WorldClaw: Agentic 3D Open-World Generation at Scale

WorldClaw presents a fully agentic, coarse-to-fine framework for generating large, freely explorable 3D worlds from open-ended text. Planning agents convert prompts into structured specifications covering regions, terrain, assets, materials, and spatial relations. The system builds a globally coherent terrain foundation using semantic layouts, reusable assets, generative or procedural materials, and region-aware height fields. Detail-heavy areas receive terrain-conditioned composition, editable textured-mesh reconstruction, and placement recovery. Render-based agents then refine terrain, objects, appearance, and contacts. The authors report coherent spatial organization, compelling local detail, and editable instance-level assets across diverse open-world prompts.