This is an experimental initial release. This LoRA focuses specifically on coating the penis with various bodily fluids.
In my own dogfooding, it shows some subtle improvements compared to not using the LoRA at all, though it’s still far from perfect.
Description
Changes from v0.1.1 to v0.2.2
The training dataset has been expanded from 26 images to 90 images.
61 base images and 29 cropped close-up images were used, covering wet penises from a variety of angles and situations.
Some stylistic bias is still present. v0.2.2 uses works from 23 different artists, compared to 18 artists in v0.1.1.
The amount of bodily fluid may vary depending on the FT model being used.
Training settings have been adjusted:
v0.1.1: 16 dim / 8 alpha, 5e-5 learning rate, 1300 steps with cosine scheduling
v0.2.2: 8 dim / 8 alpha, 1e-4 learning rate, 600 steps constant + 400 steps cosine
Training resolution remains unchanged at 512².
dim8/alpha8, UNet LR 1e-04, AdamW8bit, warmup 0, constant 600 + cosine 400steps, 512² reso, 90 images.