Today we’re sharing an update on our quest to push the frontier of spatial intelligence and generate persistent, navigable, and controllable 3D worlds. We are also introducing Marble, a limited access beta preview of our model today at marble.worldlabs.ai, where users can view and create 3D worlds. (View Highlight)
Given an image or text prompt, our model generates a 3D world that you can explore for as long as you wish — no time limits, no morphing, no inconsistency. Compared to our previous results, our generated worlds are bigger, more stylistically diverse, and have cleaner 3D geometry. (View Highlight)
Marble makes our model available for users to start viewing and building worlds today. Enthusiasts and builders can export generated worlds as Gaussian splats and use them in downstream projects. This is particularly easy to do with our open-source rendering library Spark which seamlessly integrates Gaussian splats into Three.js for building web-based 3D experiences, and renders efficiently on desktops, laptops, mobile devices, and VR headsets. (View Highlight)
The ability to navigate and interact with a consistent 3D world lies at the heart of many workflows and use cases for visual creators. Our generated 3D worlds allow for free viewpoint navigation, in the user’s browser, at zero cost. In contrast to the limited 3D experience provided by depth maps or point clouds, our worlds have richer geometric complexity and produce more complete 3D worlds, allowing users to look behind and beyond what is visible in the input view. (View Highlight)
Room-sized worlds are a great start, but creators’ imagination often extends beyond these boundaries. We want to share a glimpse of what’s possible as users compose their generated spaces together to build out much larger environments. Given the stylistic coherence and geometric consistency achieved by our model, it’s already fun to be able to navigate, browse, and imagine many use cases for larger worlds like those shown here. (View Highlight)
