鹿鸣的角色lora.训练只使用了视频和Flux2Klein调整后的图片(也就是把衣服去掉的状态也被训练了),没有使用纯生成的内容."寄明月"的背景似乎有点和角色缝一起了,没什么办法.

如果有可能尽量返点图,我会收集起来继续训练,因为鹿鸣的素材多样性很差,很多东西是过拟合的,得靠纯纯的运气把素材抽出来..

触发词:lumi

white long-sleeved short-sleeved top, light grey denim shorts with a back tie and cutouts on the sides, white hair ornament

chinese fan, chinese dress, chinese gloves, chinese stockings, chinese hair ornament, chinese low heels, jiangnan background

bunny ear, side ponytail, yellow socks, sports shoes, white shorts

purple hair ornament, bunny ear, loafer, white mid-calf socks, purple bow tie

训练用的基础模型是:lizanima_v20(什么情况,我回去找引用,结果那个模型下架了...呃,那看着办吧,Anima模型应该都还好.)

cfg:4.5 steps:24 sampler:er_sde scheduler:FlowMatchEulerDiscreteScheduler