Notice how the poses are very select. This was purposefully done. There is a term called generality, so they mote poses the ai can make the less accurate it will look. Mostly ai cant draw hands well, it gets the number of digits wrong a Lot. So instead these engineers used a very small sample of poses, and probobly 5000 images of random people. I have not looked into the paper but it would have been a good idea to use a synthetic image generator to capture a large variety of randomizations.