• 该模型基于 Animagine XL V3.0 Base 底模训练

• 用法：

  • 无需提示词触发。

  • 推荐大模型：Animagine XL V3.1。

  • 不使用 VAE。

  • 采样器：Euler a，采样步数 28 步。

  • CFG Scale：6-7。

  • T2I 分辨率：不低于 768，推荐1280*。

  • Lora 权重：1.0。

  • 开启 ADetailer 可以更好地修复手和脸。

  注意事项：

  • 衣服在训练时作为训练内容，可能难以替换。

  • 如果出现动作promote难以推理，请更换或者使用相似的promote。

  该 Lora 模型的训练数据集截取自番剧《GIRLS BAND CRY》第二集&第三集，采用标准优化器 AdamW 训练。不推荐使用其他模型推理，个人推测可能是因为训练集画风一致造成，此次训练未使用正则去规范画风，旨在还原动漫角色。

• This model is based on the Animagine XL V3.0 Base model.

• Usage:

  • No prompt required.

  • Recommended base model: Animagine XL V3.1.

  • No VAE.

  • Sampler: Euler a, 28 sampling steps.

  • CFG Scale: 6-7.

  • T2I resolution: Not less than 768,1280* is recommended..

  • Lora weight: 1.0.

  • Enabling ADetailer can improve the generation of hands and faces.

• Notes:

  • Clothes may be difficult to replace as they were part of the training data.

  • If the action is difficult to infer, please replace it or use a similar action.

The training dataset of this Lora model is intercepted from the second & third episodes of the anime "GIRLS BAND CRY" and trained with the standard optimizer AdamW. It is not recommended to use other models for inference, which I personally speculate may be because the training set has a consistent style, and no regularization was used to normalize the style during this training, which aims to restore anime characters.