如果找到了正确的方法，训练这个画风还是很简单的在Anything和NAI上表现良好（建议直接使用NAI即可），在其他底模上画风拟合性能未知，有兴趣可以自己尝试一下。没怎么测试在其他大模型上的泛化性，尝试在NAI下，与自训练的白圣女模型共同出图，得到了一张Sample。可以看到在我自己的人物lora上展现了良好的泛化性。即使是练的稍微差一点的江江也能完全驾驭的住建议与easynegative一起使用，提示词加入highres（不加应该也不会有多大影响）It may be quite easy if you find the right way to train it.Good performance was achieved on Anything series model and NAI model (recommend using NAI). The performance of art style is unknown on other base models, you can try it yourself if interested.Lack testing generalization on other base models. I have attempted using NAI model, and got a sample image with my own trained Cecilia character lora, which seems really nice. Even my poorly-trained Suzushima Arisu character lora can work well too.I recommend to use with easynegative embedding. ‘highres’ should be added in prompts (as trigger), and it’s ok if you don’t add it in.

Files with SHA256: 6888f9ee76353d5d7b5768f72032e485531c1bdcad0632f98b2d5bb64db8ff46