Synthetic data & world models

Create the data
the real world
rarely gives you.

Train on rare defects, varied materials, and conditions that take years to develop. Our synthetic environments turn building geometry, weathering, and damage into controlled training examples.

Build a dataset with us Our approach

Synthetic environment

Material

Output

The data problem

Build coverage across the conditions that matter.

A defect can look different on every material, under every light, and from every viewpoint. Synthetic environments let us vary those dimensions together, targeting the combinations a model needs to learn and generating precise labels for each example.

Material Geometry Condition Exposure Viewpoint

01 / CONSTRUCT

Construct the building

Define building geometry, material assemblies, scene context, and camera configurations.

02 / VARY

Vary the conditions

Explore appearance, weathering, defect patterns, lighting, and the factors a model needs to distinguish.

03 / ANNOTATE

Generate aligned labels

Generate aligned labels for the task: material classes, segmentation masks, geometry, or condition attributes.

04 / VALIDATE

Measure real-world transfer

Evaluate transfer on held-out observations. Measure where the model generalizes and where more data is needed.

Our research direction

A world model
for how buildings change.

Our world-model research connects what a building looks like with its physical state and how that state changes. Materials, environmental conditions, and degradation are part of the same learning problem.

The aim is a shared foundation for material recognition, anomaly detection, and failure prediction. Controlled simulation lets us explore the transitions from an intact surface to visible wear and damage.

Explore the research

Condition over time

Surface change adds condition information to the same building.

Built for your task

Designed around
the task you need to solve.

Discuss a data partnership

Specialized datasets

Start with the feature you need to learn. Scope material classes, defect categories, viewpoints, and output labels around your model’s requirements.

Model collaboration

Work together on targeted recognition tasks, training strategy, and the synthetic-to-real gap.

Evaluation by design

Evaluate on real imagery kept separate from training. Compare across materials, locations, and image quality to understand where additional training data helps.

A dataset with physical context

One scene.
Many training signals.

Every rendered image comes from a defined physical environment. Geometry, materials, camera position, and condition can all become aligned training signals for the same example.

Scene geometry: Roof form, building dimensions, exterior features, and surrounding structures.
Material & condition: Surface classes, weathering, defect location, and severity variations.
Capture conditions: Camera viewpoint, distance, lighting, and environmental appearance.
Training outputs: Rendered imagery, segmentation labels, geometry, and scene metadata.

What’s missing from your dataset?

Bring a difficult visual task, a hard-to-find defect, or an underrepresented material. Let’s explore what we can build together.

Explore a partnership

Create the data the real world rarely gives you.