这是一次中国风题材的尝试，目前用于国画等传统国风的lora已经有不少小伙伴在研习了，但叙事性较强的CG类国风好像还没有看到，出于自己项目的需求，我炼制了这一国风CG大师丹。训练素材多为国内一流的CG大拿作品，出于目前cg界目前对aigc界的敌视，鄙人就不一一列出艺术家了。其作品多为东方玄幻、神话等题材，多为场景类，人物也有但不多。

该模型适合与revAnimated、chilloutmix等表现力比较成熟的模型配合使用可以呈现较为不错的人物为主的叙事效果，场景类与Anything_v5、final-pruned等综合性更强的底模配合。

This is an attempt at Chinese style themes. Currently, there are many friends studying Lora, which is used for traditional Chinese painting and other traditional Chinese styles. However, it seems that I have not yet seen CG style with strong narrative power. For the needs of my own project, I have refined this Chinese style CG master pill. The training materials are mostly top-notch CG masterpieces in China. Due to the current hostility of the CG community towards the AIGC community, I will not list the artists one by one. His works mostly focus on themes such as Eastern fantasy and mythology, with scenes and characters, but not many.

This model is suitable for use with mature expressive models such as revAnimated and chilloutmix, which can present a good narrative effect with characters as the main focus. The scene class is similar to Anything_ V5, final run, and other more comprehensive bottom mold combinations.

我喜欢把它叫做妖丹。

因为个人喜欢中国神话题材的创作，所以对佛魔妖兽类常有遐思不断。目前midhourney上同质化严重，且对一些个性化审美表达上有欠缺。比如最有想象力和表现力的中国神话题材，支持度极差，私炼模型是迫不得已。此模型尚有极大欠缺，比如人脸手脚等细节依然有较大概率会崩掉，一个是因为本身这个模型训练是针对东方式叙事场景的强化，所以人物是次要的，其次在场景中出现人物也往往不是主要叙事元素，导致占比太小没有足够像素来表现。众所周知，如此就难免对人物细节塑造欠缺了。

所以通过调整镜头为特写或者中景就能极大概率改善此问题。如果是叙事也可以通过图生图模式中，controlnet的tile+Ultimate SD upscale放大的同时重绘细节，也能基本解决该弱点。

其优点也算突出，即表现中国风题材上终于有了叙事能力较强同时又能兼顾东方审美特有的想象力，找到叙事与叙情的一种平衡。目前来说，它确实能解决我不少工作上的问题，至少它是很好的中间优化环节的一种思路。

以上仅是一点个人经验补充，欢迎诸位反馈更多问题讨论更多解决思路。

I like to call it the Demon Pill.

Because I personally enjoy the creation of Chinese mythological themes, I often have endless thoughts about Buddha, demons, and beasts. At present, there is a serious homogenization on midhour, and there is a lack of personalized aesthetic expression. For example, the most imaginative and expressive Chinese mythological themes have extremely poor support, and it is imperative to refine models privately. This model still has significant shortcomings, such as the high probability of details such as facial features, hands, and feet collapsing. One reason is that the training of this model is aimed at strengthening Eastern narrative scenes, so characters are secondary. Secondly, when characters appear in the scene, they are often not the main narrative elements, resulting in a small proportion and insufficient pixels to represent them. As is well known, this inevitably leads to a lack of detail in shaping the characters.

So adjusting the lens to a close-up or mid shot can greatly improve this problem. If it is narrative, the weakness can also be basically solved by redrawing the details while zooming in with controlnet tile+Ultimate SD upscale in the graphical model.

Its advantages are also quite prominent, that is, it has finally developed a strong narrative ability in depicting Chinese style themes while also taking into account the unique imagination of Eastern aesthetics, finding a balance between narrative and emotional narration. Currently, it can indeed solve many of my work problems, at least it is a good way to optimize the intermediate process.

The above is just a personal experience supplement. We welcome everyone to provide feedback on more issues and discuss more solutions.

请注意：

1. 本模型仅用于交流学习，请勿拿来商用。

2. 我常用的CFG范围是5~7.5，Lora权重通常在0.6~0.9。

3. Clip skip设为1或2时，生成的效果会有所不同，大家可以多尝试看自己喜欢哪种风格。

4. 本Lora模型的训练集主要为国内cg大佬作品，生成欧美人像和建筑等效果可能会受影响。

5. Please Note:

   This model is only for communication and learning purposes, please do not use it for commercial purposes.

   My commonly used CFG range is 5-7.5, with Lora weights typically ranging from 0.6 to 0.9.

   When the Clip skip is set to 1 or 2, the generated effect will vary. You can try more to see which style you like.

   The training set of this Lora model is mainly based on the works of domestic CG masters, and the generation of European and American portraits and buildings may be affected.