Turquoise XL

A game artwork style model like Blue Archive that based on Animagine-XL-3.1

Prompt Guidelines/プロンプト記法

Almost same as Animagine-XL-3.1( https://cagliostrolab.net/posts/animagine-xl-v31-release )

Animagine-XL-3.1( https://cagliostrolab.net/posts/animagine-xl-v31-release )と同じです。

Note/注意

Halos may be generated unintentionally due to the dataset.

データセットの影響で意図せずヘイローが生成される場合があります。

A separate VAE is not required.

別途VAEを使用する必要はありません。

Training information:

Finetuned Anmagine-XL-3.1 by repeating the training and merging a LoRA two times with sd-scripts.

sd-scriptsで二回loraの学習とマージを繰り返すことでファインチューンされました。

Basic training parameters:

Resolution:1024x1024

Bucketing: enabled,min512,max2048

LR scheduler: cosine_with_restarts

Optimizer: Lion

Keep tokens:0

Network dim/alpha: 16/6

Conv dim/alpha:8/2

Gradient checkpointing:true

Grad acc steps:1

Batch size: 2

1st LoRA parameters:

Epochs:6

Unet learning rate:0.0001

TE learning rate:8e-06

LR scheduler cycles:5

Shuffle caption:true

Seed:42

LR warmup steps:250

2nd LoRA parameters:

Epochs:16

Unet learning rate:5e-05

TE learning rate:4e-06

LR scheduler cycles:9

Seed:43

LR warmup steps:250

Dataset information:

Dataset size: 333

Repeats: 2

Training script(source code)

sd-scripts( https://github.com/kohya-ss/sd-scripts )

Notice/おしらせ:

This model is licensed under the Fair AI Public License 1.0-SD( https://freedevproject.org/faipl-1.0-sd/ ).

If you make modify this model, you must share both your changes and the original license.

このモデルはFair AI Public License 1.0-SD( https://freedevproject.org/faipl-1.0-sd/ )でライセンスされています。

このモデルを改変する場合、変更点および元のライセンスを明記する必要があります。