This document provides context on the Spatial Intelligence landscape and how World Labs' "Large World Models" (LWM) differ from traditional 3D workflows.
Traditional 3D modeling relies on manual vertex/texture creation (Blender/Maya) or photogrammetry (RealityCapture). World Labs introduces Generative Persistence:
| Feature | Predictive LLM (Sora/Veo) | World Labs (Marble) |
|---|---|---|
| Consistency | Visual only (hallucinates) | Geometry persistent |
| Exploration | Fixed camera path | Free-roam 3D |
| Output | 2D Video Pixels | Gaussian Splats / Meshes |
| Handoff | Video Editing | Unity / Blender / Resonite |
While models like Sora generate "latent" representations that look like video, they lack a coordinate-consistent world state. World Labs' Marble pipeline builds a 3D world that exists independent of the viewer's camera path, enabling true Spatial Computing.
The move to 3D Gaussian Splatting (3DGS) via Spark 2.0 solves the "uncanny valley" of 3D reconstructions. Unlike traditional meshes which struggle with transparency (foliage, smoke, glass), splats represent the scene as a cloud of volumetric primitives, capturing realistic lighting and micro-details.
To get the best results from the Marble API:
- Technical Detail: Use the "Prompt Refinement" feature in the dashboard. Descriptions should focus on materials, global illumination, and spatial configuration.
- Azimuth Awareness: When using
multi_imagelifting, ensure you provide images from distinct cardinal or ordinal directions (e.g., 0°, 90°, 180°, 270°). - Scale Control: For architectural interiors, specify "human-scale" and "accessible layout" to ensure the generated Splat doesn't have "holes" in common walking paths.
World Labs is moving towards Industrial Foundations, where generated worlds aren't just for viewing, but serve as training grounds for robotics and autonomous systems.