这个版本我修正了原来上传错文件的问题,并且经过尝试权重给0.6-0.8左右,采样方法用2m++系列效果最好,半身图时不用开After Detailer,开启有可能会把独眼的特征给重绘了,解决这个问题的办法就是把lora拖到重绘提示词框内。背景泛化能力比较差,如果tag有street就一定生成的是夜景。不想要夜景就尽量不要输入street这类tag。
i have fixed the issue of uploading incorrect files in this version,and after trying to give weights of around 0.6-0.8,the best sampling method is to use the 2m++series,you don't need to open the after detailer when taking a bust,opening it may redraw the features of one eye,the solution to this problem is to drag lora into the redraw prompt word box,poor background generalization ability,if the tag has a street, it must generate a night scene,if you don't want a night view, try not to input tags like street。