<h1 id="heading-2321">Le Terrible | ル・テリブル | 可怖 LoRA</h1>I finally managed to get one right with this character. I wonder if I should try make one for L'Indomptable.<h2 id="heading-2948">Weight:</h2>Recommanded weight 0.8-1.0 (I only tested it at 1.0)You won't have to use a lower weight than 1.0 unless you want to use multiple LoRAs<h2 id="heading-2949">Recommended prompt:</h2>Add <code>one_eye_covered, eyepatch, flower_eyepatch</code> to generate this character with her eyepatch. I haven't trimmed those tags from the dataset to allow this LoRA to generate this character without that eyepatch.<h3 id="heading-2950">Skins/outfits:</h3>Default outfit:<pre><code>LeTerribleDefault, blue_bowtie, breasts_apart, center_opening, clothing_cutout,gauntlets, kepi, vambraces, white_background, white_dress, short_over_long_sleeves, bodystocking_under_clothes, bob_cut, eyepatch, pom_pom_(clothes), short_dress, armored_boots,</code></pre>Swimsuit skin (Sunlight, Seaspray, and Sanctity):<pre><code>LeTerribleSwimsuit, blue_bikini, eyepatch, flower_eyepatch, frilled_bikini, long_sleeves,necklace, off_shoulder, one_eye_covered,polka_dot_bikini,ribbon,white_jacket, crown_hair_ornament</code></pre>Nightgown skin (Silence, Slumber, and Solace):Cancelled:<ul><li>There is not enough drawing of it (huge risk of overfitting and pose fixing)</li><li>It tend to fuse with other skins/outfits</li></ul>Thoses drawing are still in the dataset but that LoRA was not trained with them.<h2 id="heading-2951">Model used for the example:</h2>V5: YesMix v1.6That character is owned by Manjuu Games and Yostar.

Le Terrible | ル・テリブル | 可怖 (Azur Lane) 

<pre><code>40 images
10 repeats 
6 epochs
2400 steps (batch size:1)
scheduler: cosine
Text encoder learning rate: 1e-5 
Unet learning rate: 0.001</code></pre>

Files with SHA256: daad6f4b0e52cdc0f04147e27b1b61260c2c583c812d40a4c79f290a41ea5463