WorldClaw: Large-Scale Agentic 3D Open-World Generation

Tencent Hunyuan3D Research has released WorldClaw, an agentic coarse-to-fine framework that generates large-scale, navigable, and editable 3D worlds from a single open-ended prompt.

The planning agent converts the prompt into the following structured specifications:

• Regions and terrain
• Assets and materials
• Spatial relationships between objects
• Global layout and region-specific level of detail

It then combines semantic layout, reusable assets, generative and procedural materials, and region-aware height fields to construct globally consistent terrain. Areas requiring detailed depiction are selectively enriched to ensure both overall scale and local detail.

Generated terrains and objects are maintained as separate instances, enabling subsequent editing and reuse. The released results include 11 worlds, covering diverse environments such as snowy villages, canyon settlements, tropical islands, and polar outposts.

Each world can be viewed with isometric layout, aerial orbit views, and ground-level walking perspectives, along with RGB, instance masks, surface normals, and depth channels. The research paper, GitHub repository, and a generated world exploration page are also provided.