LoRA for character of Arisa Narumi from Puella Magi Suzune Magica, made for Anima base v1
Model was trained mostly on her costume so other ones might be hard to prompt (not that there was a lot of them in first place).
Since Anima base v1 is a recent model, prompting might take a lot of time, testing on capable GPU like RTX 30/40 series without turbo or other modifier LoRAs yields better results (I'd also recommend waiting for finetunes or previously mentioned Anima Turbo).
_________________________________________________________________________________________________
Base Arisa: narumi arisa, 1girl, medium breasts, pink hair, hair between eyes, pink eyes, two side up, hair ornament
For magical girl outfit: narumi arisa, 1girl, striped thighhighs, elbow gloves, red skirt, striped clothes, zettai ryouiki, lace-up boots, medium breasts, pink gloves, neck ribbon, miniskirt, pink hair, hair between eyes, white belt, pink eyes, pleated skirt, clothing cutout, brown footwear, sleeveless, cross-laced footwear, two side up, frills, red ribbon, cleavage cutout, frilled skirt, knee boots, leotard, red bow, hair ornament, red bowtie, bare shoulders, black thighhighs
For school uniform: narumi arisa, 1girl, pink hair, short sleeves, two side up, pink eyes, medium breasts, pleated skirt, school uniform, purple neckerchief, serafuku, white shirt, purple skirt, sailor collar, white kneehighs, collarbone, hair between eyes, hair ornament
Casual outfit (try other colors): narumi arisa, 1girl, thighhighs, striped, skirt, striped thighhighs, hood, long hair, hoodie, pleated skirt, zettai ryouiki, short sleeves, chain
You can add to each @senri gan for their style since it's recognizable
_________________________________________________________________________________________________
LoRA can have trouble generating her belt buckle.
Character and spin-off series made by GAN
Description
V1 Anima Base upload
Comments (2)
果然在anima服饰细节训练上遇到问题的不止我一个哈哈哈,感觉anima虽然对整体的画面以及脸部特征、服装配色之类的拟合很快,但是对于一些细节的地方(像是花纹、图形、小饰品)出bug的概率比illustrious还高,增加素材数量和训练步数可以降低些许出错概率,好处就是这个模型不容易过拟合
虽然它还挺新的,但我已经喜欢上它在特定情况下展现出的灵活性了。虽然还需要一段时间才能看到改进,但我很期待Turbo版本的推出,这样我就能用它进行画质提升,而且不会出现明显的错误。目前生成任何内容并调整设置对我来说都太耗时了。
我确实找到了一篇介绍训练方法的文章,但整个过程大约需要10到12个小时。虽然这种方法可能有效,但耗时太长,所以我会继续关注相关更新,并及时采用新的训练方法。
抱歉如果有翻译错误,毕竟我是临时翻译的 :D








