⬇️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, “<a target="_blank" rel="ugc" href="https://www.digitalcreativeai.net/en/post/original-character-lora-noobai-character-training">How to create an original character LoRA [NoobAI XL Training] NoobAI XL Character training</a>”Trigger word: dcai-girl, short hair, orange hair, brown eyes, animal earsCostume: blue dress, skirt, belt, white thighhighs, boots<h2 id="about-lora-training">About LoRA Training</h2>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「<a target="_blank" rel="ugc" href="https://www.digitalcreativeai.net/ja/post/original-character-lora-noobai-character-training">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</a>」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<h2 id="lora">LoRA学習について</h2>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 WebUIのDataset 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スケジューラーと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 16、Convolution 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。

Files with SHA256: 5a2f000abc35a3d7c8db6c953132fd5a9a64222e8a9a5b033e9f84c048ba46cf