My first attempt at creating an action LoRA — it’s still difficult to train LoRAs for Z-Image Turbo, but the tests are promising. The big issue is that the size of the member is sometimes disproportionate, and you need to make several generations to get a realistic result. It doesn’t work well if you add other LoRAs.

It doesn’t have a trigger word as such, it just uses a prompt like: pov photo of a girl with curly hair, green eyes, wearing a white bride dress, doing a blowjob, she is sucking a penis, grabbing the penis with her hand. cinematic light from a window.

To generate these examples I used Euler with beta57 — it seems to give better results than other combinations.
Resolution: 1088×1440 for horizontal images
Resolution: 1440×1088 for vertical images