A Femboy 'Concept' LoRA. Turns the subject(s) of a generated image into an erect femboy/crossdressing femboy virtually without fail, and creates a strong lean towards:
Visible assholes/gaping assholes 'before' and 'after' sex.
Lean-yet-toned body/androgynous body aesthetic.
Anal sex with an off-screen partner.
Group sex with the femboy subject(s) as the focus.
Having the subject(s) wear heels, toeless legwear, chokers, and ball bras.
Demure posturing such as a hand on one's chest, and presenting with spread legs.
It cannot do flaccid penises or chastity/it has very little training with them.
Important Use Notes:
You must use the activation word 'femboi' in your prompt, and ideally it'll be the first word (its strong enough to replace girl/1girl/femboy entirely for labeling a human subject).
Use EasyNegative in your negatives (Repository here).
Highly recommend that you do gens in dimensions between 512x512~640x640, and then use hires fix to upscale+re-gen them, as it has far more success making coherent images within these ranges compared to 768 and beyond.
This lora should function well at 1.0 strength, especially in simpler prompts, but you can try adjusting it 0.6~1.2 for effect, as well as tweaking the keyword directly (such as femboi:1.1).
Adding 'bent over' or 'from behind' as 2nd in the prompt with 'femboi' as 1st should consistently generate ass shots/'cocktails'.
Use 'open-toe high heels' or 'wedge sandals/sandals' rather than high heels, as the lora has been trained to work with open-toe footwear specifically.
Adding 'spasm' near the end of a prompt will cause those 'currently orgasming' lines to appear around the subject(s).
Adding 'heart eyes' will cause accurate heart icons to form around the subject(s), and when used along with 'spasm' they will further increase the chance and accuracy of an anal sex image. 'heart' as a word alone in the prompt will just add the accurate heart icons.
Adding 'logo:2' as a negative to a prompt can augment images. I say augment because it isnt guaranteed to enhance quality and sometimes just changes the image, but it can and may be more effective on images that seem weaker than average with art quality or composition.
This lora was trained on CounterfeitV2.5, but can likely offer solid performance on other anime-style models (my comment post near the start of comments on this lora has a quick comparison of various models)
