WorldClaw - A 3D world generation framework launched by the Tencent Hunyuan team

What is WorldClaw

WorldClaw is an Agent-driven open-world 3D generation framework launched by the Tencent Hunyuan team. Users input a single natural language description, and the system can automatically plan terrain, regions, buildings, objects, and spatial relationships, adopting a global-to-local generation strategy to construct in Blender a complete 3D scene that can be freely explored and supports per-object editing. The framework is jointly driven by multiple models such as Claude Opus, GPT-Image-2, SAM3, and Hunyuan3D, and possesses vision Agent self-check and repair capabilities, which can automatically fix contact issues such as floating and clipping.

WorldClaw's main features

• Intelligent scene planning: The planning Agent analyzes the user's description and converts open-ended prompts into a structured scene specification containing regions, terrain, objects, materials, and spatial relationships.
• Global-to-local generation: First establish the world terrain and spatial framework, then gradually fill in the details of each region, avoiding road breaks and scale imbalance.
• Independent asset generation: Objects such as buildings, vegetation, and vehicles remain independent instances rather than a single baked model, supporting individual movement, replacement, and editing.
• Visual self-check and repair: The vision Agent performs multi-view rendering checks on proportions, poses, and contact issues, and automatically fixes defects such as floating objects and clipping.
• Deep Blender integration: The entire workflow of terrain generation, asset generation, scene placement, checking, and rendering runs in Blender 5.1.1, and supports operation via the MCP interface.

WorldClaw's technical principles

• Multi-Agent collaborative architecture: The system consists of a planning Agent, a scene optimization Agent, and a visual inspection Agent, which are responsible for scene design, Blender operations, and quality verification, respectively.
• Multi-model division of labor: Claude Opus 4.8 serves as the Agent foundation model, GPT-Image-2 handles image generation, SAM3/SAM3D handles segmentation, and Hunyuan3D handles 3D reconstruction and asset generation.
• Coarse-to-Fine generation strategy: First generate a semantic height field to determine the terrain skeleton, generate object composition plans by region, and finally independently reconstruct each 3D asset and combine them.
• Local repair mechanism: For contact issues, only modify the corresponding support region to avoid damaging the overall terrain structure.

How to use WorldClaw

• Input description: Describe the desired open-world scene in natural language.
• Intelligent planning: The planning Agent automatically breaks down the prompt and generates a structured scene specification containing regions, terrain, objects, and spatial relationships.
• Terrain skeleton generation: The system generates a semantic height field, determining the spatial foundation of terrain regions such as mountains, rivers, and coasts.
• Regional asset generation: Generate independent 3D assets such as buildings, vegetation, and vehicles by region, and place them in the corresponding positions.
• Visual inspection and repair: The vision Agent performs multi-view rendering checks and automatically fixes issues such as abnormal proportions, floating, and clipping.
• Export and edit: Obtain the complete scene in Blender, freely adjust object positions, replace assets, or modify materials.

WorldClaw's core advantages

• From prompt to complete world: A natural language description that can automatically plan and generate a complete explorable 3D scene including terrain, buildings, vegetation, and facilities, without the need for manual asset-by-asset modeling.
• Independently Editable Assets: All objects remain independent textured Meshes and instances, supporting individual movement, replacement, and material modification without regenerating the entire world.
• Intelligent Self-Check and Repair: The visual Agent proactively checks proportions, poses, and contact issues, automatically repairing common 3D defects such as floating and clipping to improve generation quality.
• Deep Blender Integration: The entire pipeline runs in Blender 5.1.1, supports direct operation via the MCP interface, and is easy to integrate into existing 3D workflows.

WorldClaw's project address

• Project Official Website：https://tencent-hunyuan.github.io/Hunyuan3D-WorldClaw/
• arXiv Technical Paper：https://arxiv.org/pdf/2608.05248

Comparison of WorldClaw with similar competing products

WorldClaw's Application Scenarios