Tencent unveils WorldClaw for editable 3D worlds from text

Why it matters

WorldClaw targets the expensive step between generating a convincing 3D image and delivering an editable scene that artists can revise, reuse and move into game production.

Tencent Hunyuan (@TencentHunyuan) unveiled WorldClaw on August 11th, an agent-driven workflow that converts text prompts into large, explorable 3D environments assembled from editable terrain and individual textured assets.

https://x.com/TencentHunyuan/status/2087068591296536755

The distinction matters for game and simulation work. Tencent says WorldClaw produces explicit 3D geometry rather than a video constrained to a generated camera path or a scene represented through Gaussian splatting. The output is designed to let creators move freely through the environment, reposition objects and take assets into established 3D production tools.

Tencent demonstrated the system in a video posted on X. A research paper submitted on August 5th names Tencent Hunyuan researchers Chunchao Guo, Jinpeng Li, Yang Li and Zilong Huang as the authors. Guo leads Tencent Hunyuan's 3D work, which has expanded from generating individual objects toward assembling persistent environments.

How WorldClaw constructs a world

WorldClaw divides scene creation into three stages instead of asking one model to produce an entire environment in a single pass.

First, planning agents convert an open-ended prompt into a structured specification covering regions, terrain, objects, materials, visual style and spatial relationships. The workflow fills in details needed for construction while retaining constraints explicitly supplied by the user.

A terrain agent then generates the global foundation. It creates a semantic layout map, separates the planned world into regions and builds a height field using landform operators for features such as peaks, dunes, terraces and erosion. Materials, rocks and vegetation are distributed according to each region's intended terrain and ecology.

WorldClaw spends additional computation on regions that require detailed objects. It renders a selected area as a 2D image, uses an image-editing model to add the planned content and reconstructs the inserted objects as separate 3D meshes. The workflow then recovers each object's position and places it back onto the terrain.

Agents connected to Blender inspect the result from multiple viewpoints. They can adjust object scale and orientation, repair poor geometry, smooth terrain and address visible defects such as floating objects or excessive intersections with the ground. Those checks run until the scene passes the workflow's criteria or reaches a preset iteration limit.

That coarse-to-fine design is WorldClaw's central technical bet. Large terrain and regional structure are established once, while expensive asset generation and refinement are directed toward selected locations. The scene remains a collection of manageable components rather than being flattened into one generated visual artifact.

Tencent assembled the workflow from several AI systems

WorldClaw is an orchestration system built around multiple foundation models and conventional 3D software. According to the paper, Tencent's experiments used Claude Opus 4.8 as the agent model, GPT-Image-2 for image generation, SAM3 and SAM3D for segmentation and reconstruction, and Hunyuan3D for higher-quality geometry and textures.

The researchers ran their experiments in Blender 5.1.1 on a server equipped with four Nvidia H20 GPUs. Large objects received 2,048-by-2,048 physically based rendering texture maps, while smaller assets used 1,024-by-1,024 maps.

Those requirements frame WorldClaw as a research and production workflow rather than a lightweight real-time generator. Tencent's paper reports qualitative demonstrations and comparisons, without presenting generation times or a broad quantitative benchmark for scene quality.

The demonstrated prompts produced a tropical pirate stronghold, a river canyon with tribal settlements, a desert battlefield and a snowy mountain valley containing futuristic facilities. Additional examples on the WorldClaw project page include medieval settlements, volcanic environments and a gemstone mining site.

Tencent extends its 3D push from assets to environments

WorldClaw builds on Tencent's earlier work around HunyuanWorld. The company released HunyuanWorld 1.0 in 2025 as a system for generating explorable 3D scenes from text or images. WorldClaw shifts more of the construction process to specialized agents that plan, generate, inspect and edit discrete pieces of a larger environment.

Tencent also launched its Hunyuan 3D creation engine internationally in November 2025, positioning object generation and world construction as tools for games, virtual reality, digital content, e-commerce and 3D printing. WorldClaw pushes that strategy closer to scene production, where usable topology, asset separation and placement control matter alongside visual quality.

The commercial test will be whether the workflow can reduce the manual labor required to turn generated concepts into playable levels. WorldClaw's editable output addresses a persistent weakness in generative world systems: impressive walkthroughs can be difficult to revise, reuse or integrate into a real game pipeline. Tencent's approach gives each generated object a place in a conventional scene graph, leaving artists and developers with components they can continue editing after the agents finish.