A couple if interesting things.
1) Even on their own example the resulting mesh are NOT a representation of the original object at all other than very broad shapes.
2) You have NO control over the results. You put an input and HOPES it looks like your design.
(part 1/2)
Nvidia presents EdgeRunner. The method can generate high-quality 3D meshes with up to 4,000 faces at a spatial resolution of 512 from images and point-clouds.