KoreanStyle2.5D

Updated 2023/3/15

新加入了3张韩风预览图，试了一下宽画幅，好像效果也OK，主要是想提醒大家这是一个韩风模型，然后告诉大家韩风模型需要用的Lora！最后提醒大家一定要用hires-fix对图片进行高清修复。

Newly added 3 Korean style preview images, tried a wide format, seems not bad. want to remind everyone that this is a Korean style model, and you can use some Korean style Lora to improve the quality of generated images. At last, do not forget to use hires-fix on the picture HD repair.

........................................................................................................................................................

一款韩风的2.5D模型，能够在尽量保证画风足够真实、好看的同时又保留有足够的创造力。

English introduction is below.

推荐与以下几款Lora搭配使用： fashion_girls,ulzzang-6500-v1.1,chilloutmixss20, korean_doll_likeness, 权重大概在0.2到0.3就可以形成比较好看的韩风脸。也可以与各种概念类的lora搭配，如Pillory 枷具、millking machine等，经测试效果都比较好。

还有一点比较重要的是，本模型非常推荐使用hires高清修复，能够极大提升细节质量，没有副作用，并解决有时候脸部崩坏的问题。

················································································································································································································································································································

下面我给一个生成样例展示一下如何同时结合韩风脸外加从那些概念生成的lora，这里以Pillory 枷具为例：

该Lora可以在这里下载：

https://civitai.com/models/15132/conceptpillory-jiaju

(masterpiece),(best quality),extreamly delicate and beautiful,illustration,straight-on,nude, 1girl,1boy, pillory, bound wrists, bondage, bound,,very long hair, (forehead:0.98), breasts, cleavage, navel,thighs, saliva,chains, ass,hetero, condom, full-face blush, embarrassed, sweat,city,street,surrounded, multiple others, exhibitionism, audience, background characters, crowd, public humiliation pov , <lora:ConceptPilloryJiaju_v10:0.51>, ulzzang-6500-v1.1 <lora:chilloutmixss20_v20:0.3> <lora:koreanDollLikeness_v15:0.2>

Negative prompt: EasyNegative,worst quality,low quality

Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 4030484714, Size: 512x512, Model hash: 0e9c1bdca1, Denoising strength: 0.41, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires upscaler: Latent

这个样例中既生成了比较好看的韩风脸，又也同样用好了通常只能在类似AbyssOrange那种动漫风的模型中才能用到的概念：枷具。这个也就是我创作这个模型的初衷：既保留好看的韩风脸，也不会丢失创造力。当然，预览图应该过不了审，我就不放出来了，大家自己尝试就好。

················································································································································································································································································································

先说一下为什么要创造这款模型，众所周知,chilloutMix模型是现在很流行的真人类模型，但是它非常依赖于lora,没有lora人物出来就很不好看，除此以外，由于训练集过于稀少，所以其实很多tag对它都没有用，很多衣服、挂件、姿势都摆不出来。所以我就思考，也许能够把它和一些二次元模型融合，然后让它学到二次元画中的姿势、衣服等等。所以就创造出了这款模型，画风大概是介于2、3D之间的。

与chilloutMix相比，这个模型的优点在于它不依赖于lora就可以出比较好看的图，然后2.5D画风也让人不容易恐怖谷。（当然我建议还是搭配一些好看的lora效果更好）更重要的是，它可以自由的搭配几乎所有的tag，它从二次元模型中学到了tag的含义，这样你就可以自由的搭配各种衣服、姿势、玩具等，这些chilloutmix很难做到。

跟动漫风格的模型相比，就是模型更加韩风、更加好看了。

················································································································································································································································································································

A Korean-style 2.5D model can ensure that the style is realistic and beautiful while retaining enough creativity.

It is recommended to use with the following Lora: fashion_girls, ulzzang-6500-v1.1, chilloutmixss20, korean_doll_likeness, the weight is about 0.2 to 0.3 to generate a pretty Korean style face. It can also be used with various conceptual lora, such as Pillory , millking machine, etc. and the test results performed well.

The model is highly recommended to use hires-fix, which can greatly improve the quality of details, has no side effects, and solves the problem of sometimes face or eyes collapse.

················································································································································································································································································································

Next, I will give a generation example to show how to combine the Korean style face and the concepts lora generated from AbyssOrange Here we take the Pillory as an example, the Lora could be downloaded at here:

https://civitai.com/models/15132/conceptpillory-jiaju

(masterpiece),(best quality),extreamly delicate and beautiful,illustration,straight-on,nude, 1girl,1boy, pillory, bound wrists, bondage, bound,,very long hair, (forehead:0.98), breasts, cleavage, navel,thighs, saliva,chains, ass,hetero, condom, full-face blush, embarrassed, sweat,city,street,surrounded, multiple others, exhibitionism, audience, background characters, crowd, public humiliation pov , <lora:ConceptPilloryJiaju_v10:0.51>, ulzzang-6500-v1.1 <lora:chilloutmixss20_v20:0.3> <lora:koreanDollLikeness_v15:0.2>

Negative prompt: EasyNegative,worst quality,low quality

Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 4030484714, Size: 512x512, Model hash: 0e9c1bdca1, Denoising strength: 0.41, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires upscaler: Latent

In this example, not only a good-looking Korean face is generated, but also a concept that usually can only be used in anime-style models like AbyssOrange is also used: Pillory This is the original intention of my creation of this model: to retain the pretty Korean style face without losing creativity. However, the preview image should not pass the review, so I won't release it. You can try it ourself.

················································································································································································································································································································

Let me first talk about why this model was created. As is known to all, the ChilloutMix model is a popular realistic model, but it is dependent on lora, and does not perform well without lora characters. In addition, due to the scarce of the training dataset, many tags are useless in CHilloutmix, many clothes, pendants, and poses cannot be displayed. So I thought, maybe it can be merged with some anime-style models to learn the tags of poses, clothes, etc. So this model was created, the painting style is between 2D and 3D.

Compared with chilloutMix, the advantage of this model is that it can produce better-looking pictures without relying on lora, and the 2.5D painting style is not easy to make the uncanny valley. (Of course, I suggest to use it with some beautiful loras) More importantly, it can match almost all tags freely, and it has learned the meaning of tags from other style, so that you can freely match various clothes, poses, toys, etc while chilloutmixe is hard to do.

Compared with the anime-style model, the model is more Korean-style and beautiful.