这个模型的训练集包含了来自伤物语剧场版的627张羽川翼的高质量帧采样图片。能够较好还原伤物语特有画风下羽川翼特征的同时，也能够相对自由的改变发型/穿着/动作/表情/配饰。训练的底模依然是作者常用的animefull-latest。总共训练了6个Epoch，repeats为8。每轮的Lora泛化表现略有差异，因此作者会上传包含对6个Lora在相同参数和提示词条件下出图的XYZ Plot。可以根据喜欢的图片对应的Lora进行下载。

推荐的参数是0.2-0.6。大于0.6的参数会更加趋近于剧场版的画风。但相应的，麻花辫/眼镜/直江津校服等特征将无法被有效分离。值得一提的是，虽然标签打的比较散碎，但服装/发型/脸部特征主要集中于“naoetsu_high_school_uniform,school_uniform,braid,single braid,glasses"。可以通过将你不希望体现的关键词加入到负面提示词中来控制相应特征的被表达程度(可以参考作者放出的样图中的提示词写法)。

最后，班长可爱捏！以及，

本模型生成的图片请勿用于任何商业用途！！

The training set of this model contains 627 high-quality frame-sampled images of Hanekawa Tsubasa from the theatrical version of Kizumonogatari. The model is able to reproduce the characteristics of Hanekawa Tsubasa in the unique style of the Kizumonogatari. And at the same time, it is also able to change the hairstyle/dress/movement/expression/accessories with relative freedom. The base model used for training is still the animefull-latest, 6 Epochs were trained, and the repeats is 8. The generalization performance of each Epoch of Lora is slightly different, so the author will upload the XYZ Plot which contains the same parameters and key words for all the 6 Loras. And you can download it the Lora corresponding to your favorite image in the XYZ Plot accordingly.

The recommended parameter is 0.2-0.6, the parameter greater than 0.6 will be more similar to the style of the theatrical version. But accordingly, features such as braid/glasses/Naoetsu school uniforms will not be effectively separated. It's worth noting that although the tags are scattered, the clothing/hair/face features are mainly focused on "naoetsu_high_school_uniform,school_uniform,braid,single braid,glasses". You can control the degree of expression by including keywords you don't want to be reflected in the negative cues (see the cues in the sample image released by the author).

Finally.

Please do not use the images generated by this model for any commercial purposes!!