WorldClaw

Tencent Hunyuan 3D team released a technology called WorldClaw, which allows users to generate explorable, editable 3D open-world scenes from a single text description.

WorldClaw generates terrain skeletons through semantic layout maps and region-aware height fields, giving valleys, lakes, forests, and roads continuous spatial structure.

WorldClaw currently cannot achieve real-time generation of large game worlds; the larger the scene and the more objects it contains, the higher the time and compute costs.

WorldClaw breaks tasks down into multiple agent stages executed sequentially, first converting the user's vague ideas into a detailed planning scheme that includes zone division, terrain types, object distribution, material selection, and spatial relationships.

WorldClaw achieves 3D world generation through a multi-agent collaborative architecture.

WorldClaw has self-inspection capabilities, enabling it to proactively identify and correct common 3D generation issues such as floating models and mesh clipping, and automatically fill in defects, using a closed-loop design of generate–inspect–correct.

WorldClaw's planning process relies on the structured reasoning capabilities of large language models and adopts a Plan-then-Execute approach.

WorldClaw uses a coarse-to-fine 3D world construction method, first generating a semantic layout map and region-aware heightfield, then performing local refinement.

WorldClaw uses an AI Agent pipeline with Claude Opus 4.8 as the orchestration brain, and calls models such as GPT Image 2, SAM3D, and Hunyuan3D to handle image generation, segmentation, and 3D reconstruction.

Controllability in AI-driven 3D generation is typically achieved through a hierarchical control mechanism of macro-level (text/sketch), meso-level (parametric constraints), and micro-level (manual adjustment).

WorldClaw reconstructs locally synthesized results into an editable Mesh and restores its position, scale, and orientation within the original terrain.

WorldClaw currently mainly handles terrain types expressible by heightfields, and more complex geological structures remain to be addressed in future extensions.

WorldClaw is a project whose core selling point is 'agent-driven large-scale 3D open-world generation,' completing 3D generation through the collaboration of multiple AI agents with planning, decision-making, and self-correction capabilities.

WorldClaw extends the conditional generation technique of diffusion models from 2D images to 3D scenes, using terrain height fields and semantic layouts as conditioning signals to align local content with the global terrain.

WorldClaw's first step is to have the Planning Agent understand the user's intent before directly sampling for generation.

WorldClaw's Planning Agent decomposes natural language into dimensions such as scene type, spatial layout, visual style, primary objects, and terrain category, and generates a structured specification containing regions, positions, proportions, asset lists, and spatial relationships.

WorldClaw executes Terrain Conditioned Composition, whereby local content is constrained by the surrounding terrain, viewpoint, and regional semantics.

Height fields cannot represent vertical occlusion structures (such as caves, cliff recesses, or the space under bridges), and WorldClaw currently primarily handles terrain types expressible by height fields.

WorldClaw extends the conditional generation technology of diffusion models from 2D images to 3D scenes, using terrain height fields and semantic layouts as conditioning signals.

The core breakthrough of AI-driven 3D generation solutions lies in replacing hard-coded rules with learned semantic understanding, which in theory can produce more natural and diverse results.