For better use and control, please read the instruction
(为了更好的使用以及控制,请详细阅读说明书)
Trigger words :
Maid: phRem, blue hair, 1girl, solo, roswaal mansion maid uniform, maid, detached sleeves
ONI: phRem, 1girl, solo, single horn, roswaal mansion maid uniform, weapon, detached sleeves, maid, holding weapon, holding, glowing horns, blood on face, chain, black sleeves, frills, long sleeves, frilled sleeves, oni, oni horns, blood on clothes
Chibi: phRem, Remchibi, chibi, blue hair, 1girl, solo, roswaal mansion maid uniform, maid, detached sleeves

the Chibi painting style was also trained during training, this model can also directly draw Chibi using most base models
And if you want to change the outfit, insert ‘phRem’ only as well the prompt words for the clothing you want to instead, (as far as it exist in your ckpt). And If you think you lost character details while changing clothes, please enter blue hair, blue eyes, short hair, maid, hair ornament, hair over one eye, x hair ornament, hair ribbon, pink ribbon .etc to strengthen the features
Notice & Tips:
Considering that some people may like the original style of anime, I uploaded the original Lora without removing its style. When you feel that the image appears overbaked or want to combine with other Loras, or want the style of the base mold, please use layered control and keep the MID layer (in LoRA Block Weight extension, link in: Ranting8323 / sd-webui-lora-block-weight · GitCode )(weight:1,0,0,0,1,1,1,1,1,1,1,1,0,0,0,0,0)
samples:
MID layer outcome (Meina style face)

without controlling by layer (Anime stlye)

MID layer as well realistic style outcome

advice to use Meinaalter to generate——https://civarchive.com/models/20945/meinaalter
waiting for your comments and ❤ ※ (*>.<*)~
使用超过1400张素材训练,一切努力都是为了waifu
如果爱有颜色~
触发词:
女仆: phRem, blue hair, 1girl, solo, roswaal mansion maid uniform, maid, detached sleeves
鬼形态: phRem, 1girl, solo, single horn, roswaal mansion maid uniform, weapon, detached sleeves, maid, holding weapon, holding, glowing horns, blood on face, chain, black sleeves, frills, long sleeves, frilled sleeves, oni, oni horns, blood on clothes
表情包: phRem, Remchibi, chibi, blue hair, 1girl, solo, roswaal mansion maid uniform, maid, detached sleeves

训练过了chibi画风,因此这个lora可以直接用大多数底模型画蕾姆的chibi人物
如果想要更换着装的话,只用"phRem"以及你的服装提示词就行了(前提是你的底模型里有这种服装)如果你觉得换装丢失了人物特征,有blue hair, blue eyes, short hair, maid, hair ornament, hair over one eye, x hair ornament, hair ribbon, pink ribbon这些其它提示词供选择来强化特征。
注意事项和生成技巧:
考虑到一些人可能喜欢动漫原本的风格,所以我上传了原始lora而没有除去它的画风,当你觉得画面显得overbaked或者想要结合其它lora,又或者想要底模的画风,请使用分层控制并保留中间层,一般来说大家都有这个插件,如果没有请前往 Ranting8323 / sd-webui-lora-block-weight · GitCode 安装(分层权重:1,0,0,0,1,1,1,1,1,1,1,1,0,0,0,0,0)
例图:
仅输出中间层 (Meina底模风格脸)

未分层控制 (动漫风)

保留中间层、真实画风

推荐使用Meinaalter生成——https://civarchive.com/models/20945/meinaalter
喜欢的话给个评论和点赞吧❤ ※ (*>.<*)~
Description
FAQ
Comments (8)
Hi!
Did v2 fix the issues mentioned by the other commenter?
how could I explain for that...ofc, maybe, tho sometimes the style of anime may appear prominent, and the effects of other styles of Lora may not be very obvious, I had introduced how to resolve it in instructions. that's because I want 2 add more options for users but not the skill issue, if u want a version without strongly anime style I can post one for that ԅ(¯㉨¯ԅ). in all, v1 was trained based on 140 pics and v2 was 1400 pics, v2 already has extremely strong robustness, and this time I didn't bind the clothes to the character's trigger words. so theoretically it is very flexible, hope my reply is helpful to you (*^ω^*)
Yeah works really good. Thank you for the detailed explanation
How did you achieve that art style in this lora? I think you trained using NAI, but when I try to train with it, I can never get the art style to stay even though most of the images are in the same art style.
Could it be because of your 1000+ training images and 12k steps, or did you do some magic with block weights? And I'm not talking about the :MID that is used in the preview images.
Glad to have a discussion. Could you please describe the 'art style' in more detail? I understand you're referring the illustration style, right?
Well it's kind of hard to explain, but the head shape is rounder, skin paler and coloring in general reminds me of the anime more, compared to Rem models I train with NAI. Yes I know the dataset affects all of these, but since more than 90% of my images are from the anime or official art, I don't think my dataset is the problem.
I'm basically asking if you used some uncommon settings while training, like using VAE or playing around with Down/Mid/Up Weights, Block dims/alphas or Conv dims/alphas.
I actually didn't do any special processing for the style you mentioned during the training, but I think I can still provide you with some possible reasons. In fact I only trained for about 3,000 steps, while your LORA seems to be overbaked. Yes, in fact, your LORA needs to better reflect the content of the dataset, and is closer to the original appearance of the anime. In our usual impression, Rem's face is considered to be very round, but if you train on anime screenshots like me, you will find that the artists often draw the face longer and sharper when Rem is in the background (long range), and only make the face rounder when she is closer to the frame. When I use your LORA to generate images, I can also achieve the artistic effect you may want for close-up shots. So I suggest you try to filter out the training set with close-up or rounder faces, and appropriately increase the repeat.
Regarding color, we can also find many scenes in anime with dusk, night or indoor environments, where the environmental color makes the original skin tone different from what we normally see, and I have excluded a large portion of this type of dataset.
@phoenix_ms My current rem/ram model is a bad example since it's old (and not even trained on NAI), I was talking about a new version I have been trying to train. I basically only have close-ups since screencaps get blurry really quick.
I feel like I got a somewhat desired effect by basically doubling the steps and doing other weird things, but thanks anyway for answering my questions
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.












