This post shows all of the 96 possible combinations of eye and mouth shapes that were trained. "Winking" is also possible, but not shown in these examples.

The same seed and settings were used for all examples. The keywords were weighted to 1.5 and the lora was weighted to 1.1. The checkpoint used was Juggernaut X, but I don't recommend using that checkpoint with this lora. It seems to only generate overly simplified backgrounds and has very limited clothing and face diversity. If you want to use Juggernaut, I recommend version 9 or earlier.