WorldClaw builds explorable 3D worlds from a prompt, as real assets

Tencent's WorldClaw describes a way to build an editable 3D environment from a text prompt with a coarse-to-fine process. The research paper explains that planning agents first establish a terrain base for the whole scene, and render-based agents then refine terrain, objects, appearance and contacts. For regions that need detail, the system reconstructs editable textured meshes.

Tencent's announcement says the scenes are built from editable 3D assets rather than video or Gaussian splatting, so individual objects can be edited in conventional creation workflows. The contribution is a method for coordinating scene construction from broad layout to local detail.