This LoRA is designed to leave a single afterimage from a fellatio, primarily retaining details around the mouth area.
Some afterimages also retain details extending to the eyes and head, but I plan to release those as a separate project.
Training Dataset
59 images in total:
31 full-body images
28 cropped images focusing on the afterimage details
Known Issues
When using an FT model, afterimages are less likely to appear, and their details tend to become blurry. This is likely due to compatibility with the training images, but the LoRA barely works with textures such as those found in MiaoMiao Harem.
The cheeks tend to appear sunken even without using the
":>="tag. This is caused by bias in the training dataset.Does not work with
from below, as there are no directly bottom-up angle images in the training dataset.
Description
Initial.
dim16/alpha16, AdamW8bit, UNet LR 1e-04 693steps(1024² 59 images constant 531 steps + 512² 80 images cosine 162/269 steps), .