WorldClaw Generates Large-Scale, Agentic 3D Open Worlds from Text.

Researchers have introduced WorldClaw, an innovative agentic framework designed for generating expansive, freely explorable 3D environments directly from open-ended text descriptions. This system addresses the complex challenge of simultaneously ensuring global spatial coherence, detailed local content, and the creation of explicit, reusable assets. WorldClaw operates through a coarse-to-fine methodology. Initially, planning agents interpret a text prompt to create a structured specification encompassing regions, terrain, assets, materials, and their spatial relationships. Subsequently, it constructs a globally consistent terrain foundation using semantic layouts, reusable assets, and generative materials. For areas requiring greater detail, WorldClaw generates terrain-conditioned compositions, reconstructs editable textured meshes, and accurately places them within the terrain. Render-based agents then further refine the terrain, objects, appearance, and contact points, resulting in large-scale scenes with coherent organization and visually compelling, editable content.