A realistic concept model for female peeing, wetting, and desperation.
Training details
Trained on slightly over 2000 images, all of which were manually tagged. This model involves multiple loras and custom model merges. The most prominent model component is BigLust (and by extension bigASP), therefore prompting the model should generally be how you would prompt those models.
Recommended generation settings
It was trained on various aspect ratio buckets centered around 1024x1024 pixels. 896x1152 is a good portrait resolution to use. CFG 4 or 5. For samplers I like DPM++ 3M SDE as a well-rounded, neutral option. For more detailed and generally higher-quality images, I recommend doing a 1.5x img2img upscale pass, with 0.5-0.7 denoise. All the showcase images do this.
Prompting guide
This is a bit long but I'll try to describe everything in detail. All tags are formatted in bold. Training was done with shuffled tags so any order should work, but I try to put more important tags first.
Core concepts
The core tags are peeing, wetting self, and desperation. Pretty self explanatory. You can also try leaking, which is like wetting self but much less intense, but the model may not have learned that one as well.
Poses
Standing, squatting, crouching, sitting, kneeling, lying. The model is very responsive to the pussy tag. If the girl's pussy should be visible, make sure to have that tag included. Holding crotch is used with desperation for the "potty dance" type of pose.
There are also several tags that work like modifiers to poses. Arm support for leaning or supporting using one or both arms. Knees up can be used with sitting. Leg up can be used with standing, for having one leg raised up on some object or ledge. Leaning for leaning against a wall or other object. Bent over for bending far forward at the waist, usually used with squatting or crouching. Looking down was added as a tag because of how prevalent it was in the data set, and it leaked through strongly if not explicitly tagged. Spread pussy, self explanatory. Touching self is kind of catch-all tag, used for images where the girl is pulling up on her vulva to aim her pee (but not completely spreading her pussy). Also used for having a hand on her thigh, ass, etc in a sexy, provocative pose.
Angle
From above, from below, from side, from behind, head out of frame, close-up.
Facial expression
Relief, embarrassed, parted lips, open mouth, smile, closed eyes.
Clothing position
$CLOTHING_ITEM down, where $CLOTHING_ITEM is one of panties, jeans, pants, shorts, leggings, bikini. $CLOTHING_ITEM aside, with panties, shorts, or bikini. Might work, but not learned very well: panties around one leg, shorts around one leg.
Peeing into
Peeing in $OBJECT, where $OBJECT is one of cup, bowl, vase, bucket, toilet, urinal, urinal backwards, sink, clothes. Peeing in clothes means clothes pulled down but peeing into them. E.g. "jeans down, peeing in clothes". If peeing into a container, holding container means the girl should be holding the container with her hands (but the model has not learned this very well due to a lack of training images).
Location
Indoors, outdoors, nature, urban, shower, bathtub, next to car, next to car with open door. Every image was tagged with indoors or outdoors. Fully nature (no buildings or urban infrastructure) or fully urban scenes were tagged respectively. Over edge is a less common tag, not learned as well, and means peeing over the edge of something, like a cliff, dock, boat, etc.
Quality tags
All images were tagged with rating_1, rating_2, or rating_3 for the overall quality of the image. 1 is worst and 3 is best. Less attractive women (per my own judgement) were tagged with average or below average, so the model typically defaults to more attractive women. Some training images are still frames from videos, they were tagged with source:video or source:4k_video.
Lighting / vibe
Night for photos taken at night. Implies flash photography was used. Candid was tagged for any image that looked candid, "caught peeing", etc. Has subtle but noticeable effects on multiple aspects of the image.
Known issues
The model is a bit overfit. I even measured this with custom changes to Kohya's sd-scripts that track validation loss on a held-out set of images. When comparing different versions, the overfit one always looked a bit better to me, and had better conceptual knowledge. This was a purposeful decision I made. Ideally there would be even more training data and the model could learn all the concepts strongly enough without overfitting.
Desperation and wetting conceptual knowledge is a bit lacking compared to peeing. These images were also from a very limited number of data sources, so the model has picked up on some quirks of those specific sources.
Some types of images are insufficiently tagged with looking down. In particular, from behind has a very strong tendency to have the girl looking down where the back of her head is not visible. I will fix this in a future version.
Peeing in various types of containers, especially while holding them, is not great. Need more training data.