Trained on Juggernaut XL v8. All training images were with (mostly Caucasian) women 20-30 years old, without logos or text; any logos/text that you get are from the base model, not this LoRA. Does not render Asian or African pubic hair accurately (looks too Caucasian). Steps: 29280, Epochs: 60, Repeats: 4, Network dimension: 128. Took 13 hr on RTX4080. I tested earlier epochs, and (maybe) because of high network dimension, there still isn't much overfitting. There are plenty of guides on dimension/epochs/steps/repeats, but its still mostly a black box to me.
Recommended strength: 0.3-0.8. <lora:SDXL_hairy:0.5> CFG scales 3-8.
Can work without prompts, but the phrase "hairy pussy, female pubic hair, vagina, clitoris, ..." in your prompts works best.
Works very well with Omnium, Photopedia XL, RealismEngine SDXL, and pretty well with SDXXXL, NightVision, NewReality, RealisticStockPhoto, RealisticDigital, SocaRealism, ThinkDiffusionXL, SD XL (base), Juggernaut, RealVisXL3, WyvernMix, Copax.
I had poor results using this LoRA with EnjoyXL, ImperfectXL, TD-UltraReal.
Best to add into your prompt a few other LoRAs for more realism, like Real Humans, Polyhedron_all SDXL, Freckles & Skin XL, Perfect Eyes XL, Teeth XL, NaturalBodyXL, and even additional hairy pussy LoRAs like Really Hairy Women or Hairy Girls SDXL.
I've had excellent upscale results with 4x-UltraSharp, and terrible with the default "latent".
Description
same as v1.5, but with 32 dimensions instead of 128. Trained on the same 123 images, 4 repeats, 20 epochs, 4920 steps, 2.5 hr on RTX 4080. "Only" 223Mb (instead of 891Mb for non-lite version). So far, my tests show nearly identical images between lite and full.
Note that the file name is v1.5_LITE, but v1.4_LITE shows up when you call it in the prompt; just a typo on my part that I don't know how to fix easily.