This is my first time training a LoRA; please kindly leave some feedback. このLoRAをトレーニングするのは初めてなので、どうかフィードバックをお寄せください。

Sample grid images are NOT curated/cherrypicked or edited, NSFW contents are inpainted with corrections though. サンプル画像は一切厳選されず、ランダムに選ばれています。過激な内容については手動で編集を加えています。

TRAINING:

I manually tagged 277 images, with roughly 40 unique screenshots of her from the anime, manually cleaned to have text and Adonis removed at 512x512 on NAI; I tried Anylora at first but got much better results with the former. As with any NAI checkpoint, use Booru tags for prompts, which will help NSFW as well. Those 40 images were then split up into smaller sections for focused training, totaling around 220. The remaining 57 images are for NSFW, hands, feet, and various locations. I excluded all scenes where she casts spells or Adonis has flashbacks of her in Episode 04 because of the poor image quality.

I used Prodigy and dampened the learning rate so I could get more steps in and it seemed to have helped considerably, finally settling on v6 at 1939 steps, 1 epoch. 2200 steps was overfitting and 1700 seemed undercooked. In hindsight, I think some details like the gold trim on her cape and bracers could have been better trained but I think I will wait for more images to release before training another version.

INSTRUCTIONS:

For text2image, generate images at 512 is safe, 576 has been getting me the best results though with not much loss in prompt adherence. This LoRA works fine at 1.0 but you may want to lower it slightly for better body planning. 'chloemorgan' is the main trigger word,'icewitchoutfit' for clothing or 'nude' for NSFW but be sure to add additional Booru tags to help out, especially for NSFW. Most facial features should generate consistently, including blond ponytail, blue eyes, and blue ribbon; I did not need to promps for these ever. I did have to retrain a few times for:

• armored boots, bracers, leather belt, gold-trimmed cape, long sidelocks, and midriff, fur trim. If any of these look strange, try adding them. Even without these, you should be able to get pretty close with just 'icewitchoutfit' alone. If her hands become full on gauntlets, try 'bare hands' and negatively prompting 'gloves'

• I highly reccomend adding 'white background/blue background/black background', paired with 'simple background', to your negative prompts, whichever one you tend to get. In the anime, she is often standing against a plain blue background so this can result in some boring scenery. Her official concept art has white and black backgrounds as well.

• Not sure what causes it, but her body proportions can vary between correct and somewhat thin. The most obvious area to look at would be her breast size; if it's too small, the rest of her is probably too thin. You can prompt it out or use a LoRA like I did to adjust it.

• I have generated over 100 images of her with 4 different checkpoints and all of them consistently resulted blond ponytail, blue eyes, and blue ribbon with her sidelocks there 95% of the time, along with most of her outfit being generally the right color. If they do not appear, simply prompt for them. You can prompt 'eyes open, eyes closed, mouth open, mouth closed, smiling, angry, shocked, crying, happy'.

• Her face will vary the most when doing NSFW due to the supplemental images I added to the dataset in order for it to generate lewd images. Hands shouldn't be too terrible hopefully.

手順 (ChatGPTと翻訳する)：

テキストから画像への変換では、512での画像生成は安全ですが、私には最高の結果を得るためには576が効果的でした。このLoRAは1.0で十分に機能しますが、より良いボディプランニングのために少し下げると良いかもしれません。「chloemorgan」は主なトリガーワードで、服装には「icewitchoutfit」、またはNSFWには「nude」を使用しますが、特にNSFWの場合は追加のBooruタグを追加しておくと良いでしょう。ほとんどの顔の特徴は一貫して生成されるはずで、これには金髪のポニーテール、青い目、青いリボンも含まれます。これらの特徴については、これまでにプロンプトを入力する必要はありませんでした。ただし、次の特徴については何度か再トレーニングが必要でした：

• 「armored boots, bracers, leather belt, gold-trimmed cape, long sidelocks, and midriff, fur trim」。これらの特徴が奇妙に見える場合は、それらを追加してみてください。これらを追加しなくても、「icewitchoutfit」だけでかなり近い結果を得ることができるはずです。彼女の手がガントレットでいっぱいになった場合は、「bare hands」を試し、「gloves」に対してネガティブなプロンプトを入力してみてください。

• 私は「white background/blue background/black background」と「simple background」をネガティブなプロンプトに追加することを強くお勧めします。どちらか一方の背景が得られることが多いかを選んでください。アニメでは、彼女はよく単色の青い背景に立っていることが多いので、これでは退屈な景色になる可能性があります。彼女の公式コンセプトアートには白い背景と黒い背景も含まれています。

• 原因はわかりませんが、彼女の体の比率は正しいところとやや細身のところとで変わることがあります。最も明らかな箇所は胸のサイズです。もし小さすぎる場合、彼女の残りの体もおそらく細すぎるでしょう。これを調整するためには、プロンプトを使ったり、私のようにLoRAを使ったりすることができます。

• 彼女に関して、4つの異なるチェックポイントを使用して100枚以上の画像を生成しました。すべての画像で、金髪のポニーテール、青い目、青いリボン、そしてほとんどの場合、横髪が95%の確率で表示され、彼女の服装のほとんどが一般的に正しい色で表示されました。これらの特徴が表示されない場合は、簡単にプロンプトを追加してください。「eyes open, eyes closed, mouth open, mouth closed, smiling, angry, shocked, crying, happy」のようなプロンプトを入力することができます。

• 彼女の顔は、わいせつな画像を生成するためにデータセットに追加した補完画像の影響で、NSFWの際に最も変動する可能性があります。手はあまりひどくならないことを願っています。

Not many nude images of her out there obviously, but I tried to substitute with other blond characters I have painted in the past (largely not AI-generated). About 26 images are NSFW, mostly focusing on three poses: spread legs, from behind, and legs raised. There are additional hand poses for holding a staff and NSFW-related ones such as 'penis grab'. Soles of the feet were added for specific NSFW poses and some 'from behind' poses as well. You will still benefit greatly from adding NSFW LoRAs for specific concepts or poses; I just wanted to make sure mine could do it without too much fuss.

Backgrounds are probably limited as she is mostly against the plain blue sky in most scenes in the anime. Added a few cityscapes from the anime to supplement but I have no idea if it made any difference. Snow, ice or winter motifs tend to generate often, if you negative prompt them out, they sometimes turn into fur, especially if it's near her head.