Load up your generations with this cum fiesta of a LoRA! It seems to work well on whatever checkpoint and subject you’re using, including characters from other LoRAs and embeddings, realistic or anime, etc. Checkpoints that already understand nsfw are more flexible but not required.
To create the examples, I mostly started with well-known prompt/model combos from top checkpoints and then added some cumshot keywords at the end to show how the LoRA generalizes to lots of styles.
Save it as a style so you can add some pizazz to whatever you’re generating at a moment’s notice:
(taking a cumshot on her face, cum string, ejaculation, cum on face, cum on tits, cumshot:1.2) <lora:cumshot_49:0.4>
There are two main knobs to control the intensity: prompt strength and LoRA strength.
Increasing prompt strength just adds more and more cum until around 2.0 strength where your images are more cum than anything else.
It seems most useful at around 0.4 LoRA strength, which covers faces etc. in cum, but around 0.9 strength is more likely to actually make an action shot with a dick firing off a volley. Beware, in-between numbers like 0.7 generated some floating dickblobs in my testing! Avoid the dickblob valley.
Trained entirely on images of professional adult actresses, as freeze frames from 1080p+ video. About half are POV kneeling shots, with some side view or laying down or multiple women or multiple men, and about all are frames where there’s actually a big rope of cum shooting out of a dick. Surprisingly flexible given the limited variety of inputs.
While you can replace “cum on face” with “cum on stomach” etc., the scale of the cum ropes stays the same and the realism drops off when the subject is zoomed out. There’s probably some way around this.
Hentai checkpoints respond fine to the trigger words without loading this LoRA, but it does seem to help achieve a targeted, ropey look in models that tend toward the drips all over look.