A prompt to build an explorable 3D World with WorldClaw

From a single prompt, WorldClaw assembles an explorable 3D world with coherent terrain, separate assets, and editable meshes.

From an open-ended description, WorldClaw builds a large-scale 3D environment made up of terrain and distinct objects. The scene does not rely on video or Gaussian Splatting: its elements are represented as geometry and textures that can be edited independently and explored from different viewpoints.

The workflow moves from global structure to local detail. Agents first translate the request into a structured specification covering regions, terrain, materials, objects, and spatial relationships. A second stage builds the overall terrain from a semantic map and region-specific parameters before progressively populating areas that require more detail.

For these regions, the system first creates a visual composition adapted to the terrain, isolates individual objects, reconstructs them as textured meshes, and then calculates their position, scale, and orientation within the scene. Additional agents inspect the renders to correct elements such as object size, pose, and contact with the ground.

The paper also details some of the models used behind this orchestration. GPT-Image-2 is involved in generating layout maps and certain visual references, while Hunyuan3D’s image-to-3D capabilities are used to produce reusable assets. Materials combine generated textures with procedural systems that can be executed in Blender.

Eleven environments are shown on the project page, ranging from multi-biome villages to islands and Arctic outposts. Each scene keeps its instances separate and can be viewed through different render passes, including depth, surface normals, and instance masks. Tencent positions this explicit representation for editing, asset reuse, and rendering, animation, or 3D engine workflows.