妲己 - 穿越火线HD | DaJi - CrossfireHD | Pony

1、概述|General description

此lora训练的是CFHD中的角色妲己，使用游戏内截图进行训练。训练使用的底模为ponyV6XL，因此在其他pony模型中应该也有不错的泛用性。训练集中主要包括角色的脸部特写和膝盖以上的半身照，通过P图的方式移除了角色身上的武器装备，并抠图后替换为黑色背景。游戏内角色如图所示。

This is a character from CFHD.I trained this lora with ingame screenshots. The training base model is Pony Diffusion V6 XL ,so it should has good versatility in other Pony checkpoints. The training set mainly consists of facial close-up and cowboy shot,I also removed weapon equipments and adjust to black background with photoshop.The character in the game is shown in the picture.

2、使用方法|Use guide

• 触发词:djt

• 训练时使用的提示词(可以作为正向/反向提示词框中调整角色特征):animal ears,paint dress,white panties, white headband,cleavage cutout,closed mouth

• 推荐lora权重为 0.7-0.8，单人图片可以调高，多人图片建议调低。

• 如果单人图脸部效果不好，推荐在文生图过程中直接启用ADetailer插件进行修复。如果是多人图片，推荐把文生图的结果（不启用ADetailer）发送至图生图重绘，然后再调用ADetailer插件或蒙版插件手动修复每个人物的脸部。

• Trigger word:djt

• Trained prompts(can be used as postive/negative prompts to control the characteristic):animal ears,paint dress,white panties, white headband,cleavage cutout,closed mouth

• Recommanded lora weight is 0.7-0.8.When generating solo girl image, you can raise the lora weight.When generating muti person image, i suggest you to lower the lora weight.

• Recommand using ADetailer to fix face automatically when generating solo girl text-to-image pictures.When generating muti person images, I suggest you to send the text-to-image picture to image-to-image and repaint it, meanwhile use ADetailer or mask plug-in to fix the faces manuly.

3、注意事项|Note

• 建议使用pony类型的主模型出图，此lora能否搭配SDXL主模型出图我没有进行测试。

• 在我测试过程中，当lora权重高于0.8时，且提示词输入dress类别,裙子上会产生训练集照片上的花纹。如果不希望看到花纹，可以将权重调整至0.8或以下，或者加入更多的提示词限制服装样式。

• 由于训练集图片使用PS软件去除了角色身上的武器装备，这些部位的纹理是通过软件修复的，因此不太清晰，建议在提示词中输入你自己想要的服装样式。

• 如果直接下载并套用我展示图里的设置出图，记得检查lora/embedding/checkpoint是否一致（特别是展示的涩图），否则无法得到到类似的图。

• Rcommand to use pony style checkpoints.I have not tested if this lora works with SDXL checkpoint.

• In my test, there will be pattern on the dress if the weight is higher than 0.8. It is recommand to set weight to less than 0.8 or add more prompts to improve this situation.

• Since the dress texture of the training set is incomplete(processed by Photoshop), I suggest you to specify clothes style in prompts.

• If you want to apply the same set as my posts,remember to check if there is any other difference in lora/embedding/checkpoint.