总之，要是想要更稳定的出图用v2.0好一些，如果想要更好的风格效果用v1.0In conclusion, if you want more stable image output, v2.0 is a better choice. If you want a better style effect, v0.1 is the way to go.V2.0Strength recommendation:0.7~1Prompt：ro1ghe2v2版本稍微解决了一下出图不稳定的问题（比如再没有写提示词的时候会频繁出现乳沟和吊带袜的吊带），但是因为用了新生成的数据集，风格上有些许少许的差别，在跟一些风格相差太大的角色上使用时，图像质量有小概率会变低，在跟其他风格混合的泛用性上不如v1The v2 version has somewhat resolved the issue of unstable image generation (for example, when no prompt words are provided, the breasts and suspenders of the suspenders will frequently appear). However, due to the use of a newly generated dataset, there are slight differences in style. When used with some characters whose styles are too different, the image quality has a small probability of deteriorating. In terms of its versatility when mixed with other styles, it is not as good as v1.V1.0Strength recommendation:0.8Prompt：ro1ghe2这个风格lora模型，是在交流群中的一个朋友通过多个模型混合出的风格，我嫌每次都要调用多个模型太麻烦了。所以就要了他生成的图集，训练了一个风格lora，目前在我常用的模型上使用的效果还算可以，但也有些瑕疵，因为获得的图集内有些元素忘记用提示词标注了，导致很多很多特征时不时的冒出来，需要填写反向提示词来规避，但有些也规避不掉，我想办法在之后的版本调整下，目前模型跟ckpt模型自带的角色是要效果较好，其他的角色lora与之相比，有一定概率效果没那么好This style Lora model was created by a friend in the discussion group by blending multiple models. I found it troublesome to call multiple models each time. So I asked him to generate a collection of images and trained a style Lora model. Currently, its effect when used on the models I commonly use is acceptable, but there are still some flaws. Because some elements in the obtained image collection were forgotten to be labeled with prompt words, resulting in many features popping up from time to time. To avoid this, reverse prompt words need to be filled in, but some cannot be avoided. I tried to adjust it in the subsequent version. Currently, the model and the role provided by the ckpt model have better effects, while the Lora of other roles has a certain probability of not being as good as that of the model.

Files with SHA256: 59481f93d2a52101bea8c324a5353f699ec5c69284ff44ac4c2c00853f97dce4