WorldClaw: Agentic 3D Open-World Generation at Scale
Chunchao Guo, Jinpeng Li, Yang Li, Zilong Huang
cs.AI, cs.CV
2026-08-05
WorldClaw turns text into a global terrain plus independently editable textured meshes: planning agents specify regions, a render-guided loop refines objects and contacts, and the output drops into game engines.
Generating a freely explorable 3D world from one open-ended sentence has to satisfy three goals that fight each other: coherent global space, rich local content, and assets that can be edited and reused. Most existing methods either synthesize pretty single views or produce a single monolithic mesh whose objects cannot be moved individually.
Tencent's Hunyuan team makes a deliberate trade: instead of training one end-to-end world generator, they orchestrate existing foundation models with agents into a pipeline that decouples global organization from local instances.
WorldClaw is a coarse-to-fine agent framework in three stages.
The first stage is intent analysis and planning. An intent-analysis agent and a scene-planning agent convert the open text into a structured scene specification listing regions, terrain types, required assets, material styles, and spatial relations. The second stage is global terrain generation: guided by a semantic layout map, it builds a region-aware height field procedurally, producing continuous terrain with elevation and regional semantics, then adds reusable assets and generative or procedural materials. The third stage is regional object generation and placement: the local terrain is rendered to a 2D image, an image-editing model composes objects onto it, and an image-to-3D model reconstructs each as an editable textured mesh placed back on the terrain.
A render-based refinement loop closes it. The agent re-renders from diagnostic viewpoints and checks object pose, mesh quality, scale, and terrain contact (floating, excessive penetration, unstable support), then performs object-terrain co-deformation within the local support region, iterating until checks pass or the budget is hit.
The pipeline uses Claude Opus 4.8 as the agent, with GPT-Image-2, SAM3, SAM3D, and Hunyuan3D, executed in Blender 5.1.1 on 4 H20 GPUs. The output is independently editable textured meshes plus a global terrain that drops into game engines.
The paper reports no quantitative metrics. All results are qualitative: four example worlds (tropical island, canyon, desert, snow mountain) shown from global, regional, and walk views, with instance, depth, and normal renderings. The comparison against SynCity, Marble, MajutsuCity, WorldGen, and GPT-5.6 Sol is also qualitative, arguing WorldClaw is better on terrain organization, content richness, free-viewpoint stability, and instance editability.
| Dimension | Evidence offered |
| Terrain and regions | Continuous global terrain, clear elevation, connected semantic regions |
| Content richness | Buildings, vegetation, vehicles placed by regional function |
| Free viewpoint | Global terrain plus independent object meshes, stable geometry over long traversal |
| Instance editability | Independent textured meshes, asset reuse, engine-ready |
The positioning is clear: production. Many 3D methods produce impressive but unusable scenes, while WorldClaw emphasizes editable meshes that enter game engines with reusable assets, a step toward utility. Decoupling global terrain from local instances maps directly onto the real need for worlds that are both globally coherent and locally controllable.
For practitioners, it offers a reference template for agent orchestration: use a strong model to plan and dispatch, chaining procedural terrain, image editing, and image-to-3D tools.
The most glaring issue is the total absence of quantitative evaluation. Every conclusion is a qualitative comparison; "better than baselines" rests on looking at pictures, with no objective metric or user study, which limits how convincing it is.
The authors list three limitations themselves. First, heavy dependence on strong underlying models: in experiments open-source language and image models often failed, either producing terrain programs that did not run or meet requirements, or losing object appearance and pose, so capable models like Claude Opus 4.8, GPT-Image-2, and Hunyuan3D were required. Second, code generation is unstable: turning natural language into Blender programs is still hard, and errors in scale, parameters, or node connections show up directly in the scene, requiring repeated render-inspect cycles. Third, efficiency overhead: generating and reconstructing objects one by one with multiple refinement rounds raises latency and compute with object and iteration counts, so even simple scenes go through a long pipeline.
More fundamentally, this is engineering orchestration, not a new model. Its ceiling is the ceiling of the underlying foundation models, and 3D reconstruction quality directly bounds the scene's look.