Gaussian splat#

.splat read + write lazy: zero-copy

Summary of the specification#

.splat is the 32-byte-per-Gaussian binary layout used by the WebGL Gaussian Splat Viewer and compatible tools. It has no header: the file is a flat array of fixed-width little-endian records, and the splat count is the file size divided by 32. Each record is three float32 positions, three float32 scales, four uint8 colour and opacity bytes, and four uint8 packed rotation quaternion components - 32 bytes exactly. The format carries a point cloud only; there is no connectivity of any kind.

Specification at a glance#

header	none; count = file size / 32
record size	32 bytes, little-endian
position	x, y, z as float32
scale	scale_0, scale_1, scale_2 as float32
colour	color_r, color_g, color_b, opacity as uint8
rotation	rot_0..rot_3 as uint8, packed quaternion
connectivity	none

Reference implementation ↗

Reading#

import polyxios as px

cloud = px.read("scene.splat")
cloud.vertices               # (n, 3)
cloud.vertex_attrs["opacity"]
len(cloud.element_types)     # 0 - the format carries no elements

Writing#

px.write(cloud, "out.splat")

This codec takes no format-specific options.

Quirks worth knowing#

The eleven non-positional fields are read into and written from vertex_attrs under their record names: scale_0..2, color_r/g/b, opacity, rot_0..3.
A missing vertex_attrs entry is written as zeros rather than raising, so a plain point cloud can be written out as a .splat.
element_types is always empty. Anything expecting faces will find none.
lazy=True maps the file and hands back views of it: the file is a run of 32-byte records, so the positions are a strided float32 view three columns wide and each attribute a strided view of its own column, read-only, nothing decoded or copied. An eager read converts the positions to float64. The mapping needs a path or a handle over a regular file at its start; an in-memory buffer or a gzipped file raises LazyReadError.