








It is a character LoRA of Senko-san from Sewayaki Kitsune no Senko-san.
This LoRA was trained on Pony Diffusion V6 XL and supposed to work with this checkpoint.
Both v1 and v2 were trained on the same dataset, the difference is only hyperparams used for training.
I also had some other tries - the last epochs with different hyperparams can be found on my Huggingface page.
senko_ponyxl_v1 (senko_ds6_ponyxl_lr3e-4_constant_adafactor_dim32_alpha16):
Seems to me it is a bit overtrained as it tends to change composition and work a bit worse with artist tags and style-loras (at least in scope of tests I did before publish these loras).
senko_ponyxl_v2 (senko_ds6_ponyxl_lr1e-5_constant_adamw8_dim32_alpha16):
Works better with different styles and keeps original composition.
Seems to me PonyXL has some kind of bias to ???? shoulders when we try to draw miko clothing, it doesn't even dependent on this lora. To make things a bit better, I would recomend you to use this positive:
miko, japanese clothes, ribbon mesh sleeves
and add this to negative (you can use different weight of course):
(collarbone, ???? shoulders:1.5)
I suppose this lora can have a bias in direction of drawing part of original miko suit if you try to prompt other clothes. In case you face with that issue, I would recommend you to put this in negative prompt:
miko, japanese clothes, ribbon mesh sleeves
It also may worth to add apron to negative.
1. 轉载模型僅供學習與交流分享,其版權及最終解释權归原作者。
2. 模型原作者如需認領模型,請通過官方渠道联系海藝AI工作人員進行認證。我們致力於保護每一位創作者的權益。 點擊去認領
