This LoRA was trained on 76 training images in which the breasts appear to move back and forth between two points through the use of a solid form and afterimages.
As a result, afterimages tend to appear in bottom-to-top, top-to-bottom, diagonally upward, or diagonally downward directions.
One important limitation is that, in v0.0.0, most of the training images feature vertical up-and-down movement. Therefore, this LoRA is not capable of reliably reproducing intense side-to-side movement from a direct side view.
Description
Changes from v0.0.0
Training image count and resolution: The dataset has been reduced from 76 images at 512² resolution to 40 carefully selected images at 1024² resolution.
Images with clearly defined nipples in the afterimage were selected, while training images with blurry or poorly defined afterimages were removed.
Reduced tendency to enlarge
small breastsandflat chest: This was caused by the lack of training images featuring smaller breasts. Of the 40 newly selected images, 10 featuresmall breasts, and 3 featureflat chest(including 1 cropped image). The remaining 27 images continue to feature the large-breast afterimages used in the previous version.
Unlike v0.0.0, the effect where the afterimage becomes stretched between the original image and the afterimage, creating a ghost-like smear frame, is more subdued.
Depending on the situation, the previous version may produce a stronger or more desirable effect, so keep this in mind when choosing between the two versions.
dim16/alpha16, AdamW8bit, UNet LR 1e-04 constant 800 steps, 1024² resolution, 40 images.