this is really impressive. i'm surprised that text-to-image models know enough about the way the world looks that they can be used to robustly generate coherent 3d models, that aren't full of weird non euclidean bulbous artifacts ๐ณ twitter.com/poolio/status/…