Crystalia:Kizuna kirameku koi iroha suzakuin tsubaki 絆きらめく恋いろは 朱雀院椿

This is a LORA to reproduce "Suzakuin Tsubaki", one of the characters in "Kizuna Kirameku Koi Iroha".I created this character because no one had created this character with the Pony model.Please see the sample for the level of reproduction. The face, eyes, hair, etc. are well reproduced, but the clothes are unstable.My mistake is that I did not turn off the left-right reversal learning during training, so the ponytail sometimes appears in the opposite direction, but hair ornaments and ribbons usually appear in the same direction as the ponytail, so reversing the image left-right will often solve the problem.Instructions: Use the following tags(As for bikini, it is spelled differently, but I made a mistake and wrote it like this to make it learn. If you write this, you can still output. It might be the same without it.)Character herself: crystaltsubaki, yellow eyes, white hair, very long hair, bangs, side ponytail, hair between eyes, medium breastsuniform: blue hair ribbon, hair flower, hair ornament, white cape, red aiguillette, collared shirt, short sleeves, high-waist skirt, frills, red necktie golden buttons, black pantyhose, white footwearcivilian clothes: red hair ribbon, black hairclip, white blouse, short sleeves, neck ribbon, corset, red skirt, pumps, pumpsswimwear: black hair ornament, blue hair ribbon, golden hairclip, bead bracelet, navel, shiny skin, black frilled bilinimaid's uniform: blue bows, neck bow, waist apron, detached collar, detached sleeves, bare shoulders, white gloves, maid headdress, side-tie panties, blue bow, blue hair between eyes, collarbone, thighhighsbunny: playboy bunny, black hair ribbon, rabbit ears, rabbit tail, black hairband, detached collar, wrist cuffs, red bow, bare shouldersnude: completely nude, shiny skin, (Shiny skin is my preference. If you don't want it, please remove it.)fox ears: fox ears, animal ear fluff, (if you don't want fox ears, please remove them)Bunny and Maid are not reproducible due to the small number of study materials. I think I have enough study material for the uniforms, but the cape area tends to fall apart, probably because of the way the tags are attached.================================================================Here are the other words we learned (i.e. these words will affect the generated results)morning, (which will have a strong light intensity, like in a daily scene in a game)evening, (will have a sunset color like in the adult scenes in the game)night, (I used this tag for the night scene in the game)shade, (This tag is used for CG with shadows on the face. If you don't want shadows, put this tag in the negative.)bokeh, (glowing particles that randomly appear on the screen like CG in a game)text, (Some of the images in the study included text, so we learned them as Text. Please put this tag in the negative.)unworn clothes, (Some of the training images had clothes taken off and scattered on the floor, so we tagged them as such. (In this case, the scattered items, such as skirts, were also tagged with their own tags.)bed, bed sheets, pillow,tatami,katana, (these were also included in the background, so they were also tagged in the study)orgasm,(I have learned this tag for the face when climaxing in game)embarrassed, (I have learned this tag for the embarrassed face in the game)Blush (reddish cheeks when slightly embarrassed)nose blush (red cheeks up to the nose; I think this is what most of the NSFW study material was about)dark-skinned male, solo focus, (We use this tag for images in which the main character (male) is depicted in the game. (In other words, by including these tags, it is easy to generate images of men whose faces do not appear in the game.)mosaic censoring, (CG of the game had mosaic. This tag is given. If you don't need mosaic, please put this tag in the negative)================================================================ConfigurationDownload the sample, read the PNG information and transfer it to Txt2Img, etc. The recommended settings for Step and CFG are those of the base model used to generate, not of this LORA.Euler aClip skip2================================================================Additional NotesI used the data of the standing figure in the game for the training image that shows the whole body, but it was necessary to make it more close-up, or the character tends to collapse when trying to draw the whole body at the time of generation. (full body,)Nevertheless, as shown in the sample, there are cases where the character does not collapse even when almost the entire body is shown, so perhaps it is better to avoid using this tag.It is easy to get similar compositions and backgrounds for certain body positions in NSFW, perhaps due to the bias of the study material.