Character LoRA trained for WAI v17 / Illustrious-SDXL.This LoRA is designed to recall:- character identity- long pink hair- blue eyes- cat ears and cat tail- blue flower hair ornament- gold headband and ornate headwear- white and gold Egyptian-inspired outfit- translucent veil / long robe- gold jewelry and accessoriesRecommended trigger word:textpepe_xingyi<h2 id="recommended-settings">Recommended Settings</h2><ul><li>Recommended weight: <code>0.7</code></li><li>Usable range: <code>0.6-0.8</code></li><li>Base model: trained and tested on <code>WAI-Nsfw-Illustrious-17</code></li><li>Works best with WAI / Illustrious-style SDXL models</li><li>Not intended for Pony, realistic XL, or non-Illustrious checkpoints</li></ul>Weight notes:WeightBehavior<code>0.6</code>More flexible prompt following, slightly weaker identity recall<code>0.7</code>Best balance, recommended default<code>0.8</code>Stronger identity and outfit recall, slightly more pose/background bias<code>0.9-1.0</code>Not recommended by default; can overfit pose/background<h2 id="basic-prompt">Basic Prompt</h2><pre><code>&lt;lora:pepe_xingyi_recommended_e08:0.7&gt;, pepe_xingyi, pepe (arknights), 1girl, solo, cat ears, cat tail, long tail, long pink hair, blue eyes, blue flower hair ornament, gold headband, ornate headwear, egyptian-inspired outfit, white and gold outfit, white dress, sleeveless dress, long open-front robe, translucent veil, sheer fabric, gold jewelry, armlet, bracelet, anklet, upper body, looking at viewer, simple background</code></pre><h2 id="full-body-prompt">Full Body Prompt</h2><pre><code>&lt;lora:pepe_xingyi_recommended_e08:0.7&gt;, pepe_xingyi, pepe (arknights), 1girl, solo, cat ears, cat tail, long tail, long pink hair, blue eyes, blue flower hair ornament, gold headband, ornate headwear, egyptian-inspired outfit, white and gold outfit, full body, standing, looking at viewer, detailed outfit, simple background</code></pre><h2 id="portrait-prompt">Portrait Prompt</h2><pre><code>&lt;lora:pepe_xingyi_recommended_e08:0.65&gt;, pepe_xingyi, pepe (arknights), 1girl, solo, cat ears, long pink hair, blue eyes, blue flower hair ornament, gold headband, portrait, close-up, looking at viewer, smile, simple background, detailed eyes</code></pre><h2 id="negative-prompt">Negative Prompt</h2><pre><code>low quality, worst quality, blurry, bad anatomy, bad hands, extra fingers, missing fingers, extra limbs, deformed, disfigured, bad face, asymmetrical eyes, text, watermark, logo, signature, cropped head, cropped feet, duplicate, multiple girls, wrong hair color, wrong eye color, wrong outfit, black outfit, simple black dress, missing cat ears, missing tail, nsfw, nude, naked, exposed breasts, loli, child, underage</code></pre><h2 id="training-info">Training Info</h2><ul><li>Dataset: 113 cleaned PNG images</li><li>Resolution: 1024</li><li>Caption prefix: <code>pepe_xingyi, 1girl</code></li><li>Base model: <code>WAI-Nsfw-Illustrious-17</code></li><li>Training script: <code>sd-scripts / sdxl_train_network.py</code></li><li>Network: LoRA</li><li>UNet only: yes</li><li>Network dim / alpha: <code>32 / 16</code></li><li>Dropout: <code>0.05</code></li><li>Optimizer: <code>AdamW8bit</code></li><li>Mixed precision: <code>bf16</code></li><li>Repeats: <code>6</code></li><li>Batch size: <code>1</code></li><li>Gradient accumulation: <code>2</code></li><li>Epochs: <code>10</code></li><li>Selected checkpoint: <code>epoch 8</code></li></ul><h2 id="evaluation">Evaluation</h2>The selected checkpoint was chosen after testing:<ul><li>epochs: <code>4, 5, 6, 7, 8, 9, 10</code></li><li>weights: <code>0.6, 0.7, 0.8, 0.9, 1.0</code></li><li>seeds: <code>2026062601, 2026062602</code></li><li>prompts: upper body, full body, portrait, sitting indoor</li><li>total test images: <code>280</code></li></ul>Best overall checkpoint:<pre><code>pepe_xingyi-000008.safetensors</code></pre>Reason:Epoch 8 at weight 0.7 gives the best balance between identity recall, outfit accuracy, prompt following, and reduced overfit. Epoch 9/10 are stronger but more prone to pose/background bias.

Files with SHA256: ef8efdb0accc528d40882e3b540be75f34f68c6a2a9e5e21b62165aabf925d92