This LoRA is designed to generate a single tongue afterimage depicting a tongue moving either horizontally or vertically.

For v0.0.0, I trained the LoRA on 30 images in total:

• 25 training images, with manga panels and multi-scene illustrations cropped down to a single panel where necessary

• 5 cropped images focusing on the specific area targeted by the LoRA

Of the 30 images, 26 feature 1girl, while the remaining 4 feature 2girls in cooperative fellatio. As a result, reproducibility tends to decrease when more than one person is involved.

Changes from 2way-afterimage_tongue_v0.0.0

The training images were completely replaced, the training resolution was increased, and the optimizer was changed.

Previous version:
dim16/alpha8, Lion8bit, UNet LR 1e-05, cosine_with_restarts[2] 1,400 steps, 10 images (512² resolution), 1,400 steps

v0.0.0:
dim16/alpha16, AdamW8bit, UNet LR 1e-04, cosine 1,200 steps, 30 images (1024² resolution)

The most noticeable improvement is the reduction in image quality degradation. The new version preserves image quality more effectively compared to the previous LoRA.