训练与文本编辑条件:
三人次数不同,图片数量不同。
二姐和三妹次数相同,服装三妹变化大。
尽量排除三人趋同,老二描述特征减少。
老大反推详细内容,固定头像,
老二固定风格:汉服,
老三固定头像、半身像、全身像。
本来想做三人LORA,结果参数很好LOSS=0.1,然而是个失败品,所以单生成了一个低学习率的(loss=0.078)大姐,意外好用。
第二版大姐远景还是不尽如人意,相反新生成二姐V2最好版叙述为:erjie,three sister,hanfu,没想到意外成功,基本上可以明确的是图像中央渲染最大的图形就是固定目标,文字只是辅助,太多图和文字只会造成训练的低学习率,给图AI明白你要做什么,训练时背景最好是白色或单色调。
原图是664X1080,推荐生成图大小640X768会有不错效果。
畸形的情况可以调整CFG=4,或等待后续LORA发布,新的模型会添加各种角度图片,以及提高出图稳定性。
I originally intended to do a three person LORA, but the parameter was very good, and the loss=0.1. However, it was a failure product, so I generated a low learning rate (loss=0.78) elder sister(dajie), which unexpectedly worked well. I will upload the LORA for my two younger sisters tomorrow and the day after.
The original image is 664X1080, and it is recommended to generate an image size of 640X768will have a good effect.
The abnormal situation can be adjusted to CFG=4, or waiting for subsequent LORA release. The new model will add various angle images and improve the stability of the image