This LoRA creates an afterimage effect, making parts look like they are rapidly moving back and forth between two positions.
In v0.0.1, you can expect the effect to work reasonably well with basic positions, but it is currently a bit finicky and highly experimental. Due to gaps in the training data, the effect rarely works for fellatio or handjobs (these are included in the dataset, but with so few patterns that they are barely reproducible). It also completely fails to activate from a side angle (from side).
I plan to address and fix these issues in the next version.
Description
arm_v0.0.0 (Initial Release) This is the initial release of the arm and wrist afterimage LoRA. The dataset specifically targets actions where the arm or hand appears in duplicate due to afterimages, focusing on "waving," "handjob/fellatio gestures," and explicit "handjobs."
Important Notes:
Dataset Variety & Effect Strength: Because the training images feature a wide variety of camera angles, the model successfully avoids overfitting. On the flip side, however, it can be somewhat difficult to trigger the LoRA's primary goal: a "highly detailed afterimage." You can slightly increase the success rate by boosting the weights of the prompt, for example:
(handjob, 2way-afterimage, afterimage, motion blur:2).Regarding "waving": This effect is exceptionally difficult to trigger. You will likely need to boost the weights of the trigger words and run multiple generations (high trial count) to get the desired result.
dim16/alpha8, unet 1e-05(lion8bit/cosine_with_restarts[2]), 40images / 1000 steps