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

Meta today unveiled WorldGen, an end-to-end research system that turns a single text prompt (e.g., “cartoon medieval village” or “sci‑fi base station on Mars”) into a fully navigable, interactive 3D world in minutes. Unlike prior single‑view approaches that produce high fidelity only around a central viewpoint, WorldGen generates stylistically coherent, geometrically consistent scenes that span roughly 50×50 meters, include navmeshes for free roaming, and export directly to standard engines like Unity and Unreal. The result is render‑efficient, playable environments suitable for gaming, simulation, and social VR — pointing toward major time and cost savings for 3D content creation. Technically, WorldGen combines procedural reasoning, diffusion‑based 3D generation, and object‑aware scene decomposition in a staged pipeline: planning (procedural blockout and navmesh extraction), reference image generation, image‑to‑3D reconstruction to build a base mesh, navmesh‑conditioned scene generation, initial texturing, part extraction via an accelerated AutoPartGen, and iterative refinement (image enhancement, mesh refinement, texturing). The system maintains global consistency across the scene rather than extrapolating from one view, but it remains research‑stage — not yet public — with known limits on world size and latency that Meta plans to improve.