Not to belittle this or anything (it does look good and show promise), it feels like they somehow generate several consistent (but discrete) views of a given world, then feed all that to the good old pose estimation + gaussian splatting workflow. Whenever you leave the generated area (which isn't exactly huge on the few I tested) you get tell-tale signs of GS.
Yeah, if the entire point is that you can move around inside those worlds, I'd have expected a bit more "walkability" - maybe a few different viewpoints that each have their own Gaussian splatting? Right now, it dissolves pretty quickly once you change the location.
Yeah, it's more of a somewhat 3D-drawing of a frame that you can navigate inside, rather than a world up that happen to fit with whatever image you use as an input, but makes sense as a standalone world when you walk around. For being a "world" model, it doesn't seem to grasp physical space very well.
The interior scenes look and walks great, but any scenes with/in exteriors seems kind of bad.
