大家好，好久不见了。这次带来的是画风lora的2.0更新，我在这次的新版本中使用了很多我新学到的技巧，使它相较于1.0和1.1版本有很大的提升。

使用说明：

weight：0.7-0.9,过高的权重容易带来手部崩坏，但是不同模型的效果不尽相同

改动说明：

更换了训练的底模，使模型有更好的泛用性

调整了训练集，从2000张图片中挑选了画风较为明显的120张图片。根据1.1版本的反馈，大家并不喜欢太久远的画风，所以这次没有使用很多古早时期的图片作为训练集。

为训练集增加了tag，我在网上查阅了很多资料，对于画风训练有两种看法。一种是只保留触发词，另一种是为能被定义的全部元素都打上标签，这次我尝试了后者，效果差强人意，还有优化空间（其实是太烂了不想手打tag）。

优化说明：

相比于1.1版本，画风更为明显，即使只使用1girl也能有很明显的效果

修复了1.1版本色彩不够鲜艳，画面色调较暗的问题，而且进行了增强，2.0版本的画面有更强的色彩对比度。

缺点：

手的部分容易崩坏，尝试了很多种方法都没有较好的效果。可以尝试多增加一点负面词降低手部崩坏的概率

这个模型我应该会继续优化到3.0，如果觉得我做的还不错的话请点一下❤，：D

以下是机翻，如果有翻译错误敬请见谅/The following is translated by machine. Please forgive me if there are any translation errors.

Hello everyone, long time no see. This time I bring you the 2.0 update of the Lora style, and I have used many new techniques that I learned in this new version, which significantly improves upon versions 1.0 and 1.1.

Instructions for use:
Weight:0.7-0.9. excessive weight is easy to cause hand collapse, but the effect of different models is different.

Changes:

Replaced the base model used for training to improve its general usability.

Adjusted the training set by selecting 120 images with more distinct artistic styles from a pool of 2000 images. Based on feedback from version 1.1, it was found that people didn't like very old artistic styles, so we avoided using many vintage images for the training set this time.

Added tags to the training set. After researching online, I found two approaches to style transfer: one is to retain only trigger words, while the other is to label all definable elements. This time, I tried the latter approach, and the results were satisfactory but still have room for improvement (to be honest, it was too tedious to manually tag everything).

Enhancements:

Compared to version 1.1, the artistic style is more pronounced, and even using just 1 girl can produce noticeable effects.

Fixed the issue of colors being less vibrant and the overall tone being too dark in version 1.1. Furthermore, enhancements were made to increase color contrast in the visuals of the 2.0 version.

Drawbacks:

The hand parts are prone to breaking. I have tried various methods, but none has yielded satisfactory results. It may help to introduce more negative keywords to reduce the probability of hand parts breaking.

I plan to continue optimizing this model until version 3.0. If you like it, please give me a ❤️. :D

misaki kurehito/深崎暮人style 2.0