<h1>Le Malin (Azur Lane) LoRA</h1><h2>12-05-2023/V7 Update:</h2><ul><li>Improve Le Malin Muse accuracy</li></ul><h2>16-02-2023/V6 Update:</h2><ul><li>Reduce the model file size to 9MB.</li></ul><ul><li>Use a way lower network dim (8 instead of 160) and Alpha (1 instead of 160) to prevent overfitting</li><li>More flexible</li></ul><ul><li>Rebalanced dataset (I uploaded it)</li><li>Trained on NAI instead of AnythingV3</li><li>Add a new token for each canon outfit (using booru token was a bad idea): <code>LeMalinDefault</code><code>LeMalinLapin</code><code>LeMalinMuse</code><code>LeMalinSundress</code><code>LeMalinSwimsuit</code></li></ul>I will probably make a V7 that use noise offset.I don't recommend using V5 (overfitted)This is a LoRA, this mean that you need to update Voldy/AUTOMATIC1111's webui to use it (use <code>git pull</code>) .<h3>Recommended Weight:</h3>1.0 unless you want to use multiple LoRAsFor Le Malin Muse: 0.8-1.0 (because of background overfit)<h3>Recommended prompt:</h3>That LoRA can reproduice all her game skin and version (as 12/05/2023).Default Skin:<pre><code>double bun, underboob cutout, white dress, white pantyhose, le_malin_(azur_lane), pauldrons, epaulettes, pleated dress ,cape</code></pre>For V6 and onwards:: You need to replace <code>le_malin_(azur_lane)</code> by <code>LeMalinDefault</code>Sundress :<pre><code>double bun, long hair, short dress, sleeveless, sleeveless dress, sundress, white dress, white_flower, white_rose, halter dress, black camisole, black collar, spiked collar,le_malin_(the_knight's_true_nature)_(azur_lane) </code></pre>For V6 and onwards:: You need to add <code>LeMalinSundress</code>Swimsuit :<pre><code>double bun, long hair, swimsuit, white bikini, le_malin_(wednesday's_retreat)_(azur_lane), halterneck, criss-cross halter, white bikini,arm scrunchie, arms behind head,black collar</code></pre>For V6 and onwards:: You need to add <code>LeMalinSwimsuit</code>Le Malin µ :For V5 and V6:<pre><code>ballerina, ballet dress, white pantyhose,white_footwear, wings, ballerina, double bun, long hair,tutu,white leotard,le_malin_(muse)_(azur_lane)</code></pre>I suspect that this LoRA overfit it. (V5)It had more repeat than the other skin/version because there is actually few SFW art of that skins. (V5)16-02-2023/V6 Update: You need to add <code>LeMalinMuse</code>. That outfit is a bit less accurate on V6...but on the other hand it is less overffited. (no background overfitting,more flexible)12-05-2023/V7 Update: Use this instead :<pre><code>LeMalinMuse, ballerina,ballet dress, white pantyhose, double bun, tutu, le_malin_(muse)_(azur_lane) , idol, idol_clothes, see-through,wing_hair_ornament, long sleeves, covered_navel, star \(symbol\), </code></pre>Bunny girl:<pre><code>white choker, white leotard, white pantyhose, white wrist cuffs,twin braids,le_malin_(listless_lapin)_(azur_lane),fake animal ears, sleeveless, clothing cutout, rabbit ears,</code></pre>(It is quite ironic that this borderline NSFW skin has way more SFW arts than Le Malin µ).For V6 and onwards: You need to add <code>LeMalinLapin</code><h3></h3><h3>Model used for the example:</h3>V7 examples: YesMix v16V6 examples : AnythingV4.5V5 examples : AnythingV4.5 and Pastel Mix .<h3></h3>That character is owned by Manjuu Games and Yostar.

Le Malin (Azur Lane) (All skins)  (9MB update)

Training settings:<pre><code>286 images
variable repeats (15 for Le Malin µ,10 for her default/swimsuit/sundress skin and 4 for her bunny girl skin, 4 for non canon outfits) 
4 epochs
9220 steps (batch size:1)
scheduler: cosine with restart
Learning rate: 1e-4
Text encoder learning rate: 5e-5 
Unet learning rate: 1e-4 
max bucket resolution: 512 
clip skip: 2 
dim/alpha: 160 
batch size=1 
Base model: AnythingV3, this mean that this LoRA work on any model derived from that particular model like AnythingV4/4.5, Pastel Mix, AOM 1/2,etc...</code></pre>

Files with SHA256: 1cae84a998f0282cdcb3cc30599d4b89f3b912868bdbfe4965276ad46d95b667