Arknights Lemuen
Trained on NAI 7GB model with two official arts from HORTUS DE ESCAPISMO event. The model itself can reproduce Lemuen's upper body detailed quite well, including the halo and wings. For lower body, the results are relatively inconsistent and I can't make it better.
I strongly recommend using LoRA Block Weight extension with this model. I personally use this preset: 1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1 with 0.8~0.9 weight.
If you don't feel like using the extension, you can try weight 0.6~0.8. Be aware that low weight sometimes can't reproduce her wings very well. If you find her wings and halo missing, try add halo or halo above head, energy wings to your prompts.
Like I mentioned above, this model's trainset is very limited. Try tweak the checkpoint model, weight, vae or use with controlnet to make your work better. I personally find anylora works better with it, you can find it on this site.
明日方舟 蕾缪安
训练底模为NovelAI的 7GB模型,素材为两张空想花庭的活动资源,一张立绘和一张站立的CG。一般来说,这个lora还原安姐上半身细节还不错,尤其是面部,基本上拉高点权重就能拟合到立绘上了。不过下半身效果就不是很稳定,因为本身立绘就是坐轮椅,我尽力抠图又手画补全了一部分,又用换头加上了薇尔丽芙的下半身作为素材,结果还是不太行,也没办法了。
个人强烈推荐配合用 LoRA Block Weight extension 这个插件,搭配我上面英文部分的那个预设,使用0.8到0.9的权重效果比较好,当然你要是想换装或者色色,那就拉低点。
如果你就是不想用这个插件,那建议用0.6到0.8的权重。但是比较低的权重会导致翅膀和光环还原不够好,建议加上 halo 或者 halo above head, energy wings 这种提示词补全特征。
最后,如上文所言,模型本身训练集极其有限,建议多尝试调整底模,VAE还有controlnet。我个人比较推荐用 anylora 这个模型,站内一搜就有。注意这个模型本身没什么特殊画风,所以出来的图画风会跟训练集,也就是跟立绘非常相似,如果你想不喜欢这个画风说明你这个人只能度过相对失败的一生可以换个底模并拉低权重,还有调整lora block weight的预设。