我的Pixiv个人主页 / My Pixiv personal homepage: https://www.pixiv.net/users/17321274
这是一个来自游戏《永劫无间》的角色——顾清寒-霜凝雪。本Lora训练图集不包含任何真人图片。底模使用国风V3.3。参考Lora权重0.75。和国风3.3兼容性较好的模型有：ReV Animated, DreamShaper, GCM-Game Concept Map等，你可以尝试使用这些模型进行AI绘图。
This is a character from a China mobile game called Naraka: Bladepoint——Gu Qinghan-Frost snow. This Lora training atlas does not contain any images of real people. The basic mold uses GuoFengV3.3.Reference Lora weight 0.75.Models with good compatibility with Gufeng 3.3 include: ReV Animated, DreamShaper, GCM-Game Concept Map, etc. You can try to use these models for AI drawing.
Lora召唤词 / trigger words：YJWJguqinghanSNX
参考正面tag如下——Reference Prompt：
YJWJguqinghanSNX, 1girl, solo, white hair, long hair, blue eyes, breasts, jewelry, lips, makeup, see-through, dress, science fiction, hair ornament, gloves, long sleeves, thigh boots, cityscape, night, looking at viewer, mature female, hand on hip,
参考负面tag如下——Reference Negative prompt：
(worst quality, low quality:1.4), monochrome, zombie,undefinedEasyNegative, lowres, bad anatomy, bad hands, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,too many fingers, thick lips, oversized head,facing away,extra arms,extra ponytail,bad eyes,fused eyes,distorted eyes,multicolored eyes,small breasts,loli,distorted background, overexposure,
为了更好地展现人物服装特征，我的模型展示图往往使用半身照。事实上，对于绝大多数角色，我都是使用全身素材进行训练。如果你喜欢全身照，请删去cowboy shot后添加full body并尝试添加这些词汇以展示人物的鞋子:high heels 或 boots 或 sandal 或 shoes等。
In order to better show the character's clothing characteristics, my model display images often use cowboy shot. In fact, for the vast majority of roles, I train using full-body footage. If you prefer full-body photos, try adding these words after adding full body to show the person's shoes :high heels or boots or sandal or shoes.
版权说明：我的Lora训练均使用网络公开图片进行训练，训练后的Lora成品不推荐用于商业用途。如果你用于商业用途，一切风险由你自己承担，我不承担任何责任！
Copyright Note: My Lora training is trained using public images on the Internet, and the Lora is not recommended for commercial use. If you use for commercial purposes, all the risk is borne by you, I do not assume any responsibility!
如果你有喜欢的人物角色，希望将他们做成Lora，你可以联系我帮助你训练Lora模型，不收取任何费用。
If you have a favorite character that you would like to make Lora, you can contact me to help you train Lora models at no charge.
我的个人AI交流群：828795031
入群密码：S82KFCV50
群内不搞黄，不收费，不教萌新（萌新请自学），欢迎进来交流技术。
Description
2023/09/18
FAQ
Comments (13)
<lora:永劫无间顾清寒-霜凝雪4005:1:1:lbw=1,0,0,0,0,0,0,0,1,1,1,1,1,0,0,0,0>
Pretty good results with these weights. However it's a bit hard to get good face still, as lora has big effect on style.
days later i will try some real people checkpoint and see what happen
Nice, let me know how it goes. Maybe it's worth trying on AI model too. Overall outfit works quite well with your LORAs but face is hard to get right due to large effect on style.
<lora:永劫无间顾清寒-霜凝雪4005:1:0.8:lbw=1.5,1,1,0,0,0,0,0,1,1,1,1,1,0,0,0,0>
I improved the weights a bit to get overall better quality. 1.2 LORA weight with 0.8 on hi-res fix makes outfit really accurate but requires some inpainting for face.
