烟上月の写实摄影_清月流光|Realistic_moonlight|

Introduction in Chinese
这是一个致力于写实风格的sd1.5人像微调大模型，旨在改善sd1.5对于写实人象的不足之处，加入了胶片写真人像进行训练。本模型（以下简称清月流光）尤为擅长刻画真实人像，同时也能生成各种写实类型的背景、角色IP形象、动植物等等一切基于现实存在的事物。泛化性非常好，兼容几乎所有的写实类lora。（炼丹交流答疑Q群：892762841，欢迎大家前来交流心得，我会经常为大家释疑解惑）

清月流光不干扰任何脸型的生成！切换提示词生成的每一张脸几乎都是不同的脸型。这意味着你可以轻松使用Adetailer插件对图片进行指定人脸的更换生成，以及使用任意脸型lora改变出图的脸型特征。（如果你不了解这个插件，可以去B站搜一下ADetailer，学一下使用方法）也可以直接随机抽卡，抽出你的SSR哦~完全不必担心每张脸都是固定的

清月流光是标准的fp16版本，任何sd1.5的大模型删掉无意义的垃圾数据都只有2g左右（对出图或者训练模型，没有任何影响的数据，称为垃圾）有的大模型虽然是5g甚至7g，但多出来的那部分对于出图或者训练是没有任何作用的，删了完全不影响。

喜欢清月流光的话，可以给我的模型点个收藏关注下我哦，这对我是莫大的支持！非常感谢♪(･ω･)ﾉ如果有生成好看的图片，可以给我返个图展示一下你的赛博女友~

关于采样方法，不做要求，我测试过几乎都能用，但我强烈推荐Restart！这个优化器是新出的算法，只需10-15步就能有非常好的出图效果，潜力极大，且兼顾了速度和质量。我的封面也基本都是用的这个采样方法。如果出图脸部崩坏，请使用Adetailer插件进行修复。sd1.5本身底模训练分辨率只有512，崩脸无法避免。说个题外话，很多采样方法每次都会在出图过程中加入新的随机噪声，导致图片无法复现，这是很正常的现象。想要百分百复现原图，需要使用euler a出图

如果条件允许，我强烈建议你使用我主页另一个基于SDXL训练的写实大模型——十指流玉。得益于sd官方增加了数倍的训练集数量，以及更高的图片分辨率，SDXL在写实方面对比传统的sd1.5有指数级的提升。我同样为十指流玉加入了大量高分辨率胶片摄影图集进行训练，无论是光影还是照片质感，都有非常大的提升。对提示词的响应更加“听话”，指哪打哪。

有需要定制模型、商业合作、一对一教学的，可以加我Q：928075483。抖音小红书同名——烟上月，账号会经常发布一些漂亮的赛博cos小姐姐，感兴趣可以去关注下

English introduction

This is a realistic style of sd1.5 portrait fine tuning large model, designed to improve the shortcomings of sd1.5 for realistic human figures, adding film portraits for training. This model (hereinafter referred to as Moonlight) is particularly good at portraying real people, but also can generate a variety of realistic types of background, character IP images, animals and plants, and everything based on reality. The generalization is very good, compatible with almost all realistic class lora. (Alidan Exchange Q group: 892762841, welcome to exchange ideas, I will often for you to solve doubts)

Moonlight does not interfere with the formation of any face shape! Almost every face generated by switching prompt words is a different face shape. This means that you can easily use the Adetailer plugin to change the face of the image, and use any face shape lora to change the face shape of the image. (If you don't understand this plug-in, you can go to Station B to search ADetailer and learn how to use it) You can also draw the card directly at random and pull out your SSR Oh, you don't have to worry about each face is fixed

Moonlight is the standard fp16 version, any sd1.5 large model delete meaningless junk data is only about 2g (on the map or training model, no impact on the data, called garbage) Some large models although 5g or even 7g, but the extra part of the map or training is not any effect, deleted completely does not affect.

If you like the moon, you can give my model a collection to pay attention to me, which is a great support for me! Thank you so much. ♪(· ω ·)ノ if you generate a nice picture, please send me a picture to show your cyber girlfriend ~

As for the sampling method, there are no requirements, I have tested almost all work, but I highly recommend Restart! This optimizer is a new algorithm, only 10-15 steps can have very good results, great potential, and both speed and quality. I use this sampling method for most of my covers. If your face is broken, use the Adetailer plugin to fix it. sd1.5 itself has a bottom mode training resolution of only 512, and face breakage is unavoidable. As a side note, many sampling methods will add new random noise to the drawing process each time, resulting in the image will not be reproduced, which is a normal phenomenon. To reproduce the original drawing 100%, use euler a to draw

If you can, I highly recommend you to use another realistic large model based on SDXL training on my home page - Ten Fingers Liuyu. Thanks to the sd official increased the number of training sets several times, as well as higher picture resolution, SDXL in terms of realism compared to the traditional sd1.5 has an exponential improvement. I also added a large number of high-resolution film photography collections for training, both light and shadow and photo texture, have a very big improvement. More "obedient" in response to prompt words, pointing where they go.