Hacker News Reader

This is likely drawing from the set of pre-generated faces. If so it is violating the attribution requirement of the license.

Are you sure? I would think the pre-generated list would weed out the obviously incorrect/malformed ones that you sometimes get.

Karras et al 2018 dumped a random sample, not a screened sample. (This avoids cherrypicking accusations.)

Agreed, I got the EXACT same image twice, so I think that must be what is happening.