Synthetic data

Training data at simulation speed

Generate fully annotated EO imagery at scale — across any domain, any condition — without a single real-world collection mission.

Generation modes

Mode 1

Scene builder

A visual interface for composing simulation environments and time series scenarios. Place assets, configure sensor parameters, and export annotated imagery — no code required.

Mode 2

Procedural static data generation

A code-first pipeline for large-scale dataset generation with maximum randomization. Define distributions over scene parameters — asset placement, lighting, weather, sensor noise, backgrounds — and sample thousands of labeled images programmatically.

Fully parameterized

Every scene variable — asset pose, scale, occlusion, lighting angle, atmospheric conditions — is configurable as a distribution.

Massive scale

Run headless across cloud infrastructure. Generate orders of magnitude more data than the scene builder with a single script.

Higher variance

Wider randomization ranges and combinatorial sampling cover edge cases the scene builder's manual workflow can't efficiently reach.

from nulllabs import SceneGenerator

gen = SceneGenerator(domain="maritime")
gen.randomize(
weather=["clear", "overcast", "fog"],
n_assets=(1, 8),
sensor_resolution=["1080p", "4K"],
)
gen.export(n=10_000, output="./dataset")

10,000

labeled images

in

20 min

any domain

Domains

Ground autonomy

Urban terrain, open desert, forested environments with configurable occlusion and lighting.

View examples →

Maritime autonomy

Open water, littoral zones, sea state variation, and weather-driven visibility degradation.

View examples →

Aerial ISR

Nadir and oblique views across altitudes, with atmospheric haze and sensor noise profiles.

View examples →

Counter UAS

Small UAS signatures against cluttered sky backgrounds at a range of aspect angles and ranges.

View examples →

Asset pipeline

3D neural reconstruction

Any real-world asset can be added to our simulation database. We use neural reconstruction to convert video or photo captures into high-fidelity 3D models — no manual modeling required.

Capture

Photograph or video an asset from multiple angles. Standard camera hardware is sufficient.

Reconstruct

Our neural reconstruction pipeline produces a detailed 3D model with accurate geometry and appearance.

Deploy

Assets are added to the simulation database and immediately available for synthetic data generation across all domains.