded, which is what every reference viewer does and is also what is correct here: a capture's colour comes from a spherical-harmonic DC term, which is linear radiance, and this engine shades and composites in linear and grades once at the end. What would make it wrong is a capture authored by a tool that baked a display transform into the DC term — which produces a washed-out cloud, not a subtly wrong one.
`opacities`readonly	`Float32Array`	One per splat, 0 to 1, already through the logistic if the format stored a logit.
`sh1`readonlyoptional	`Float32Array`	The l=1 spherical-harmonic band, nine per splat, or absent for a capture that has none. More Interleaved by basis function and then by channel — `Y(1,−1).rgb`, `Y(1,0).rgb`, `Y(1,1).rgb` — which is not how a `.ply` stores them. That file is channel-major, all fifteen of red before all fifteen of green, and the reader transposes on the way in for the reason every other reordering in this package happens there: the file's own order is on screen at exactly one place, and a shader that had to know it would be a second copy of a convention nothing checks. Raw coefficients, signed, in the same units the training run wrote — the band constant is the shader's to apply, like `SH_C0` is the reader's for the DC term.