如果想发挥lora模型的最大性能请根据下面的内容,因为这是我经过大量测试得出的结论,以防获得不佳的效果。
1.基础模型请使用blueboys2D,VAE请使用animevae,正向提示词中应该是male而不是boy(我的训练集中的标签也是male),并且记得把“(rouge:1.5),girl”放入反向提示词中。以上的这些做法都是为了尽量展现饮月的男性特征(很多男性lora模型会偏向中性,这是我不想看到的结果)
2.采样方法要使用Eular a,否则会影响出图的质量,step=40
3.放大算法使用:R-ESRGAN 4x+ Anime6B,35,0.3
4.分辨率的长和宽最好一致(我的训练集为768*768)
5.总触发词为YINYUE,头部的所有特征我加入到了“face,yinyue_horns”中,上半身的所有特征加入到了“clothes,belt,sleeves ”中,下半身的所有特征加入到了“pants,boots”中,请根据你想生成肖像或者全身照合理安排这些提示词。同时,你可以使用“looking to the side,looking at viewer”控制饮月脸的朝向,“from above,from below"控制视角,from front position,from side,from behind”也对应了人物不同视角。
If you want to maximize the performance of the Lora model, please refer to the following content, as this is the conclusion I have drawn after extensive testing to prevent poor results.
1. Please use blueboys2D for the basic model and animevae for VAE. The forward prompt word should be 'male' instead of 'boy' (the label in my training set is also 'male'), and remember to put '(route: 1.5), girl' in the reverse prompt word. The above practices are all aimed at showcasing the male characteristics of the drinking moon as much as possible (many male Lora models tend to be neutral, which is a result I don't want to see)
2. The sampling method should use Eular a, otherwise it will affect the quality of the plot, step=40
3. Use of amplification algorithm: R-ESRGAN 4x+Anime6B, 35, 0.3
4. The length and width of the resolution should be consistent (my training set is 768 * 768)
5. The total trigger word is YINYUE. I have added all the features of the head to "face, yinyuehorns", all the features of the upper body to "clothes, belt, sleeps", and all the features of the lower body to "pans, boots". Please arrange these prompt words according to your desired portrait or full body photo. At the same time, you can use "looking to the side, looking at viewer" to control the orientation of the drinking moon face, "from above, from below" to control the perspective, "from front position, from side, from behind" also corresponds to different perspectives of the character.

Description
FAQ
Comments (6)
您好,我想请问作者大大,崩铁的游戏实机画风也是这个模型的一部分是吗?也就是用2D类的底模也会画出这样偏游戏模型的画面是吗~
是因为我的训练集几乎全部来自于游戏实机的截图,训练的初衷是尽可能想还原人物特征。用2D底模做训练和相关训练参数的目的是探求在不损失特别大的特征前提下去除部分游戏模型的3D感和增加一些细节
@Link_Star_T 嗯嗯。我反而很好奇这种游戏实机画风应该怎么训练,很想有个崩铁画风lora或者原神画风lora
好牛!感谢分享!!!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.