I also noticed that the outfit is a bit too modular in the training data. In particular "long sleeves" causes her to have huge, loosely hanging sleeves and "see-through" accidentally makes some parts of outfit transparent. I was able to deal with sleeves by removing the tag and putting "bare shoulders" in negative. To be honest this particular outfit is basically bodysuit. Should be very tightly form-fitting and have shiny look (not cloth).
For reference - usually face should look like this in the model I'm using (used nurse LORA + only one layer from Ji Yingying LORA to get only hairstyle).
@Malekith My training was conducted according to the pictures you provided. The faces of the people in your pictures have very obvious eye makeup, so it is also reflected in the model. For most game or anime characters, facial features are an important distinguishing factor, so I didn't remove them. As for the coat you mentioned, I'm not sure if ai can understand this kind of transparent coat, and I don't know how to fill in the corresponding tag words, perhaps the word "transparent raincoat" would be more appropriate. Yesterday I tested the effect on this game character with another live action model and it was not significantly different from what I have used so far. I was using https://civitai.com/models/148173/realisticmoonlightor
@Malekith The feature of guofeng 3.3 I am currently using is that the face is relatively fixed. The advantage is that it will hardly break the face, and the disadvantage is that it is difficult to fit some facial details with special features.I trained a character for Honor of Kings who had a strong facial feature, but it didn't show up on this model.
@blacksnow Mind uploading the other version? I would like to try it too, maybe I can try different weights and improve results a bit.
The main problem with face was lighting and facial features style (and it's similar on your other LORAs). Makeup etc. all worked well.
Regarding outfit what I mean is that when you train LORA without outfit tags, the outfit will be absorbed into LORA activation tag. So for example with Ji Yingying black and white outfit I could get entire outfit without tags for clothes etc. However when you split outfit into outfit tags then SD will mix up understanding of tags a bit - for example when using new skin LORA for Ji Yingying - using shorts tag sometimes makes normal denim shorts, black pantyhose gives her normal pantyhose on both legs (it's supposed to be only on one leg) and so on.
For example here https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/54c5bff6-2f07-4c55-a96f-5a65cfbcd1fa/width=1920/justina3x.jpeg
we see that outfit was fairly accurate, but the outfit material was incorrect - it looks like normal, loose cloth instead of skin-tight bodysuit. And on the side of her dress there is "see-through" part, that is not supposed to be there - only her laboratory coat should be see-through.
So overall it would be maybe better to only add tags for hair, makeup etc, and keep all outfit related tags absorbed into LORA activation tag. Or maybe reduce problematic tags like "long sleeves", "dress", "see-through". Alternatively maybe outfit tags should be changed - probably something like white latex, latex dress, skin tight, shiny clothes.
Thanks for responses.
@blacksnow Managed to mostly resolve issues with face by using a trick with secondary LORA with reverse weights (valued flipped from 0 to 1 and other way around on every layer). I uploaded some pics with good results.
@Malekith I will try different way to fix this problem, thank you
@blacksnow Hi again, would you mind sharing your Kohya_ss config file for LORA training? I am trying to run some training myself, so far first try was fairly decent, but I'm still working on face - I'm training again at higher resolution (before was 512 only). I know the training parameters can be looked up from LORA files, but I am not very experienced with training, so it would be helpful to have clean config file.
https://cdn.discordapp.com/attachments/1168270869601198230/1168537511912546405/grid-0011.jpg?ex=6552205d&is=653fab5d&hm=b6bd6a56b70b870a57089f7bf2b9d9e90eb377eacbdf61289958e8f7ddbc64bd&
@Malekith I will send you tomorrow.I have set up an additional training set in training, using nude figure material to enhance postural generalization, as well as to complement nude figure details. It might not be in line with how you train.
@blacksnow Thanks, greatly appreciated. If you could share your additional photoset for poses etc. it would be awesome too. Once I train something decent I'll share on Civitai.
I recently made Naraka SD discord server with friend from EU community. Not sure if you use Discord, but you can take a look if you want to.
https://discord.gg/TTtK8E2w
@Malekith ok im in
@Malekith my discord id: blacksnow2100