Sign In

DCAI Girl - NoobAI XL (ε‑Pred)

4

25

0

0

Updated: Jul 27, 2025

characteranimecharactororiginal

Verified:

SafeTensor

Type

LoRA

Stats

25

0

Reviews

Published

Jul 26, 2025

Base Model

NoobAI

Training

Steps: 2,500
Epochs: 5

Usage Tips

Strength: 1

Trigger Words

dcai-girl
short hair, orange hair, brown eyes, animal ears
blue dress, skirt, belt, thigh strap, white thighhighs, boots

Hash

AutoV2
5A2F000ABC
default creator card background decoration
DCAI777's Avatar

DCAI777

⬇️日本語は下部にあります。

This LoRA is the final result of a character LoRA created for the DCAI blog article, How to create an original character LoRA [NoobAI XL Training] NoobAI XL Character training

Trigger word: dcai-girl, short hair, orange hair, brown eyes, animal ears

Costume: blue dress, skirt, belt, white thighhighs, boots

About LoRA Training

The training was conducted using noobai‑XL v1.1 eps-pred. At first, 50 base images were created with VRoid Studio, then converted them into two distinct styles using ComfyUI with Animagine XL 4.0, resulting in 100 styled images. These were repeated five times to create a dataset of 500 images in total. Captions were edited using the Dataset Tag Editor in A1111 WebUI.

Training was performed with Kohya_ss GUI, combining the Cosine Annealing learning rate scheduler and the Prodigy optimizer. To further improve quality, training was conducted in bf16 precision with the networks.lora + Convolution configuration. The model was trained with the following parameters: Network Rank (Dimension): 32, Network Alpha: 16, Convolution Rank (Dimension): 32 / Convolution Alpha: 16, Epochs: 5, Total Steps: 2500, and Batch Size: 1.

For more details, please refer to the full blog post.

🔸🔸🔸🔸🔸🔸🔸

このLoRAは、DCAIのブログ記事「オリジナルキャラクターLoRAの作り方【NoobAI XL学習編】NoobAI XLのキャラクター学習」で制作したキャラクターLoRAの最終成果です。

トリガーワード:dcai-girl, short hair, orange hair, brown eyes, animal ears

コスチューム:blue dress, skirt, belt, white thighhighs, boots

LoRA学習について

学習にはnoobai‑XL v1.1 eps-predを使用しています。ベースとなる画像はVRoid Studioで50枚作成し、それらをComfyUIを用いてAnimagine XL 4.0スタイルに2パターン変換。合計100枚のスタイル変換画像を生成し、それを5回リピートして、最終的に500枚の学習用画像を構築しました。キャプションの編集にはA1111 WebUIDataset Tag Editorを使用しています。

トレーニングは Kohya_ss GUI を用い、Cosine AnnealingスケジューラーProdigyオプティマイザー の組み合わせで実行。さらに、bf16での学習とnetworks.lora + Convolutionの設定を加えて品質アップを試みました。Network Rank(Dimension)32 / Network Alpha 16Convolution Rank (Dimension) 32 / Convolution Alpha 16、Epoch数5 / 総ステップ数2500 / バッチサイズ1の設定で学習しています。

詳しくは、ぜひブログ記事をご覧ください。