Saori, burgundy hair, low ponytail, side-swept bangs, reddish-brown eyes, large breasts,mature female

Additionally, through testing, I found that using CFG Scale 4 / Steps 30 tends to produce relatively stable results.

However, these settings are only based on my own testing and should not be considered absolute or fixed values. There are certainly people with more experience who may know better settings or methods, so please treat these values as a reference rather than a strict rule.

Saori had already been living as the sexual slave of a university professor before she married. With her master’s permission, she married her current husband, but her submissive relationship with the professor—and later his son—continued even after marriage. When Saori’s son accidentally discovers evidence of her hidden past, his parents take him to meet her master and reveal the truth: behind her ordinary life as a wife and mother, Saori has remained bound to this long-standing master-slave relationship. Her husband has known about and accepted it all along.

Although the review itself is short, the content is quite intense and the artwork is in color, so I used this work as a test case to train and experiment with a LoRA model.

Something rather unexpected happened with this model.

V1 was trained with the goal of reproducing the original character as closely as possible, and the preview results looked very good. However, after generating and testing more than ten images, the actual results were much less consistent than expected, especially in terms of composition and character reproduction.
So, please do not download the V1 model itself. The images uploaded under V1 are only meant to serve as a reference for what the original character looks like.

On the other hand, V2 was originally a version I was planning to delete because I was not satisfied with it at first. However, after testing it several times, it unexpectedly produced much better and more stable results than V1, so I decided not to discard it and uploaded it as a new version instead.

For reference, V1 was trained for 2000 steps, while V2 was trained for 500 steps. Ironically, the 500-step V2 ended up performing better in actual generations, which was quite surprising to me.