中文版：
分享一下我在Tensor.art训练此Lora的过程。
https://tensor.art/models/806735905185588017
1.数据集处理
图片均来自游戏内截图，裁剪至合适的比例。由于要训练衣服的Lora而不是角色的，所以选择了包含6个角色的共20张图片，其中有一组左侧右侧和背面视角的图片。
上传至Tensor.Art, Tensor.Art会自动打标。再给每张图片的标签加上christmas clothes作为触发词。由于自动打标没有识别角色，也加上了角色的触发词，例如shinosawa hiro \(gakuen idolmaster\)。
2.参数设置
底模用Illustrious, 训练网络模块Lora，触发词用christmas clothes。
图片单张重复15次，训练轮数10，每1轮保存一个。Clip Skip 设为2，其他参数按默认。
样图提示词 1girl, christmas clothes
3.开始训练
训练完成后选择样图效果好的Lora即可发布。
English Version:
Sharing my Lora training process on Tensor.art:
https://tensor.art/models/806735905185588017
1.Dataset Processing
The images are all screenshots from the game, cropped to an appropriate aspect ratio. Since I wanted to train a Lora for the clothes and not the characters, I selected a total of 20 images including 6 characters, with a set of left, right, and back view images.
Uploaded to Tensor.Art, which automatically tagged them. Then I added "christmas clothes" to each image tag as a trigger word. Since the automatic tagging didn't recognize the characters, I also added the character trigger words, such as "shinosawa hiro \(gakuen idolmaster\)".
2.Parameter Settings
Base model: Illustrious, Network Module Lora, trigger words: "christmas clothes".
Image repetitions: 15 per image, training epochs: 10, saving every 1 epoch. Clip Skip: 2, other parameters using default in Tensor.art.
Sample image prompt: "1girl, christmas clothes"
3.Start Training
After training, select and publish the Lora with the best sample image results.
