Gaussian Splats

DroneDB supports 3D Gaussian Splatting as a visualization format. Gaussian Splats provide photorealistic 3D scene reconstruction from drone imagery, with efficient streaming via the SPZ and RAD formats.

Supported formats

Format	Extension	Description
PLY (splat)	.ply	Gaussian Splat in PLY format (content-detected)
SPZ	.spz	Compressed Gaussian Splat (Niants SPZ format)
SPLAT	.splat	Binary Gaussian Splat format
KSPLAT	.ksplat	KSplat compressed format

note

PLY files are automatically classified as point clouds, 3D models, or Gaussian Splats based on their content.

Build pipeline

DroneDB converts Gaussian Splat sources into a streaming-optimized format:

1. Source (PLY, SPLAT, SPZ, or KSPLAT) is indexed
2. SPZ compressed archive is generated as a temporary input (model.spz)
3. build-lod converts the SPZ into a RAD level-of-detail container (model.rad) for progressive streaming

The RAD format enables the viewer to load coarse splats first and progressively refine the scene as the user navigates. After a successful build, the intermediate .spz file is removed; the only delivery artifact is model.rad (plus bounds.json and georef.json when georeferencing is available).

Prerequisites

LOD streaming requires the build-lod binary. The DroneDB build scripts compile it automatically from the vendor/spark submodule if a Rust toolchain (cargo) is installed. Without build-lod, DroneDB can still serve plain SPZ files, but no RAD LOD streaming is available.

CLI

# Index a Gaussian Splat file
ddb add model.ply

# Build SPZ + RAD for streaming
ddb gsplat model.ply

The gsplat command converts the source file to SPZ format. When build-lod is available, the build pipeline then generates the RAD LOD container. A GsplatGeoref sidecar file (georef.json) is created with SRS and offset information following the ODM convention.

You can tune spherical-harmonics quantization with the DDB_GSPLAT_SH_BITS environment variable:

# Reduce SH bit depth to save bandwidth (values 1-8)
DDB_GSPLAT_SH_BITS=5 ddb gsplat model.ply

# Control first-band and remaining-bands bit depth independently
DDB_GSPLAT_SH_BITS=5,4 ddb gsplat model.ply

In Hub

Gaussian Splats are rendered in the browser using the built-in splat viewer:

• Progressive loading: Coarse splats load first, refining as you navigate
• Camera auto-framing: The viewer automatically frames the scene using the RAD bounds
• Georeferencing: Splats are positioned in the correct geographic location using the sidecar georef data

Georeferencing

Each Gaussian Splat entry includes a georef.json sidecar file with:

• SRS: Coordinate reference system (EPSG code or WKT)
• Offset: Translation offset for positioning the splat in geographic space

This follows the ODM (OpenDroneMap) convention for Gaussian Splat georeferencing.

Workflow example

# 1. Index the Gaussian Splat output from ODX or another processor
ddb add model.ply

# 2. Build streaming format
ddb gsplat model.ply

# 3. Push to Hub
ddb push

The Hub viewer will automatically detect the Gaussian Splat entry and render it with the 3D splat viewer.