These results are good, but looking at them only reinforces my belief that really good generation of images of real complex subjects requires more than training

These results are good, but looking at them only reinforces my belief that really good generation of images of real complex subjects requires more than training on images (e.g. relationships between images and 3d structure, and understanding relationships between parts of structures, and how items rather than image parts co-occur). E.g. in the first example, where her collar bone should be, are we looking at a garment or skin? We know those are not the same; the model does not. In the second example, the collar of the boy's garment looks odd -- and perhaps like it would be challenging to put on/take off. The model doesn't have an understanding of what a garment is, how it relates physically to the wearer, etc. Glasses should be symmetric most of the time. A light patch on one side of a person's head is perhaps more likely to be due to lighting conditions than really asymmetric and uneven white hairs. These are all things that a human knows from operating in a world full of humans wearing stuff, moving in 3d spaces, etc.