The goal of this LoRA is to depict full penile insertion together with the deformation of the surrounding flesh caused by close physical contact.
In other words, it is designed to reproduce effects such as ass ripple, ass press, and pussy press resulting from balls deep.
Initial release notes(v0.0.1)
For v0.0.1, I trained the LoRA on 60 images in total:
50 images showing full insertion to the base from various angles
10 cropped images focusing on the deformation of the surrounding flesh
The overall result feels somewhat like a jack-of-all-trades, master of none.
I compared the results with no LoRA, this LoRA, and LangdonAlger's balls deep LoRA. For full insertion to the base and completely concealing the penis inside the vagina, LangdonAlger's balls deep still appears to perform better.
For the deformation of the surrounding flesh, the Anima base model can sometimes produce interesting results. However, FT models often fail to reproduce the deformation. This is likely due to the insufficient number of cropped close-up images.
For the next version, I plan to significantly increase the proportion of cropped close-up images in the training dataset. Based on what I learned from developing my previous two LoRAs, increasing the proportion of close-up images should help the LoRA learn the deformation details more effectively.
As of v0.0.1, this LoRA still needs to be used alongside LoRAs specialized in localized deformation, such as an ass ripple LoRA.
Description
Changes from v0.1.1
This update aims to address the overfitting to POV compositions observed in v0.1.1, which caused the male character's POV crotch to appear excessively, even when not explicitly requested.
To mitigate this issue, the number of POV training images in the dataset was adjusted, and captioning errors were corrected.
The results appear to show an improvement: unwanted pov crotch appearances are less frequent than in the previous version. However, the issue has not been completely resolved and may occasionally recur.
The main changes are as follows:
Reduced the training dataset from 80 to 70 images.
Adjusted the training configuration:
v0.1.1: LoRA, dim32/alpha32, CAME, UNet LR 3e-05, cosine scheduler, 960 steps, 80 images, 1024² res.
v0.1.3: LoRA, dim32/alpha32, CAME, UNet LR 2e-05, constant scheduler, 700 steps, 70 images, 1024² res.