米山舞(YoneyamaMai) 2D Style LoRA

中文 | English 使用前阅读 | Read it before use

English is not so good, please forgive me.

by fizzballs: https://i.imgtg.com/2023/05/21/OOPtIc.jpg

Standard Style: Artist: 米山舞 (标准插画风格) | YoneyamaMai artist's standard painting style - v1.0 | Stable Diffusion LoRA | Civitai

--

前言

这下彻底速通了！

仿米山舞的作画风格的LoRA模型。

看到预览图后，部分朋友可能会问：“米山舞的画风不是那种水一样的头发吗？然后很立体的那种感觉呀？”先别着急下这样的结论，至少这个模型不止有这种画风，在下文中，我也会介绍如何利用本LoRA模型生成这种画风，我们慢慢往下看。

Introduction

This is a LoRA model that imitates YoneyamaMai's art style.

After seeing the first preview images, some friends may ask: "Isn't Yoneyama Mai's art style that kind of water-like hair? And a very three-dimensional feeling?" Don't jump to conclusions yet. At least this model is not limited to that style. In the following paragraphs, I will also introduce how to use this LoRA model to generate this style. Let's keep reading.

--

关于这个模型的训练

这个LoRA模型我删了重练数次，前后训练超万步，不得不说，试着练过的朋友肯定都知道，这是个很难练的画风。因为米山舞喜欢用大透视的作画风格，动作也多半不寻常，部分作品的画风偏抽象。一不小心就会过拟合，头发跟水一样、画风极度抽象、四肢各种畸形，哪怕降低权重，也就只有头发跟水一样的画风了，这不是我想达到的效果，我需要的是他的画风与2D模型的有机结合。现在你下载到的这个LoRA是我训练得比较好的一个版本，即使是这个版本，仍然会存在一些问题，但我真的尽力了，它用上了我目前掌握的最先进的训练方法。

Training of this model

I've trained this LoRA model multiple times, with over ten thousand iterations. Anyone who has tried training it knows that it's a difficult art style to master. YoneyamaMai likes to use large perspectives and unusual movements in his art style, and some of his works have an abstract feel to them. It's easy to overfit and end up with hair that looks like water or an overly abstract style, or deformed limbs even if you reduce the weight. This is not the effect I am aiming for. What I need is a combination of his art style and a 2D model. The LoRA model you are downloading now is one of the better versions I've trained, but there are still some issues. Nevertheless, I've done my best, and this model uses the most advanced training methods I learned.

--

关于这个模型的画风

这个LoRA模型的画风在更多的时候表现为强光影、多线条的2D风格，这是从米山舞参与制作的几个MV和部分插画中学习到的，我认为，这个风格作为主要画风更加合适，你也可以想办法为它添加更多细节，但这也需要你有手段控制其线条的流畅性。

（这张图在LoRA的预览图中也有，你可以在那里找到生成参数，下同）

Art style of this model

The art style of this LoRA model is often characterized by strong lighting and multiple lines, with a 2D style. This is learned from several MVs and illustrations that YoneyamaMai has been involved in creating. I think this style is more suitable as the main art style, and you can add more details if you can control its line's fluidity.

(You can find the generated parameters for this image in the preview image of LoRA, as shown below)

当然，如果想要实现各种水一样的头发，可以添加这几个词条用于生成图片，我把他们拆分了出来，而不是混入主要画风：liquid hair, floating_hair, silver_hair, wind, liquid, water

这样，你就可以得到“那种”画风了：

Of course, if you want to achieve various water-like hair styles, you can add these few keywords for generating images: liquid hair, floating_hair, silver_hair, wind, liquid, water. In this way, you can generate that kind of art style.