(Conversely, you can output a picture that is quite close to the CG in the game)================================================================If I have extra BUZZ I will eventually make some modifications (e.g. left-right flipping, etc.)I haven't learned the combat outfits, and I haven't learned the weapons. This is because I thought that too much learning would interfere with character reproduction.This LORA was created using CIVITAI's training service. These outfits were created using a single dataset, but I understand that there are procedures for creating LORAs with multiple dataset structures? Please let me know if you are familiar with this.I do not know how to learn about characters with special eyes.I have created several LORAs, but I have not been able to get them to work, and I decided to release this one because, although it has some problems, it has met my goals to some extent. I would appreciate any feedback you can give me so that I can improve it.　AutismMix LightningStep8 was used to generate the sample images.　===============================================================絆きらめく恋いろは　に登場するキャラクターの一人「朱雀院椿」を再現するLORAです。Pony系でこのキャラを作っている人がいなかったので作成しました。再現度はサンプルをご覧になってください。顔や目、髪型などの再現性は高いですが服は不安定といった感じでしょう。私のミスなのですがトレーニング時に左右反転学習をオフにしなかったため、ポニーテールが反対に出現することがありますが、髪飾りやリボンもたいていポニーテールと同じ方向にでるので、画像を左右反転すれば解決することが多いです。　使用方法：以下のタグを利用してください。(ビキニに関しては綴りが違うのですが、ミスでこうかいて学習させてしまいました。これを書いても出力できます。なくても同じかもしれないです）キャラクター本人：crystaltsubaki, yellow eyes, white hair, very long hair, bangs, side ponytail, hair between eyes, medium breasts,制服：blue hair ribbon, hair flower, hair ornament, white cape, red aiguillette, collared shirt, short sleeves, high-waist skirt, frills, red necktie, golden buttons, black pantyhose, white footwear,私服：red hair ribbon, black hairclip, white blouse, short sleeves, neck ribbon, corset, red skirt, pumps,水着：black hair ornament, blue hair ribbon, golden hairclip, bead bracelet, navel, shiny skin, black frilled bikini,メイド服：blue bows, neck bow, waist apron, detached collar, detached sleeves, bare shoulders, white gloves, maid headdress, side-tie panties, blue bow, hair between eyes, collarbone,　thighhighs,バニー：playboy bunny, black hair ribbon, rabbit ears, rabbit tail, black hairband, detached collar, wrist cuffs, red bow, bare shoulders,ヌード：completely nude, shiny skin, (Shiny skinは好みです。いらなければ消してください）狐耳：fox ears, animal ear fluff,バニーとメイドは学習素材の枚数が少ないため再現性がやや低いです。制服については学習枚数が足りているはずなのですが、タグの付け方がわるかったのかケープの周辺が崩れやすいです。================================================================その他の学習をしたワードを紹介します（つまりこれらのワードは生成結果に影響を与えるでしょう）morning,（ゲーム内の日常シーンのような光の強い明るさになります）evening,（ゲーム内のアダルトシーンのような夕焼け色になります）night, (ゲーム内の夜のシーンにこのタグを使用しました）shade,（顔に影がかかっているCGなどにこのタグをつけています。つまり影が嫌ならこのタグをネガにいれる）bokeh,（ゲーム内のCGみたいに画面にランダムに出現している光る粒子）text,（学習画像に文字が含まれているものがあったので、Textとして学習させました。ネガに入れてください）unworn clothes, (学習画像の中に服を脱いで床に散乱しているものがあったので、このようにタグをつけました。その際Skirtなどの散乱している物自体もタグをそれぞれつけています）bed, bed sheets, pillow,tatami,katana,（これらも背景に含まれていたため学習させています)orgasm,(ゲーム内の絶頂時の顔についてこのタグを学習させてます）embarrassed（ゲーム内で照れているときはこのタグを学習させています）Blush（少しだけ照れているときの頬の赤み）nose blush（鼻まで赤くなる頬の赤み。NSFWの学習素材はたいていこれだったと思います）dark-skinned male, solo focus,（ゲーム内に主人公（男性）が描写されている画像はこのようにタグを学習させています。つまりこれらのタグをいれることでゲームのような顔が映っていない男性が生成されやすいです）mosaic censoring,（ゲームのCGはモザイクがありました。このタグを与えています。モザイクが不要ならこのタグをネガにいれてください）================================================================設定サンプルをダウンロードしてPNG情報を読み込んでTxt2Imgなどに転送するのが確実でしょう。StepとCFGについてはこのLORAではなく生成に使用したベースモデルの推奨設定です。Euler aClip skip2================================================================補足事項全身が映っている学習画像にゲーム内の立ち姿のデータを利用したのですが、もっとアップにする必要があったのか、生成時に全身を描画しようとするとキャラクターが崩れやすいです。(full body,)とはいえ、サンプルにあるようにほぼ全身が映っていても崩れない場合もあるようなので、このタグの使用を避けるのがいいのかもしれません。学習素材の偏りが原因なのかNSFWの特定の体位について似たような構図や背景が出やすいです。（逆に言えば、ゲーム内のＣＧにかなり近い絵が出力できます）================================================================もしBUZZが余ったらいずれ修正を行うつもりです（左右反転など）戦闘服の衣装を学習させていませんし、武器も学習させていません。あまり多く学習させるとキャラクター再現に支障がでると考えたためです。このLORAはCIVITAIのトレーニングサービスでLORAを作成しました。これらの服装はひとつのデータセットでつくったのですが、複数のデータセット構造でLORAを作成する手順もあるそうですね？詳しい方は是非教えてください。特殊な目を持つキャラクターの学習方法がわかりません。私はいくつかLORAをつくっているのですが、どうにもうまくいかず、このLORAは問題点があるものの、ある程度目標をクリアできたので公開することにしました。フィードバックなどをいただければ改善につながるのでよろしくお願いします。サンプル画像の生成にはAutismMix LightningStep8を利用しています。