Tencent’s WorldClaw Builds an Editable 3D World From One Prompt
AI agents plan the terrain, build separate textured objects, and use Blender to inspect and repair the finished scene.
Tencent’s Hunyuan3D team has introduced WorldClaw (opens in a new tab), a research workflow that builds a large 3D world from one written prompt.
The result is a collection of editable parts. A creator can move one building, replace a tree, or reshape the ground instead of rebuilding the whole scene.
It plans the world before it builds
WorldClaw first turns the prompt into a plan for the regions, terrain, objects, materials, and their positions. That shared plan keeps the later steps working on one connected world.
Next, it draws a color-coded map and turns it into one continuous height field. The height field controls how high or low the ground sits, helping mountains, rivers, and plains meet without visible breaks.
The system spreads rocks, plants, and materials across the land. AI agents then render the scene in Blender, inspect the result, and adjust rough terrain, misplaced objects, textures, and lighting.
Every object stays editable
When a region needs a village, vehicle, or dock, WorldClaw makes a flat reference image over the existing terrain. It cuts each object out and rebuilds it as a separate textured 3D mesh.
Because each object keeps its own shape, position, scale, and texture, a creator can change one part without touching the land around it.
Another inspection checks for objects that float, sink, lean, or arrive at the wrong size. WorldClaw can move the object and reshape the small patch of ground beneath it before checking again.
Tencent shows 11 generated worlds, including a pirate island, river canyon, volcanic lair, and snowy village. Its paper compares rendered views with five other methods, but does not include measured speed, cost, or a user study.
It is a workflow, not a finished game
WorldClaw creates 3D assets for a game workflow. It does not add the navigation, physics, or interactions that make a world playable. Tencent lists game-engine integration as future work.
The research setup is also demanding. Tencent used Blender 5.1.1 on a server with four NVIDIA H20 GPUs, along with several AI models. The paper says each added object and repair round increases the time and cost.
There is no public WorldClaw app yet. Its GitHub repository currently holds the paper and visual assets, not the code needed to run it.
A runnable release and game-engine handoff would show whether one sentence can become a world that creators can keep shaping.
More from Capture & Create.
Browse all stories
Spatial Web · 4 min read
Safari Lets You Drag 3D Models Into Your Space on Vision Pro
A USDZ file and one HTML element are enough to make a web model interactive, stereoscopic, and draggable at real size.
AI-made 3D · 3 min read
Meshy 7 Turns a Single Photo Into an Impressively Detailed 3D Model
We tested Meshy 7 with our own images. The detailed, fully rotatable results show how far image-to-3D generation has come.
Gaussian splats · 1 min read
Intangible’s Browser-Based 3D Studio Gains Gaussian Splat Support
The update lets creators place scans of real places beside other 3D objects and move a camera through the scene.