All text translated into English using DeepL :-(
Ysera seems to have less popularity than her sister. However, I wanted to do a LORa on her as well. For the dataset, 61 images were taken from rule34.
However, I'm still new to training and don't understand how it works at all. So I just follow other people's guides.
This LORa is trained on Autismmix_confetti.
Words Trigger: YseraCute
Support words: colored skin, purple skin, pointy ears, green hair, long hair, horns, blue eyes, glowing eyes
But you are free to change hair length, chest size, etc.
Recommended strength: 0.8-1.0
I've noticed the following problems:
1. LORa has a big impact on generation style. (Although it itself is also good at changing the style with other LORa)
2. due to the fact that in the dataset she has almost all pictures with a tiara on her head, which looks like extra horns, extra horns can be generated without the tiara.
3. sometimes the eyes break. I recommend fixing this with ADetailer.
4. sometimes draws a white glare on the lower lip.... I have no idea how to fix it T_T
p.s: The second picture from the example shows the result with LORa on style and some fixes in ADetailer and Photoshop. The two HornGrab pictures are made to demonstrate that LORa is very amenable to generation style changes. On one LORa to PersonalAmi style, and on the other the style changed ONLY from LORa to HornGrab.