The Herta
The Herta v1.0
使用110张截图作为素材训练，参数按照官方建议设置。在Anima-Standalone-Trainer上进行训练。
Using 110 screenshots as the training material, and setting the parameters according to the official recommendations. The training was conducted on Anima-Standalone-Trainer.
我在这里分享我的参数设置。
I am here to share my parameter settings.
Learning Rate LoRA 学习率 0.00002
Optimizer 优化器 AdamW8bit
LR Warmup Steps 学习率预热步数 200
Text Encoder LR 文本编码器学习率 0
LR Scheduler 学习率调度器 Cosine
Weight Decay 权重衰减 0.01
Tag Dropout Rate 标签丢弃率 0
Caption Dropout Rate 描述丢弃率 0.05
Shuffle Captions 打乱标注 False
Network Dim (Rank) LoRA秩 32
Network Alpha LoRA缩放系数 32
Train UNet Only 仅训练UNet True
实际上我选用的是第20轮的模型，因为后续的模型生成的图片有些奇怪。
In fact, I chose the model from the 20th round because the pictures generated by the subsequent models were a bit strange.
崩坏星穹铁道的大黑塔
The Herta of Honkai:star
权重0.8-1.0
Weight: 0.8 - 1.0
采用游戏截图来训练，之后会加入官方人物模型的渲染图作为补充，这个版本并不完美，对她的发型发饰等还原比较好，但是官方服装的一些细节不能很好的还原，还可能存在一些未知的问题，我的时间有限没办法把所有的问题都测试出来。
The training is conducted using game screenshots. Later, official character model renderings will be added as supplementary content. This version is not perfect. It does a good job in restoring her hairstyle, accessories, etc., but some details of the official costumes cannot be accurately restored, and there may be some unknown issues. My time is limited and I am unable to test all the problems.
只需要输入The Herta就能复现角色，输入其他服装则可以替换为其他服装，anima模型的泛化性很不错，不太容易过拟合到官方服装。
Just input "The Herta" to re-create the character. Inputting other costumes can replace them with other outfits. The generality of the anima model is quite good; it is not prone to overfitting to the official costumes.
我采用的是自然语言和danbooru标签混合的方式来训练，用自然语言描述人物的动作背景服装，danbooru标签来对服装细节、神态、镜头等进行补充。
I used a combination of natural language and Danbooru tags for training. I described the characters' actions, backgrounds and costumes in natural language, while the Danbooru tags were used to supplement details of the costumes, expressions, and camera shots.
有什么建议可以在评论区交流，我会用来迭代我的训练参数，训练出更好的lora。
If you have any suggestions, please share them in the comment section. I will use them to iterate on my training parameters and train a better LORA.
之后会尝试推出更好的lora。
We will then attempt to introduce a better version of LoRa.