Wan2.2 T2V
Faithfully recreate 2B's in-game appearance.
Weight: High noise: 0.7-1.0
Weight: Low noise: 0.5-1.0
You can start with weight 1.0 with both model.
If you notice some "blury eyes" or " blury face" problem, try lowering the weight of the low noise model.
Sometimes the outfit doesn't generate correctly. Try adding more detailed prompts, like 'black thigh-high boots.'
Workflow included in the demo videos, try drag & drop to your ComfyUI.
Outfit prompt (If the trigger words not working properly):
Skirt: yorha2b, exclusive asymmetrical black dress costume, Juliet sleeves, long sleeves, brown stockings, black thigh boots
Skirtless: yorha2b, exclusive leotard costume, Juliet sleeves, long sleeves, brown stockings, black thigh boots
Add the prompt: " black blindold, coverd eyes" to make her wearing blindfold.
FLUX.1 D V1
Shrink down the size of the file, now it's in dim4.
Fixed the appearance of the blindfold.
Fixed some weird artifacts on the outfit.
Four outfit included: the original, the skirtless, the reincarnation, and the Kaine DLC outfit.
Weight: 0.5-0.8, lower weight will help when you want to apply other outfits.
FLUX.1 D Alpha 0.5
It's my first experimental lora based on FLUX, using the same dataset as the SDXL 3.0. Since I don't have much time to spend in training, the training resolution of this model is 512x512, and the total step is reduced by half.
Dim: 16
Weight: 0.6-0.8
Checkpoint I tested: Only on Flux.1 D
Since the total training step is reduce by half, might need extra description in the prompt to reproduce the outfits accurately.
SDXL 3.0
New Dataset, less overfitting
Added two outfit from the Nier: Automata's DLC and NieR:Reincarnation
Weight: around 0.7
Checkpoint I tested: TurboVisionXL, DreamshaperXL V2.1 turbo, realVisXL V4.0, and JuggernautXL. Some weird artifact will introduce in JuggernautXL and some checkpoint relate to it.
SDXL 2.0
Dataset overhaul, the face and outfit is more consistent.
Adding some pose images, some poses can be achieve without using the ControlNet
Recommended checkpoint: TurboVisionXL, DreamshaperXL, the preview pics also using those.
Since the dataset was captured from the 3D model, adding the photograph style prompt will help to mitigate the "3D-like" feeling.
TurboVisionXL works very good when blending the character into the background.
Outfit prompt (SDXL 2.0):
Skirt:
black dress, black hairband, black gloves, cleavage cutout, juliet sleeves, thigh boots, black thighhighs, turtleneck
no skirt:
leotard, black hairband, black gloves, cleavage cutout, juliet sleeves, thigh boots, thighhighs under boots, black thighhighs, turtleneck
V3.0/SDXL 1.0
Dataset and reg set overhaul
Lowering the weight of the original outfit, now able to change clothes without dealing with Lora block weight
Adetailer and Hires.fx help face generation by a lot, highly recommend using them.
Weight: 0.5-0.75, I'm using around 0.65 in my demo pics
Work with realistic and also anime style checkpoint
Base model: Chilloutmix-Ni
Works with most Asian face checkpoint models, some Western face models also work, but the body ratio and the face are somewhat weird. The weight for my generate pics is 0.75
V2.0
Modified dataset, removing the yellowish and greyish looking
Modified reg set, changing background is much easier than the previous version
Reproduce the original game look of 2B in NieR: Automata, mainly focusing on her face, so no blindfold is offered in this version. The 0.65-0.9 range seems to work the best. The skirt and skirtless variants can be switched by using the "black skirt" and "white leotard" prompts. Highly recommend using it with Adetailer and Hires.fix, it helps face generation by a lot. Since there's only one outfit in my dataset, if you want to change the clothes, you need to use Lora block weight plugin to lower the corresponding weight.
着重还原游戏里2B的形象, 推荐权重0.6-0.8, 可以通过black skirt和white leotard的提示词来切换无裙子有有裙子版本. 最好能配合Adetailer和Hires.fix一起用, 能极大减少脸崩的可能性. 因为训练集就一套衣服, 所以衣服基本是过拟合的, 想换衣服得用Lora block weight去降低相应的权重
Description
Modified dataset, removing the yellowish and greyish looking
Modified reg set, changing background is much easier than the previous version
FAQ
Comments (6)
Wow nice! Would you mind sharing some tips on what a dataset looks like for using in-game photos to make a Lora? I'm trying to make some for some custom characters I have
Sure! I usually turn the resolution and antialiasing all the way up in the game, for some games with really bad antialiasing, you may want to use dldsr to bump the resolution beyond the monitor resolution, then take the character screenshot in close, mid and long range. After that, I would crop and carefully select the face-focus photos and put them in the training set, usually around 30-50 pics, for the other full-body or upper-body photos, I would throw them in the regularization set and then start the training and see the result. If see the reg set element contaminate the train set, then turn down the prior-loss-weight and train it again. At some point, you will get a character Lora that will look very similar to your reference photo. The con of this method is that since we only train the face-focus photos, all other full-body photos are in the reg set, so the outfit will look somewhat different from your reference.
我保留了这层屎黄色滤镜，这样你才知道你玩的是尼尔🤣
2.0和1.0最大的区别也就这层滤镜了, 或许能再出一个分支故意地保留这层滤镜🤣
I love it, it's perfect.
It's like something out of the video game itself.
I wonder if you could please do the same with A2 (YoRHa Type A No. 2)
She is my favorite from Nier
Been very busy these days, only capable of training some models based on my current existing dataset. A2 is on my list, but I'm not sure when I can capture her screenshot and train the models.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.