Easily Create eye-catching images for 萧炎之丹（一种超级温柔的插画风格）【A soft illustration style】

【中文版】

这是本菜菜炼的第一个丹，很多参数都还没有搞明白（老爸奖励俺一台新电脑，实在是迫不及待开始炼丹），因此lora出来效果并不是很稳定。这个画风本身五官就与一般美女五官画法不一样，因此在很多大模型下生成的图片会畸形。emmm...因为本来就是菜菜第一个丹，如果大家喜欢的话，我日后会再研究一下参数，多多完善一下这个lora！!

然后，使用的话，有几个参考建议，可以让这个模型出图稍微稳定一些：

1、如果要生成理想的（我预设的风格），大模型最好使用counterfeit 2.5的版本。

2、强烈建议大家使用controlnet（我用了reference only）去控制人物！！不然无论是色彩、画风还是五官，都会有点奇怪。

3、采样的话，推荐使用Euler a，或者DPM++ 2M Karras。其他的我试的时候感觉跟自己预想的风格差异有点大，如果你们喜欢可以试一试。

4、这个模型最大的问题在于它的眼睛，经常会生成不对称或者很丑陋的半眯状态。我自己想了几种解决办法。第一种，在prompt中加上symmetrical eyes（意思是对称的双眼？我用百度翻译的），在negative prompt中加easynegative等一些生成脸部比较差的embedding加以控制。第二种，使用controlnet，这是2一样的原理。第三种，减少lora的权重，或者如果你仅仅喜欢这个lora的色彩，也可以试试分层lora，不要让lora影响到五官。第四种，生成之后通过Photoshop等后期软件对眼睛进行调整（可能这个方法还是最快的，笑）

这个画风我很喜欢，冷暖色的对比很夸张因此颜色像彩虹一样好看，颜色饱和度低，色彩有点莫兰迪色调的味道，人物的五官也与一般的模型相比多了一层优雅和忧郁的感觉（五官真是又爱又恨），线条也更加柔和。我炼的时候底模就是counterfeit 2.5，因此我不知道其他大模型之下它的效果怎么样。如果大家能炼出一些酷酷的效果，也欢迎分享。

关于这个模型的名字，萧炎是我好姐妹的爱称，也是斗破苍穹里面的一个会炼丹的帅哥（好像是？），以她的昵称来命名我炼的第一个模型以表达我对她的爱，她一定会很感动的（笑死）。

国际友友看这一版：

【English Version】

This is my first lora, and many parameters are still unclear (my dad rewarded me with a new computer, and I can't wait to start making lora), so the effect of lora is not very stable. The facial features of this painting style are different from those of ordinary beauties, so the pictures generated under many large Checkpoint models will be deformed. emmm...Because it is the first lora of mine, if you like it, I will study the parameters again in the future and improve this lora! !

Then, if you use it, there are several reference suggestions that can make this model a little more stable:

1. If you want to generate an ideal (my preset style), it is best to use the counterfeit 2.5 version for the large model.

2. It is strongly recommended that you use controlnet (I used reference only) to control characters! ! Otherwise, whether it is color, painting style or facial features, it will be a bit strange.

3. For sampling, it is recommended to use Euler a, or DPM++ 2M Karras. When I tried the others, I felt that the style was a bit different from what I expected. If you like it, you can give it a try.

4. The biggest problem with this model is its eyes, which often produce asymmetrical or ugly half-closed states. I thought of several solutions myself. The first one is to add symmetrical eyes to the prompt (meaning symmetrical eyes? I translated it using Baidu), and add some embeddings such as easynegative to the negative prompt to control the poorly generated faces. The second is to use controlnet, which is the same principle as 2. The third way is to reduce the weight of lora, or if you just like the color of this lora, you can also try layering lora, and don't let lora affect the facial features. The fourth type is to adjust the eyes through post-production software such as Photoshop after generation (maybe this method is the fastest, lol)

I like this painting style very much. The contrast between warm and cold colors is exaggerated, so the colors are as beautiful as a rainbow. The color saturation is low, and the colors are a bit like Morandi. Compared with the general model, the facial features of the characters are more elegant and It has a melancholic feel (the facial features are really love-hate), and the lines are softer. When I practiced, the counterfeit 2.5 was the base model, so I don't know how it works with other large models. If you can refine some cool effects, welcome to share.