As someone completely not involved in this project, I would predict that increasing the scene size while remaining halfway consistent isn't that difficult. Let

As someone completely not involved in this project, I would predict that increasing the scene size while remaining halfway consistent isn't that difficult. Let me elaborate by using cat-4d.github.io, one of their competitors in this field of research: If you look at the "How it works" section you can see that the first step is to take an input video and then create artificial viewpoints of the same action being observed by other cameras. And then in the 2nd step, those viewpoints are merged into one 4D gaussian splatting scene. That 2nd step is pretty similar to 4D NeRF training, BTW, just with a different data format. Now if you need a small scene, you generate a few camera locations that are nearby. But there's nothing stopping you from generating other camera locations or even from using previously generated camera locations and moving the camera again, thereby propagating details that the AI invented outwards. So you could imagine this as you start with something "real" at the center of the map and then you create AI fakes with different camera positions in a circle around the real stuff, and then the next circle around the 1st-gen fakes, and the next circle, and so on. This process is mostly limited by 2 things: The ability of your AI model to capture a global theme. World Labs has demonstrated that they can invent details consistent with a theme in this demos, so I would assume they solved this already. And the other limit is computing time. A world box 2x in each direction is 8x the voxel data and I wouldn't be surprised if you need something like 16x to 32x the number of input images to fit the GSplats/NeRF. So most likely, the box limit is purely because the AI model is slow and execution is expensive and they didn't want to spend 10,000x the resources for making the box 10x larger.