Real-world data for world models

We capture real places in 16K 360° with metric depth at every pixel, delivered as color, depth maps, and point clouds. Liberty3D is a sample corpus of landmarks across Philadelphia, New Jersey, and Washington, DC that shows what our camera and pipeline produce.

Explore Liberty3DTalk to us
( 01 )See the output

See the output

Every scene is a real place. Open any scene and switch between the color image, the depth map, and the 3D point cloud. All three come from the same capture at the same instant. Look at the depth alignment, the point cloud density, and the level of detail across the entire 360° view.

See all of Liberty3D
( 02 )How it is captured

How it is captured

Alia is a single camera that captures 360° in stereo in one shot. Depth is measured, not estimated, and it is aligned to the color image because both come from the same optics. The optical design is patented. The single-camera panoramic stereo method behind it was published at CVPR. Our processing pipeline converts each capture into color, depth, and point cloud outputs automatically.

360° RGB, single shotMetric depth, every pixelFIG. 01.1 — 360° RGB, SINGLE SHOTFIG. 01.2 — METRIC DEPTH, EVERY PIXEL
drag

THE SAME INSTANT, TWO WAYS. COLOR AT 16K, METRIC DEPTH AT EVERY PIXEL, ACROSS THE ENTIRE 360° VIEW.

World Model Data
( 03 )What it is for

What it is for

Training generative models

World models and video models learn more from data that includes real 3D structure. We provide 16K color and dense metric depth of real places, with the lighting, clutter, and complexity that synthetic data lacks.

Building 3D environments

The point clouds are a strong starting point for Gaussian splats and other reconstructions of real places. We are extending the pipeline in that direction.

( 04 )Research

We test our capture methods on public models and publish the results

RetailSMV
Which camera viewpoint helps video world models learn.
arXiv →
PRISM-World
Teaching video world models how real environments behave. CVPR 2026 workshop.
arXiv →
PRISM
A multi-view video dataset for vision-language models.
arXiv →
SABER
An action dataset for VLA models.
arXiv →
All research
( 05 )We capture where you need

We capture where you need

Liberty3D was captured across Philadelphia, New Jersey, and Washington, DC. We run capture programs in the environments your model needs to learn from.

( 06 )Also from DreamVu

Also from DreamVu

Robot training data.

VLM, navigation, and loco-manipulation datasets from synchronized ego and exo capture.

ROBOT DATA →Simulation environments.

Real venues rebuilt as simulation scenes.

SIMULATION →Custom capture.

Tell us what your model needs to learn.

TALK TO US →
( 07 )Common questions

Common questions

What does DreamVu do?
We capture real places in 16K with metric depth and deliver the data for training world models. Liberty3D is a sample of that output.
How is it different from other 360° data?
The depth is captured with the color, from the same camera at the same instant. You can see the result in 3D on this site.
Can I use it to build 3D environments?
Yes. The point clouds are a good input for Gaussian splatting and reconstruction.
Do you sell cameras?
No. We sell what the camera produces.
How do I get the full Liberty3D corpus?
Request it through the form on the World Model Data page and we will send you login details for web access.
Do you still do robot training data?
Yes. See Robot Data.

Tell us what your model needs to learn

Capture programs, research collaboration, and dataset partnerships.

Talk to us