[dougabug]

This approach is interesting in that it applies image-to-image diffusion modeling to autoregressively generate 3D consistent novel views, starting with even a single reference 2D image. Unlike some other approaches, a NeRF is not needed as an intermediate representation.

[nl]

Interesting to compare this to DreamFusion[1] which came out of Google Research (which also did this paper) today.

In DreamFusion they do use a NeRF representation.

[1] https://dreamfusion3d.github.io/gallery.html