::: 该页面为中国传统纹样loha的使用介绍，关于详细的训练过程，分层调试方法，可以到我们的wiki上去查看。https://aigc.ioclab.com/sd-showcase/chinese-ornament.html :::

中国传统纹样lora by IoC Lab

训练及数据：@shichen @ciaochaos

版本 介绍 2023.4.1 上传lora，介绍内容（本篇） 2023.4.4 更新训练方法（wiki待上传）

在图片中你可以看到有大约25个以上的版本，我挑选了其中最有特点的几个版本，你可以在version中挑选下载，或者前往huggingface下载全部的版本

https://huggingface.co/ioclab/ioc-lab-models/tree/main

使用方法：

在训练集的标注中，分为以下五个语段：

traditional chinese pattern

with circular symmetrical composition
with square symmetrical composition
with special-shaped symmetrical composition

red and blue color scheme

featuring horses and clouds
featuring flowers and branches

a flower in the center

所以你可以围绕着这五种固定句式进行扩充，如需要作为背景，可以调低权重，增加 1 girl

分层拆解：

在分层图片中可见，在高权重下，通过仅开启 midd，outd 等 lora 层的权重，可以做到相比于直接降低权重，更能保留纹样风格的人物背景图。

推荐模型版本介绍：

这里是一些效果比较特色的模型版本，但出图效果与参数，种子等都有关系，你可以尽情尝试，期待你的评论

特色模型版本

以下是一些特色模型版本：

• 4：特点是纹样更卡通化，色彩更缓和

• 7：特点是对关键词的还原更好，更符合数据集的倾向

• 10：特点是更对称，同时构图倾向会偏向于非中心对称构图

• 13：特点是相比10的颜色会更缓和，高权重下会出现第三种散布构图

• 15：特点是高饱和度和高横向对称构图

• 16：特点是高饱和度+散布构图

人物配合版本

以下是容易和人物配合的版本：

• loha01_old3: 更容易出卡通光影，构图和花纹装饰性强

• loha04_old3: 构图容易出全身，光影保留效果好

• loha06_old1: 对构图限制小，人物动作自然

• loha02_old2: 背景层次更多，对称性强

数据集标注及loha训练方法，见wiki链接

https://aigc.ioclab.com/sd-showcase/chinese-ornament.html

Chinese Ornament Wiki in English

::: This page is an introduction to the use of traditional Chinese pattern Lora, for detailed training processes and layer debugging methods, please refer to our wiki. https://aigc.ioclab.com/sd-showcase/chinese-ornament.html :::

Chinese traditional pattern Lora by IoC Lab

Training and Data: @shichen @ciaochaos

Version Description 2023.4.1 Uploaded Lora, Introduction Content (This Article) 2023.4.4 Updated Training Method (Wiki To Be Uploaded)

In the pictures, you can see more than 25 versions. I selected the most distinctive versions for you to download, or you can go to Huggingface to download all versions.

ioclab/IocLab_Models at main (huggingface.co)

How to use:

In the annotation of the training set, the following five language segments are divided:

traditional chinese pattern

with circular symmetrical composition
with square symmetrical composition
with special-shaped symmetrical composition

red and blue color scheme

featuring horses and clouds
featuring flowers and branches

a flower in the center

So you can expand around these five fixed sentence patterns. If you need to use it as a background, you can reduce the weight and add one girl.

Layered Decomposition:

As shown in the layered images, under high weight, by only turning on the weights of midd, outd, and other Lora layers, you can achieve a person-background image with a style more retained than directly reducing the weight.

Recommended Model Version Introduction:

Here are some model versions with distinctive effects, but the output effect is related to parameters, seeds, etc. You can try them out and we look forward to your comments.

Distinctive Model Versions

Here are some distinctive model versions:

• 4: The characteristic is that the pattern is more cartoonish and the color is more gentle.

• 7: The characteristic is that it better restores the keywords and is more in line with the tendency of the dataset.

• 10: The characteristic is that it is more symmetrical, and the composition tends to be non-centrally symmetrical.

• 13: The characteristic is that compared to 10, the color will be more gentle, and the third scattered composition will appear under high weight.

• 15: The characteristic is high saturation and high horizontal symmetry composition.

• 16: The characteristic is high saturation + scattered composition.

Character Matching Versions

The following are versions that are easy to match with characters:

• loha01_old3: More likely to produce cartoon light and shadow, with strong composition and decorative patterns.

• loha04_old3: The composition is easy to produce the whole body, and the effect of retaining light and shadow is good.

• loha06_old1: The restriction on composition is small, and the character's movements are natural.

• loha02_old2: More layers in the background, strong symmetry.

For dataset annotation and Lora training methods, please refer to the wiki link.

https://aigc.ioclab.com/sd-showcase/chinese-ornament.html