右上角打开“everything”模式可以查看更多封面图片
我不知道花子的泳装图为什么被吞掉！这个lora确实是泳装版本，如果你没有看到我发的图片，大概率是平台审核不通过，但是我并没有露出任何敏感的地方
I'm not sure why the picture of Huazi in a swimsuit was removed! This Lora model is indeed a swimsuit version. If you didn't see the picture I posted, it's likely that it didn't pass the platform's review process. However, I didn't expose any sensitive areas.
这个模型可以在保留角色特征的情况下实现换装，或者不覆盖任何衣物。
This model can achieve outfit changing while preserving the character's features, or without covering any existing clothing.
这是碧蓝档案的花子的泳装形态，因为目前训练的数据集有限，还不能做到很好的还原。
因为我在训练是打的是全标，因此你如果想调动该角色，有以下标签是你必须添加的。
必需的positive prompt ：
<lora:swhuazi_v1:0.8>,huazi, girl, breasts, pink hair, green eyes, solo,long hair, braid, ahoge
创作出泳装形态需要的positive prompt ：
<lora:swhuazi_v1:0.8>，huazi, 1girl, shirt, solo, large breasts, cleavage, white shirt, smile, bow, hair bow, blush, water, wet clothes, wet, see-through long sleeves,wet shirt, collared shirt, very long hair side braid
建议权重：0.8
右边这个是角色浦和ハナコ的 sdxl版本
我并不建议你下载这个模型，这个模型仅仅是一个不成熟的lora！
这个模型的起因是我想测试如果用同样的训练集，相似的参数，sd1.5和sdxl模型训练出来lora区别会多大。
从效果上来看，说不上很好，也说不上很差。
我准备了15张数据集，repeat为20，batch为8，ep为2。
我在训练时准备了263张正则化图片，这些图片搜集自pixiv。
我的训练脚本是https://github.com/bmaltais/kohya_ss。
我训练的设备是3090双卡训练。
训练一次大概是20分钟左右。
从测试上来看，这个lora的最佳权重是0.8。
这个lora会出现时而二次元，时而三次元的情况。
该lora在1000以上分辨率，效果会大大变好。
我对sdxl训练并不熟练，我正在找出导致这个情况发生的原因。
This is the swimsuit form of Huazi, but due to the limited training dataset, it cannot be accurately reproduced at the moment.
Since I used full tags during training, there are certain tags that you must add if you want to generate this character.
These are the required positive prompts:
<lora:swhuazi_v1:0.8>， huazi, girl, breasts, pink hair, green eyes, solo, long hair, braid, ahoge.
To create the swimsuit form, you will need the following positive prompts: <lora:swhuazi_v1:0.8>，huazi, 1girl, shirt, solo, large breasts, cleavage, white shirt, smile, bow, hair bow, blush, water, wet clothes, wet, see-through long sleeves, wet shirt, collared shirt, very long hair side braid.
It is recommended that you use a weight of 0.8 for these prompts.
This is the sdxl version of the character Hanako Puruha.
I do not recommend downloading this model, as it is an immature LORa model.
The reason for this model is that I wanted to test the difference between using the same training set and similar parameters to train LORa models in SD1.5 and SDXL.
In terms of performance, it's neither good nor bad.
I prepared 15 datasets with a repeat of 20, batch size of 8, and 2 epochs.
During training, I prepared 263 regularized images collected from Pixiv.
My training script is available at https://github.com/bmaltais/kohya_ss. ↗
I trained the model using a 3090 dual-card device, and each training session took about 20 minutes.
From testing, the best weight for this LORa model is 0.8.
This LORa model can sometimes produce two-dimensional or three-dimensional results.
The performance of this LORa model greatly improves at resolutions above 1000.
I am not proficient in SDXL training and am currently trying to identify the cause of this situation.