WorldGen by Meta – Text to Immersive 3D Worlds (www.meta.com)

Meta introduced WorldGen, an end-to-end research system that generates fully navigable, stylistically consistent 3D worlds from a single text prompt (e.g., “cartoon medieval village” or “sci‑fi base station on Mars”). Unlike single-view approaches that only render high-quality geometry near a central viewpoint, WorldGen conditions on a global reference/layout and combines procedural reasoning, diffusion-based 3D generation, and object-aware scene decomposition to produce large, interactive scenes (demonstrated at ~50 x 50 meters) with coherent textures, geometry, and walkable navigation meshes. Generated worlds include sound-aware connectivity and export directly to common engines like Unity and Unreal. Technically, the pipeline stages include procedural blockout and navmesh extraction, reference image generation, image-to-3D reconstruction to create a base mesh, navmesh-driven scene generation, accelerated part extraction (AutoPartGen), mesh refinement, and texturing. Innovations focus on maintaining geometric and stylistic integrity across distance, scene decomposition for reusable parts, and render-efficient outputs suitable for gaming, simulation, and social VR. WorldGen is still research-only with limits on world size and latency, but it signals major potential to democratize 3D content creation, reducing time and cost for developers and creators while pointing to future scaled, faster generation for immersive applications.