2.0版本
使用了1024*1024的尺寸进行训练，试图解决“克隆人”的问题
使用了400张单人的图片训练训练
优化的模型大小，仅有27M
触发词：1girl,kamisato_ayaka,solo,tassel,breastplate,japanese_armor,skirt,japanese_clothes,dress,armored_dress,
2.0 Version:
Trained using a resolution of 1024x1024 in an attempt to address the issue of "clone images."
Trained on a dataset of 400 individual images.
Optimized model size, only 27MB.
Trigger words: 1girl, kamisato_ayaka, solo, tassel, breastplate, japanese_armor, skirt, japanese_clothes, dress, armored_dress.
1.5版本：
本模型版本采取高质量的角色训练集训练，比上一个版本效果提升了20%
对于官方服装可能没有很好的准确度，因为在训练集中官方服装的占比相当少
神里绫华太难训练了...
1.5 Version:
This model version was trained on a high-quality character dataset, resulting in a 20% improvement compared to the previous version.
It may not have high accuracy for official costumes because the proportion of official costumes in the training dataset is relatively low.
训练模型：SDXL1.0官模
测试模型：Kohaku-XL
建议权重：0.6~0.8
Training model: SDXL1.0 Official Model
Testing model: Kohaku-XL
Recommended weights: 0.6~0.8
Description
FAQ
Comments (7)
I use all your Genshin and (love them btw), however, this one is a royal pain. I am getting a LOT of cloning and multiple people on almost every render. never had that issue using the same parameters on the other Lora's so i assume it has something to do with the way this one was trained. any advice?
I can understand how you feel. It can be mentally exhausting when multiple clones are generated frequently in images. Please send me the parameters of your images, or post them in the gallery, and I will check for any issues they may have.
我采用训练的方法和1.5基本一致，唯一区别就是素材的分辨率提高了。另外，我是在768x1024这个尺寸下训练的，也是在768x1024尺寸生成图片，使用远高于这个尺寸的出图可能会出问题（但是没有人跟我提到过）。当然，最好是能找出发生克隆人现象的原因，这会帮助我改进训练方法。所有的原神角色系列lora，我都会进行持续的改进和升级
I followed a training method similar to 1.5, with the only difference being that I increased the resolution of the materials. Additionally, I trained and generated images at a size of 768x1024, so using resolutions much higher than that might cause issues (although nobody has mentioned it to me). Of course, it would be best to identify the cause of the cloning phenomenon, as it would help me improve the training method. I continuously work on improving and upgrading all the Lora models for the Genshin Impact character series.
@Chenkin posted to gallery.
@Axelreplica lease tell me what dimensions you use to generate images. Also, I have updated my training method. Can you check if the latest 2.0 version of the model "Shinri Akira" still exhibits the phenomenon of generating clone images?
@Chenkin 1024x1024
@Axelreplica Has the issue with the updated 2.0 version of Shinri Ayaka, which was released yesterday, been resolved?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.