they already showed this tech like 1 year ago, it's basically just a real time video generator, but instead of the screen being the memory, they probably store the memory into a cube map. So if you walked behind the wall, the paint should disappear.
The painting system is new, but it does not look like it's done through AI, it must be a tool that's directly manipulating the memory (the screen pixels / depth). And to get the high FPS I can only assume it's more like photogrammetry of a static AI cube map, than a real time AI video generation (I don't see any people walking around).
Actually looking at the video again, the artifacts on the floor looks like someone just took a photo of a real wall and then just generated the rest of it using world-gen.
https://worldgen.github.io/
Overall cool tech, the problem is that there is like 10000x more 2D videos on the internet compared to 3D cubemap videos (youtube VR videos), and I assume that google is probably just converting a 2D video into a 3D video using something like worldgen so that you can look around, and that causes a lot of issues and artifacts, and they are just hiding it because it's so shitty (imagine if it ran at half speed as shown on video, you would need to lobotomize it or use expensive asic chips to make it run faster).