--

有用的提示

• 如果你需要生成偏肖像画的风格、或是主要聚焦在面部上，请不要使用竖图，而是正方形或者横图。原因很好理解，哪怕是你来画，也很能画出竖着的只有一张脸的图，AI就更加了，如果非要这样做，更多时候会得到上下两张脸的图。此外，这种肖像画的训练素材，更多使用的也是方形图或是横图。（但是不排除运气好的时候能生成非常不错的作品）

• 当线条过于强烈，或者说，线条过于密集的时候该怎么办？这时候可以添加一个或几个词条以改善这种情况，比如 3d, marker_\(medium\), watercolor_\(medium\) 等等，道路千万条，人具有能动性，很多时候换个方式思考能解决不少问题。

• 一些时候你会发现，头发“耷拉”着很难看，这时候也可以尝试添加一个或几个词条改善这种情况，比如 silver_hair, wind, hair_undone, hair_half_undone 等等，这样是不是好多了？此外，还有比较简单的解决方法，那就是使用2.5D的Checkpoint模型。在我看来，画风模型更多的时候提供的是基础，而不是将部分效果直接表现出来。好用的东西往往有其局限性，功能丰富的东西往往不太好上手。

• 如果遇到看不出LoRA画风的情况，请考虑更换Checkpoint模型，尤其是一些自带强烈风格的Checkpoint模型。如果出图相当不稳定，也应该考虑更换Checkpoint模型。

• 我训练的大多数2D LoRA 风格模型都没有经过3D Checkpoint模型的测试。不过如果有强烈的兴趣，也不是不能一试。

• 如果你在使用badhandv4 (Textual Inversion)，并且你生成的图四肢出现畸形，暂时停用它以改善这一情况。一般只有在原图的四肢画的不怎么好的时候才应该考虑使用badhandv4。

• 还有疑问直接在评论区提出就行，不过由于该模型仍然有其局限性，有些问题一时半会儿估计也不太好解决，我尽力而为，能帮上忙就帮帮。

Useful tips

• If you need to generate portrait painting styles or focus mainly on the face, please do not use vertical images but square or horizontal images. Even if you are drawing it yourself, it's challenging to draw a face in only one vertical frame, let alone the AI model. If you insist on doing this, you'll often get a picture of two faces (one on top and one on the bottom). Moreover, most of the portrait paintings training materials use square images or horizontal images. (But there's no guarantee that you won't get good results.)

• What should you do when the lines are too strong or dense? You can add one or several keywords to improve this situation, such as 3d, marker_\(medium\), watercolor_\(medium\), etc. There are many ways to solve problems by thinking differently.

• Sometimes you may find that the hair looks "droopy" and unappealing. In this case, you can also try adding one or several keywords to improve this situation, such as silver_hair, wind, hair_undone, hair_half_undone, etc. Aren't the results much better? Also, there is a relatively simple solution, which is to use a 2.5D Checkpoint model. In my opinion, the art style model provides the foundation rather than directly representing part of the effect. Useful things often have their limitations, and feature-rich things are often difficult to use.

• If you encounter a situation where you can't see the LoRA art style, consider changing the Checkpoint model, especially some Checkpoint models with strong styles. If the output is unstable, you should also consider changing the Checkpoint model.

• Most of the 2D LoRA style models I trained have not been tested with a 3D Checkpoint model. However, if you're interested, it's worth trying.

• If you are using badhandv4 (Textual Inversion), and the limbs in your generated pictures appear distorted, temporarily disable it to improve this situation. Generally, you should only use badhandv4 when the limbs in the original picture are poorly drawn.

• If you have any questions, please feel free to ask in the comments. However, this model still has its limitations, so some problems may not be easily solved. I will do my best to help.

--

模型主要信息

权重建议值：1 (or 0.5~1.4)

基于模型：Anything-5RE

训练信息：筛选出共计841张原画（含裁剪图和视频截图），迭代1072步，多分辨率、智能定向、集散词条、梯度迭代训练（可谓真是办法用尽，手动打tag这辈子是不可能的）。

其他的一些建议：EasyNegative。此外，万一畸形四肢换个种子重画就行。

About Model

Recommended weight value: 1 (or 0.5~1.4)

Based on the model: Anything-5RE

Training information: Selected a total of 841 original paintings (including cropped images and video screenshots), iterated through 1072 steps, and used multi-resolution, intelligent orientation, word scatter, and gradient iteration training (I tried every method I could, manual tagging is impossible).

Other recommendations: EasyNegative. Moreover, in case of distorted limbs, try starting over with a different seed.

--

如果你创作出了好作品，欢迎分享它。如果它无法满足你的需求，我是说，我的朋友，请不要给出一星好评。

If you've created something good, feel free to share it. If it doesn't meet your needs, I mean, my friend, please don't give a one-star 'praise'.