This LoRA was made to increase the accuracy and quality of when generating an image where a man licks the sole of a female's feet, or, Footlicking. This sort of content is misunderstood by most models, often generated inaccurately with the man's face being nowhere near the sole of the foot, or just showing the woman with her tongue out. I aimed to rectify this at least partially.

V1: Even though this was trained on NoobAiXLV-Pred 1.0, the LoRA seems to work well with other Illustrious based models -> they don't even have to be v-prediction based ones. I am unsure as to why this would be the case but it lead me not to label this one as a V-pred model.

About Training:
Generated on a dataset of 65 images, all hand selected and manually masked to take out the background and various text elements within. Training was done using OneTrainer from the V-prediction guide here: https://civarchive.com/articles/8723/vlora-or-how-to-train-a-lora-on-v-pred-sdxl-model

and settings from the guide here: https://civarchive.com/articles/8723/vlora-or-how-to-train-a-lora-on-v-pred-sdxl-model

The second program's settings worked much better than the first guide's, yet they recommended using a trainer mainly written in a foreign language. OneTrainer was made in English and also supports masked training, so I ported the applicable settings over and made that my go to training method. Said setting changes were

All Data Types = bFloat16

Optimizer = AdamW8-bit

General + Unet Learning Rate = 0.0001

Learning Cycles = 1.0
Text Encoder Learning Rate = 0.00001
Accumulation Steps = 4
LoRA alpha = Same as LoRA rank (32)