이미지: METAL LAB 생성
Summary
- Tencent Hunyuan has unveiled 'Hy3D WorldClaw,' an agentic workflow that generates large-scale 3D open worlds from text prompts
- The company said generated scenes can be freely explored and are made up of editable, game-ready 3D assets
- It drew a clear line against video generation and Gaussian splatting, emphasizing the difference from existing world-generation methods
- 이름
- Hy3D WorldClaw
- 공개 주체·시점
- Tencent Hunyuan, 2026년 8월 11일 X 게시
- 입력
- 텍스트 프롬프트
- 구조
- 단일 모델이 아닌 '에이전틱 워크플로'로 소개
- 산출물
- 편집 가능한 게임용(game-ready) 3D 에셋으로 구성된 대규모 오픈월드
- 탐색
- 생성된 모든 장면을 자유롭게 이동·탐색 가능
- 명시한 차별점
- 영상 생성 아님, 가우시안 스플래팅 아님
Walking through a world built from a single prompt
What Tencent Hunyuan posted on X on August 11 was a short video and two sentences. Enter a text prompt, and out comes a large-scale 3D open world — one you can then freely walk through. Its name: Hy3D WorldClaw.
What stands out is that the team didn't call this a "model." Instead, they described it as an "agentic workflow." That means it isn't a single neural network spitting out a result in one shot, but a process that autonomously works through multiple steps to assemble a scene. Think of it as a pipeline that separately handles laying out terrain, placing buildings and trees, and applying materials.
"Not video. Not Gaussian splatting."
Tencent led not with self-praise but with a negation: "Not video generation, not Gaussian Splatting" (Tencent Hunyuan X post). That single line is the core of this announcement.
Over the past year or two, output labeled "world generation" has largely split into two camps. One is video-generative: as the user moves, the next frame is rendered in real time. It looks convincing on screen, but no actual shape data persists underneath — turn around, and the building you just passed might have changed. The other is Gaussian splatting, a technique that reconstructs a space from multiple photos as millions of blurry point clusters. It's highly photorealistic, but the result isn't a mesh (3D surface data), making it hard to drop into a game engine and have doors open or walls move.
What WorldClaw claims to offer is a third path: output that comes out as editable 3D assets from the start. If true, that means the created world isn't just something to look at — developers could pick out and move or modify objects within it.
| Approach | Output | Editing/reuse | Weakness |
|---|---|---|---|
| Video-generative world models | Frame sequences | Essentially impossible | No persistent geometry, hard to maintain consistency |
| Gaussian splatting | 3D point sets | Limited | Not mesh-based, difficult for engine integration or physics |
| Hy3D WorldClaw (claimed) | Game-ready 3D assets | Claimed possible | Release form and quality still unverified |
A cluster of "world" announcements in just days
The context around this announcement reveals a broader trend. Over roughly ten days in early August, major labs one after another turned their attention to "worlds."
| Date | Entity | Announcement |
|---|---|---|
| Aug 4 | NVIDIA | Introduced World Action Model, a concept for robot policies based on video world models |
| Reported Aug 8 | NVIDIA | Released open physical AI foundation models Cosmos 3 and OpenMDW 1.1 |
| Reported Aug 9 | Qwen | Released Qwen-AgentWorld-35B-A3B, a world model for agent training (Apache-2.0) |
| Aug 11 | Tencent Hunyuan | Unveiled Hy3D WorldClaw, a text-to-3D open-world workflow |
Same word, different destinations. NVIDIA's and Qwen's world models focus mainly on creating environments to train robots and agents. WorldClaw, by contrast, squarely targets content meant for humans to enter and use — games and 3D production pipelines.
Tencent's 3D roadmap
Tencent Hunyuan has been shipping quickly of late. On August 9, it released Hy3, a text-generation model built on an MoE architecture, and on the same day uploaded a GGUF version of a 1.8B translation model quantized down to 1.25 bits to Hugging Face. Lightweight language models on one hand, pipeline-level 3D on the other — the team appears to be pushing both directions at once. The "Hy3D" prefix signals that this is an extension of the team's ongoing lineage of 3D-generation work.
That said, this particular post reads more like a demo and a statement of intent. Details such as the release form of weights or code, supported engines, and the upper limit on generation scale weren't included in the post.
So what actually changes
Up to now, most AI-generated "worlds" have been spectacles. Video looked good but couldn't be entered or touched; splatting reconstructed space faithfully, but the moment you tried to bring it into a game production tool, a human had to start modeling all over again. If WorldClaw's claims hold up under scrutiny, it opens a path that goes straight from text to an editable stage. For small game teams that used to spend days building out a single background section, this would mean a tool that can produce a first draft; for robotics, it could mean the ability to mass-produce training environments that drop directly onto a physics engine.
In the end, it all comes down to one question: does the same quality hold up outside the demo video, with prompts written by someone else? That answer will only come once outsiders can actually open the files.
