![SFW, (grassland), outdoors
mejiro_ardan_SF, 1girl, (solo), horse_ear, horse_tail,light smile, full body
(long hair), (light_blue_hair), purple_eyes, hair ornament, collarbone, (black and whtie frilled dress), center frills, hemline, (uneven sleeves:1.2),detached sleeves, puffy sleeves, [short sleeves:long sleeves:0.2], (uneven gloves:1.2), (one hand wears black glove and the other one wears white glove), (uneven_boots:1.2), (ankle boots), (uneven legwear:1.2)](data:image/gif;base64,R0lGODlhAQABAAD/ACwAAAAAAQABAAACADs%3D)
Quick Generate
Model Description by Creator
导言
这是一个《Umamusume Pretty Deby》《闪耀优俊少女》的目白阿尔丹角色LyCORIS，其与其他同角色lora的区别在于，添加了新的Neige Émeraude卡面立绘。在使用中使用cut off插件（https://github.com/hnmr293/sd-webui-cutoff）防止元素污染会获得更好的效果。
示例图使用了Viewermix模型进行生成。
使用方法
建议以0.75左右的权重使用。
在打标过程中，为隔离元素污染，采用了每个卡面分开设关键词的方法，以下是各个立绘的关键词：
私服：mejiro_ardan_CA
Neige Émeraude：mejiro_ardan_NE
夜風に舞うは祈りの花：mejiro_ardan_PFN
特雷森制服：mejiro_ardan_UN
胜负服：mejiro_ardan_SF
其他非特殊立绘：mejiro_ardan
训练细节
以NAI Final作为底模训练，使用了约500张图片作为训练集，将约一半图片作为回归训练集，每张图片重复训练10次，训练10个epoch，最终步数为5150步。
以下为主要设置参数：
network_dim=32
network_alpha=32
keep_tokens=8
conv_dim=4
conv_alpha=4
lr="1e-4"
unet_lr="1e-4"
text_encoder_lr="1e-5"
batch_size = 2
局限性
人物立绘中的不对称设计的还原难度非常大，基本上只能依靠随机抽卡和手动辅助达到一定的还原效果。
在出示例图的过程中没注意到Additional Network插件失效，一些LyCORIS的层没有输出，可能对最终结果有所影响。
Introduction
This is a LyCORIS of Mejiro Ardan from "Umamusum Pretty Deby". The difference between other similar characters Lora is the addition of Neige Émeraude. Using the cut off plugin（ https://github.com/hnmr293/sd-webui-cutoff ）to prevent elemental pollution will achieve better results.
Example images are generated by Viewermix model.
Usage
Suggest using a weight of around 0.75.
During the marking process, in order to isolate element contamination, a method of setting keywords separately for each card surface was adopted. The following are the keywords for each vertical drawing:
Casual clothes: mejiro_ ardan_ CA
Neige Émeraude：mejiro_ ardan_ NE
夜風に舞うは祈りの花: mejiro_ ardan_ PFN
Teresan uniform: mejiro_ ardan_ UN
Shobofuku：mejiro_ardan_SF
others：mejiro_ardan
Training Details
Using NAI Final as the base model for training, approximately 500 images were used as the training set, and about half of the images were used as the regression training set. Each image was repeated 10 times for 10 epochs, with a final step count of 5150.
The following are the main settings parameters:
network_ dim=32
network_ alpha=32
keep_ tokens=8
conv_ dim=4
conv_ alpha=4
lr="1e-4"
unet_ lr="1e-4"
text_ encoder_ lr="1e-5"
batch_ size = 2
Limitations
It is very hard to draw the asymmetrical designs in character portraits. Maybe a lot of tries and manual assistance is a certain way to achieve good results.
During the process of presenting the example image, I did not notice that the Additional Network plugin was invalid, and some LyCORIS layers did not output, which may have an impact on the final result.
