此模型原本公开在抱脸上https://huggingface.co/PWB/starlake ，现5.0已完成，一并上传。
模型制作的初心是同时兼顾良好的背景和人物，但由于介于中间，既非画风特化也非风景特化导致地位较为尴尬，这也是一开始未在civitai上传的原因。
而后因为受到欢迎，遂上传之。
迭代5.0的过程也是几经波折...暂且按下不表。
5.0的不同点：
*软化了过拟合，现在不会倾向于自动加入过度的细节。
*tag服从度上升，需要写出对应元素才会出现。
*添加了金字塔噪声，现在可以直接画明暗差大的图。
This model was originally published on Hugging Face https://huggingface.co/PWB/starlake and is now completed in version 5.0 and uploaded together.
The original intention of the model making is to take care of both the background and the characters. The differences in 5.0 are:
Softened overfitting, now it will not tend to automatically add excessive details.
The tag obedience has increased, and corresponding elements need to be written to appear.
Added pyramid noise, now you can directly draw pictures with large differences in brightness.
你可以随意使用该模型，但请不要出售模型本身获利。注意吃相。
You are free to use the model, but please do not sell the model itself for profit.
Description
FAQ
Comments (2)
Perfect results when tesing on background only. It seems that a good building architechure. How you get that?
I have tried many methods, but even now I don’t think the effectiveness of these methods is certain: including tagging background and character pictures according to the proportion of the picture and training them together; fusing models composed of background and character models in the training set; training background while regularizing characters; and mixing a small amount of real photography into the training set. In fact, most of these attempts have resulted in very bad results, and the completion of starlake is only a few of them.