The lack of facial LoRAs for Z Image Base was troubling so I trained my own to my liking.

Trained on 104 manually-tagged images. The captions were a mix of danbooru-style tags and natural language.

The trigger word is "facial" and the LoRA was intended primarily for kneeling/lying down poses where the female subject is facing the subject straight-on and close-up, though it might be more flexible than that. It should output respectable loads on the face length-wise instead of those disappointing dribbles, puddles, or nonsense directions that I often see online.