Hi3D V1.0 vs Hunyuan3D WorldClaw: One Asset Done Well, or a World Roughly

A Hunyuan3D WorldClaw scene is not one 3D model. It is terrain, buildings, vegetation, vehicles and characters, each reconstructed as a separate mesh, placed on a generated landscape and then inspected by a vision agent that sends bad regions back for repair. Hi3D V1.0 does not build worlds. It takes one coarse mesh and one reference image and spends a 96 GB GPU making that single object's geometry sharper. The comparison is therefore not scene versus scene. It is throughput versus fidelity, and the two systems sit on opposite sides of the same production line.

What WorldClaw needs before it can build a world

WorldClaw is Tencent Hunyuan's agentic framework for turning one text prompt into a large, explorable, editable 3D world. The paper, "WorldClaw: Agentic 3D Open-World Generation at Scale," appeared on arXiv as 2608.05248 on 7 August 2026, and the project page and GitHub repository followed the same week. The architecture runs in four stages:

• Intent analysis and scene planning — two language-model agents parse the prompt. One extracts only the constraints the user stated; the other fills in the details the prompt left out. The paper names Claude Opus 4.8 for this stage.

• Global terrain generation — a semantic layout map and region-aware height fields produce a continuous terrain skeleton with materials, rocks and vegetation assigned by region.

• Regional object generation — detailed regions are rendered into terrain-constrained 2D composites, segmented into instances with GPT-Image-2 and SAM3D, reconstructed as individual 3D meshes through Hunyuan3D, and placed back onto the terrain.

• Render-based refinement — a vision agent renders the scene from multiple viewpoints, looks for floating objects, interpenetration, wrong scale and wrong pose, fixes the offending region and re-renders until it passes. Assembly happens in Blender through MCP.

That third stage is the one that matters here. Every object in a WorldClaw scene is a separate image-to-3D reconstruction. A scene with two hundred distinct props is two hundred mesh generations, each conditioned on a segmented 2D composite and each carrying whatever fidelity the underlying reconstruction model can produce. The scene's visual ceiling is set by the worst asset in it, and the framework's own evaluation does not measure that. The paper reports qualitative side-by-sides against SynCity, Meta's WorldGen, MajutsuCity and World Labs' Marble — no FID, no CLIP score, no user study, and no per-object fidelity metric.

What Hi3D would contribute, and what it would not

Hi3D V1.0, released on 23 October 2026, is a refinement model built on Sparc3D. Give it a reference image of an object and a coarse mesh of the same object, and it progressively samples sparse latents, decodes them through a self-pruning sparse VAE and extracts a higher-resolution surface. The three passes take the structural grid to 64³, 128³ and 192³ and the sparse extraction grid to 512³, 1024³ and 1536³, at 25 scheduler updates per pass. It runs one object at a time.

The fit with WorldClaw's third stage is obvious on paper. WorldClaw already has a segmented 2D composite for each object and already produces a mesh through Hunyuan3D. Those are exactly the two inputs Hi3D wants: a reference image and a coarse mesh. Refining each object before it is placed would raise the geometric detail of the assets that decide the scene's close-up quality.

Three limits keep that from being a drop-in upgrade.

First, Hi3D is geometry-only. It produces no texture, material or vertex colour, and WorldClaw's output is explicitly textured, individually editable meshes meant for a game engine. Hi3D would improve the silhouette and the local surface detail and then hand the asset to a texturing step that neither repository supplies. A world built from refined-but-untextured meshes is not the deliverable WorldClaw promises.

Second, the reference image has to show the same object. WorldClaw's segmentation gives an image crop of the object inside a regional composite, which is close but not identical to a clean product reference. Whether that is good enough is an empirical question nobody has answered.

Third, throughput. Hi3D refines one mesh at a time, and the measured configuration is a 96 GB H20 with no CPU path and no batch mode. WorldClaw already runs on 4× H20 by the paper's account. Adding a per-object refinement stage in front of placement multiplies the most expensive part of the pipeline by the number of objects in the scene.

The number neither page gives you

You cannot cost this composition from the published material, and that is the most important practical finding in this comparison.

Hi3D publishes no wall-clock latency. The README gives scheduler updates, pass counts and a measured GPU, but not seconds per object at 512³, 1024³ or 1536³. WorldClaw publishes no per-object reconstruction time either. The paper's compute statement is about the whole system on 4× H20, not about the marginal cost of one more prop. Without both numbers, the arithmetic that decides whether refinement is affordable — objects per scene times seconds per object times GPU-hours — cannot be done from the sources.

