v1 uses a combination of mostly photos and some AI images. Mostly captioned using a wide variety of tokens combined with natural language, all done manually. Dataset re-iterated over 3 times, during a period of 6 months.
Description
C666D Pear Booty Pawg Moss Image v1.
This model is trained with a dataset consisting of 94 images of mixed resolutions. Some of the images are AI, but mostly used photos. It supports creating female characters with a exagerrated lower body, and long legs. The poses in the dataset are mainly: standing, sitting, squatting, bent over, from behind, front, side, etc.etc. without full nudity, explicit nsfw / R18-rated content. But many images do contain high waist lingerie, tight dresses, yoga pants, skirt and shorts, croptops and other cute and sexy stuff. So a wide variety of clothing items. And it increases the quality and sharpness with Moss-SFT.
When used with other lora the recommended usage strength is: 0.7/0.8. This is something I noticed will all Moss loras I tried using SFT. And when going above this strength, the output gets noisier and dusty.
Trained triggers: 🔥🔥🔥
lady, (used instead of 1girl).
milf, (matured woman).
pear booty pawg, (Helps support shape with other lora).
wide hips, (thickens hips/lowerbody).
thick thighs, (thickens thighs below waist).
big ass, (increases size of ass).
plump, (makes character more chubby, both both and face).
thin waist, (decreases width of waist).
thin legs, (decreases volume of thighs-to-ankles).
fat legs, (increases volume of thighs-to-ankles).
chubby face, (fattens face).
Skinny face, (thins face).
Clothing Details: 👗🩳👠
yoga pants, (fully tight-fitted pants).
leggings, (tight pants with loose pipes on ankles).
boho leggings, (the forbidden printed leggings).
pantyhose, (tight nylons and sheer mesh).
skirt, (describe length, to ankle, thighs, thighs, above waist).
dress, (describe length, to ankle, thighs, thighs, above waist).
shorts, (thong style booty shorts).
wireframe corset, (Thin corset with lining).
backless outfit, (open back).
chunky high platform heels, (high heel footwear).
stiletto high heels, (high heel footwear).
nike sneakers, (sport footwear).
transparent clothes, (works on all individual clothing pieces).
thong, (deep skin underwear).
highleg, (high-waist bottoms height).
Etnicities and skincolors with facial and body features: 👩👩🏻👩🏿
alternative glamrock girl, (alt style).
turkish, (arabian descent, dark smooth hair, fair skin).
african, (dark skin, dark frizzy hair).
romanian, (light skin, dark haircolor, european).
caucasian, (light skin, european american).
dark-skinned female = (ebony african).
spanish latina. (tanned skin latin face).
caucasian latina, (tanned skin semi-caucasian face).
american latina, (tanned skin semi-caucasian face).
latina light skin, (latin semi-caucasian tanned skin).
caucasian european, (fair skin).
caucasian american, (fair skin).
caucasian french, (fair skin).
tanned caucasian, (light tanned caucasian skin).
ebony african, (ebony dark skin).
canadian, (caucasian skin).
mature african woman, (ebony dark skin).
latina, (latin light-tanned skin).
ebony african light skin, (ebony tanned skin).
afro american, (ebony dark skin).
ebony african american, (ebony dark skin).
Testing and Usage: 🤖🤖🤖
I have tested the lora to a peak strength of +1, and haven't had any distortion or deformations yet using turbo models. So it can be used as a slider. Feel free to experiment.
the lora works best when the triggers are used in a array, seperated by comma's: e.g: "pear booty pawg, thick thighs, wide hips".
To get a long clothes that cover the legs and ass, ommit/remove the words ass and thighs. This needs to be done within a single sentence! example: her thick thighs and wide hips are covered with a transparent long dress.
The images used in the dataset are in a modern style, but captioned manually to include a wide variety of objects, locations, and other elements, so it understands a wide context of scenes and also lighting effects.
Works very well with my cellulite lora, and allows alot of control on the bodyshape when adjusting strength on both loras individually, but this model also supports cellulite by itself.
I used the same dataset as my Krea2 version, but Moss Image managed to learn more tokens and triggers during training, before overfitting, so it's somewhat more versatile.
For the examples I used Moss-SFT basemodel, CLIP: qwen3.5_9b_heretic_v2_int8_hq from Artius, VAE: flux2-vae, with different samplers, steps and CFG, so experiment, Moss is pretty flexible.
Help me power up: ⚡👍⚡
Hit the like button on the model page, and on my images if you enjoy what I create! And posting a image to the modelpage grants me ⚡50 blue buzz for each user that uploads to a model. Also saving my images to your gallery helps me out a little bit! So I would appreciate it alot if you could spare a few seconds! ⏰ I'm almost out of buzz, because I spend it all on training, and every bit helps so I can keep going! Donations are welcome too! ❤️