Tencent announces WorldClaw

Tencent Hunyuan has introduced WorldClaw, a new research project and system designed to transform text prompts into large, explorable 3D environments. Unlike traditional generative models that produce a single locked video or static image render, WorldClaw focuses on delivering instance-level assets that creators can edit and reuse.

Planning and Refinement Agents

WorldClaw, a new paper from Tencent Hunyuan, describes a text-to-3D system that builds large, explorable scenes and returns editable, instance-level assets rather than a single locked render. To achieve this, the system uses a multi-stage process where planning agents convert a prompt into a structured specification of regions, terrain, assets, materials, and spatial relations before render-based agents refine terrain, objects, appearance, and contacts.

Consistent Terrain Generation

Maintaining spatial consistency across an expansive virtual landscape is a common challenge for automated generators. WorldClaw builds a terrain foundation from semantic layouts and a region-aware height field, which the authors say holds one consistent terrain structure across a large scene. For regions that require intricate details, the system reconstructs textured meshes and recovers their precise placement on the terrain.