Tencent Hunyuan's agentic pipeline for building large, freely explorable 3D worlds from an open-ended text prompt, and the Hunyuan3D line's answer to a three-way tension: a generated world has to hold global spatial coherence, carry rich local detail, and still break down into explicit assets someone can edit and reuse afterwards. WorldClaw is coarse-to-fine and fully agentic. Planning agents first turn the prompt into a structured specification of regions, terrain, assets, materials and spatial relations. From that it builds a globally coherent terrain foundation out of semantic layouts, reusable assets, generative or procedural materials, and a region-aware height field.

Regions that need detail then get terrain-conditioned compositions, from which the system reconstructs editable textured meshes and recovers where each sits on the terrain; render-based agents make a further pass over terrain, objects, appearance and contacts. The paper reports that across diverse open-world prompts this yields large scenes with coherent spatial organization and editable instance-level assets while the global terrain structure stays consistent. Four authors, Tencent Hunyuan (Chunchao Guo, Jinpeng Li, Yang Li, Zilong Huang). The repository carries no license file and, as of early October, the paper and project page but no released weights; it had about 1,340 GitHub stars.

Paper

Authors: Chunchao Guo · Jinpeng Li · Yang Li · Zilong Huang
3dgenerationagentsworld-modelresearch

Related