This is an experimental early release. I believe Mr. MetalchromeX will likely come up with something even better than this LoRA. V1 has been released, yay!
The training data focuses specifically on "pressed-up buttocks" and "skin ripples/waves" from various angles and positions. Please note that it may not be ideally suited for "imminent penetration" scenes, as the emphasis is on the physical contact and pressure.
Description
v0.2.3 Changelog (Changes from v0.1.0)
Dataset Updates: Added 45 new training images. The breakdown consists of 21 newly added monochrome doujinshi close-up crops and 24 close-up crops of the 25 existing training images. By ensuring that closely cropped images of the target area make up more than half of the entire dataset, the goal is to achieve efficient concept learning even with low Dim and low Learning Rate (LR).
Trigger Word & Captioning Changes: Instead of using the default Danbooru tag
ass ripplestandard to Anima, the training was conducted using independent custom keywords:ar_v02andar_ext_v02. Images with a single ripple are captioned asar_v02, while those with double or multiple ripples are captioned asar_ext_v02. Trained using fixed 1-caption shuffling.Artist Tag Removal: Previously, artist-specific elements within the captions included the actual artist tags during training. These artist tags have now been completely removed.
Training Configuration Changes: Changed the learning rate scheduler from
constant(LR:5e-05) tocosine(LR:1e-04, warmup steps = 100), and lowered the Dim from 16 to 8. This adjustment aims to suppress art style contamination.
dim8/alpha8, unet 1e-04 (adamw8bit / cosine), warmup_steps=100, 70 images / 1000 steps