What can be said is the shape of the cost. Hi3D's 512³ setting runs one pass; 1024³ runs two; 1536³ runs all three. Each pass is 25 scheduler updates. Higher resolution buys detail and costs passes. Batch size is one, so the cost scales linearly with the object count and cannot be amortised across a batch. If a WorldClaw scene contains dozens of hero objects and hundreds of background props, a sensible pipeline would refine the heroes and leave the background at the coarse stage — a decision that requires exactly the per-object timing neither project has published.

Five dimensions that actually separate them

• Unit of work — Hi3D V1.0 one object vs Hunyuan3D WorldClaw one open-world scene containing many objects

• Output — Hi3D V1.0 a geometry-only mesh vs Hunyuan3D WorldClaw editable, textured meshes assembled in Blender

• Input — Hi3D V1.0 a reference image plus a coarse mesh vs Hunyuan3D WorldClaw one open-ended text prompt

• Code and weights — Hi3D V1.0 downloadable today under the Hi3D Open License 1.0 vs Hunyuan3D WorldClaw a README and a paper, no code, no weights, no API, no pricing

• Evidence — Hi3D V1.0 no numeric benchmark vs Hunyuan3D WorldClaw qualitative comparisons only, no FID, no CLIP, no user study

Availability decides the quarter

One of these is a tool and the other is a design document.

Hi3D V1.0 can be cloned, installed and run today. It needs Linux, Python 3.10, CUDA 12.0 to 12.5, a compute-capability 8.0 GPU and a source build of flash-attn, and the full 1536³ ladder has only been validated on a 96 GB H20. The licence permits commercial use, with a notification and separate-licence requirement once a product exceeds 200,000 monthly active users or USD 1,000,000 in monthly revenue, and an attribution requirement for commercial integrations. The weights are eight files on Hugging Face. There is no hosted endpoint — the model card says it is not deployed by any inference provider — and the company's commercial v3.0, with 2048³ voxels and 8K textures, is a separate product sold through its own platform.

WorldClaw cannot be run by anyone outside Tencent today. The GitHub repository contains an assets folder, a .gitignore and a README. There is no implementation, no checkpoint, no API and no pricing, and an issue asking for code has gone unanswered. The paper is detailed enough to reimplement the architecture and the project page is a good demonstration, but neither is a dependency you can schedule. If you are building a world generator, WorldClaw is a blueprint worth reading; it is not a component you can call.

That asymmetry is the practical answer to the matchup. A team that needs better object geometry this month can use Hi3D. A team that needs WorldClaw's scene-level capability this month has to build it, and the paper is the only thing Tencent has published that helps.

Where the LLM bill lands

WorldClaw's architecture is unusually honest about where the expensive calls are. Planning the scene, extracting the user's constraints, filling in the details the prompt omitted and running the visual critique that decides what to re-render are all language or vision-language calls. The framework routes them to frontier models rather than training its own. That is a deliberate bet: orchestration over training, with quality bounded by components you do not control.

It is also the layer OrcaRouter covers. The 3D reconstruction stays on your GPUs and is not something we serve — the catalogue is text, vision, video, embedding and speech models, not geometry. The planner and the critic are a different story. Putting more than 200 models behind one API means the planner and the critic do not need separate contracts, separate keys and separate failure domains; the provider's list price is passed through at 0% markup, so a vendor price change reaches the endpoint the same day; automatic failover keeps one degraded provider from stalling a world mid-generation; and the routing DSL composes several models into one call, which is what a plan-then-critique loop actually is. For a pipeline whose per-scene cost is dominated by repeated agent calls, that control surface matters more than any single model choice.

Who should act, and on what

If your bottleneck is the quality of an individual asset — a printable part, a product model, a hero prop — Hi3D V1.0 is available now and the experiment is cheap to start, provided you own a 96 GB GPU and already have a coarse mesh. Read the limitations section before the press release: the model refines structure, it does not fix a wrong silhouette, and it returns no texture.

If your bottleneck is scene assembly, WorldClaw is the more ambitious system and the less available one. Its contribution is architectural: a coarse-to-fine agent loop with a render-based critic, and a clear statement that the object reconstruction stage is where per-asset fidelity is won or lost. The obvious next experiment is to put a refiner like Hi3D behind that stage and measure whether the scene's close-up quality improves enough to justify the extra passes. Nobody has published that measurement, and until someone does, the comparison remains a hypothesis about a production line rather than a result.