The rendered results are with lowreconstruction fidelity for NeRF [1]and NeRF-W [7]only with6and12training views. For NeRF [1]with18training views, the result shows higher fidelity, but the undesired reflection is also finally rendered (labeled by green box).

The pursuit of realistic image synthesis at arbitrary views, given only a single source image, has long been a cornerstone challenge in computer vision and graphics.

Comparing images above and below the green line in every figure clearly shows the failure in generating coherent samples whose digit identities are supposed to match to conditioned images.

Notsurprisingly, GVS also inherits the challenges of both: generating RGB values from a bare semantic map is an ill-posed problem that is further complicated by the need for the different output views to be photometrically and geometrically consistent. One could tackle this problem with the sequential application ofexisting techniques.

Fortunately, recent advancements in photogrammetry techniques have greatly improved the accessibility of 3D asset creation from 2D images.

GVS also inherits the challenges of both: generating RGB values from a bare semantic map is an ill-posed problem that is further complicated by the need for the different output views to be photometrically and geometrically consistent. One could tackle this problem with the sequential application of existing techniques.