WorldClaw Generates Explorable 3D Open Worlds From Text Using a Multi-Agent Coarse-to-Fine Pipeline That Keeps Assets Editable

arXiv / HuggingFace Daily Papers (41 upvotes)·low signalXBlueskyLinkedInCopy link

Tencent Hunyuan's WorldClaw (arXiv 2608.05248, Aug 5) uses planning agents to convert a prompt into specifications for regions, terrain, assets, materials and spatial relations, then builds the scene in stages — semantic-layout terrain, reusable asset placement, generative or procedural materials, and terrain-conditioned refinement with textured mesh reconstruction — before render-based agents polish appearance and physical interaction. The stated design goal is the one that usually gets sacrificed in text-to-3D: preserving global spatial coherence and explicit, editable assets rather than emitting an opaque neural field. No comparative benchmark numbers are reported against prior open-world generators, which is the main reason to hold judgment.