World Labs – Generate persistent, navigable, and controllable 3D worlds (www.worldlabs.ai)

World Labs today launched Marble, a limited-access beta that generates persistent, navigable, and controllable 3D worlds from images or text prompts (marble.worldlabs.ai). Unlike transient or inconsistent outputs from some prior systems, Marble produces larger, stylistically diverse scenes with cleaner, more complete 3D geometry you can freely explore in a browser—no time limits, morphing, or holes. The model focuses on environments (not isolated people or animals) and supports composing multiple generations to build out much bigger coherent worlds. Technically, Marble outputs formats like spz and ply and exports scenes as Gaussian splats that integrate with World Labs’ open-source Spark renderer and Three.js for efficient web, mobile, desktop, and VR rendering. This splat-based pipeline gives richer geometry and behind-the-image continuity compared with depth maps or point clouds, enabling free-viewpoint navigation, persistent editing, and downstream use in filmmaking, games, VR, and interactive content. Early users report fast iteration and seamless stitching; an API and broader access are planned as the team refines consistency and scale. For creators and ML practitioners, Marble represents a practical step toward scalable spatial intelligence and composable 3D world generation suitable for real production workflows.