B-TIER news
Tencent WorldClaw — prompt-to-3D worlds with every object a separate asset
100k ft — What is this?
Tencent WorldClaw — prompt-to-3D worlds with every object a separate asset
50k ft — Why care?
Tencent's WorldClaw generates a whole 3D scene from a prompt where every tree, house and lamp post is an editable mesh. It chains image generation (GPT Image 2), Segment Anything to split objects, and Hunyuan 2D-to-3D, with depth, normal and instance maps. Not publicly available yet.
10k ft — Key takeaways
• Not yet usable, watch for release
• Pipeline pattern: generate image, segment objects, lift each to 3D
• Useful for game environments and robot sim training
5k ft — Tools & techniques
Hunyuan, Hunyuan3D, Segment Anything, WorldClaw
image-to-3D, instance segmentation
1k ft — Deep-dive / try it
This might be one of the coolest AI models I’ve seen all year.
Tencent’s new WorldClaw model can generate an entire editable 3D world from a prompt — not just the environment, but every individual object inside it.
Trees, buildings, carts, lamp posts, terrain, they’re all generated as separate 3D assets you can actually work with.
Under the hood, it combines image generation, segmentation, and Tencent’s Hunyuan 2D-to-3D tech to turn a generated scene into a full collection of 3D objects.
And the use cases could get really interesting.
You could imagine generating game environments in minutes, building virtual worlds for simulations, or even creating environments where robots can train before they ever enter the real world.
Unfortunately we can’t actually play with this model yet, but…
Transcript excerpt
The coolest thing that I saw this week is this new world claw model from Tencent. It's a tool where you can actually prompt an entire 3D world into existence, and every little individual element of this 3D world here is actually an editable individual asset. So for example, generate a snow line village. It generated the entire environment, but it also generated every little individual asset. The trees, the houses, the carts, the little lamp posts. There's another one, a desert battle scene. You've got the desert here and you've got all these assets, but look, there's a depth map, there's a normal map, there's an instance map, and every single one of these is a separated asset. And these aren't like 3D Gaussian splats or nurse or anything like that. What this Tencent Hanyan model is most known for is actually taking 2D images and converting them to 3D images. So it will generate the images with GPT image two, use the segment anything model to break out everything from within the images, and then use the Hanyan technology, which converts 2D images into 3D images to create all of the 3D assets that go into the world. This can be used to develop virtual environments that you can train robots in so that robots could see every possible scenario and train in virtual environments before being put into the real world. This isn't a model that we have access to actually play with yet, but I just found it really, really cool, so I wanted to share it first.