[MILLION LIVE!/ミリオンライブ/ミリシタ] Iku Nakatani/中谷育 LoRA

LoRA model of Iku Nakatani (中谷育) from media-mix project THE IDOLM@STER MILLION LIVE! (アイドルマスターミリオンライブ！).

Base model for training is animefull-latest, and the trigger word is iku nakatani (million live).

The LoRA weight of samples is 0.8; you can also try other weights. The greater the weight, the closer the character style is to the game card of THE IDOLM@STER MILLION LIVE! THEATER DAYS (ミリシタ) and the new tachie drawn after the release of the game.

Note: I am not using the official LoRA loader, so you should use the additional network WebUI plugin to use my LoRA setting.

I’ll appreciate it if you’re willing to use this LoRA model and share AI arts to other people. I’m looking forward to your wonderful work.

————————————

跨媒体企划《偶像大师 百万现场》的角色中谷育的LoRA模型

训练时使用的基底模型为animefull-latest；模型的触发词为 iku nakatani (million live)

示例图的LoRA权重为0.8，您也可以尝试其他权重值。权重越大，角色风格越接近游戏《偶像大师 百万现场 剧场时光》（MLTD）游戏卡面以及该游戏发行后新绘制的立绘

示例图是通过WebUI的可选附加网络（additional network）插件加载本LoRA模型生成的，所以prompt里并没有调用LoRA的tag，还请注意

如果您愿意使用本LoRA模型并将作品分享给其他人，本人将不胜感激。期待您的佳作。

————————————

Some confusions about the training:

In the very beginning, I used the training parameters just the same as the Momoko model, Mirai model and Yuriko model:

Base model: animefull-latest
Resolution: 576x768
Optimizer: Lion
Learning rate of unet: 1/3 of optimal learning rate calculated by DAdaptation optimizer
Learning rate of text encoder: 1/10 of learning rate of unet
Gradient accumulation steps: 64
Scheduler: cosine with restarts
Dim/Alpha: 128/64
Batch size: 2

And I set different repeat times for different materials like the following:

2D tachie: 8 times
Icon from 2D tachie: 8 times
3D tachie: 5 times

Also, I applied the multi-resolution noise with the following parameters:

min_snr_gamma: 5
multires_noise_iterations: 6
multires_noise_discount: 0.3

The result showed that AI couldn’t remember Iku’s single hair bun well. I tested it by generating 50 illustrations, and found that in Epoch 20 less than 50% of the illustrations contained the single hair bun, and in Epoch 40 the percentage rose up to 72%. It seems that the training needs going on, but Epoch 40 or even more is a bit terrible for this small-dataset training because it would be easy to overfit. I rechecked the dataset, but didn’t found anything wrong. I tried to increase the learning rate, the repeat times, and even removed the multi-resolution noise (because this may slow down the convergence of model), but the consequence basically remained the same, though the situation indeed improved.

Then, I followed the training method of LoRA Lazy DAdaptation Guide, other than changing the Dim/Alpha from 32/16 to 128/64. The result is ACCEPTABLE overall: AI understands the single hair bun and it appears in illustrations stably; there is no strange patches of color or distortions; the clothes are easy to change; the pose seems OK. However, I observes that the overfitting of style is more obvious than the previous traning method.

In my opinion, It may need more steps for the Lion optimizer to learn more features, and ＞40 Epochs in this case still seems to be safe? Efforts such as increasing learning rate or repeat times seem useful in some degree, indicating that the learning rate in the beginning should be little higer.

Of course, this is just my personal view; all in all, the model is still reliable to use, but l’m not satisfied with its performance. I want to make sense why the previous training method cannot achieve a good result in revealing Iku’s single hair bun. If anyone knows the reason, please tell me and I’ll appreciate it.