<h3 id="introduction">Introduction</h3>This LoRA is trained on <a rel="ugc" href="https://booth.pm/ja/items/8562330">Kaguya</a> (輝夜, かぐや), a VRChat avatar created by <a rel="ugc" href="https://paryi.booth.pm/">IKUSIA</a>.<h3 id="usage">Usage</h3>Recommended LoRA strength: <code>1.0</code> Trigger word: <code>kaguya (vrchat)</code>In ComfyUI, escape the parentheses:<pre><code>kaguya \(vrchat\)</code></pre>Default outfit:<pre><code>black serafuku, black sailor collar, red bow, bell, black cardigan, long sleeves, black pleated skirt, black pantyhose, black footwear, black beret, multiple tails</code></pre>The trigger alone is enough to bring out the character's identity. Add the outfit block above if you want the full original look.<h3 id="training-details">Training Details</h3><ul><li>Trainer: <code>sorryhyun/anima_lora v1.17.1</code></li><li>Dataset: 15 captioned images</li><li>LoRA rank / alpha: <code>32 / 128</code></li><li>Learning rate: <code>2e-5</code></li><li>Optimizer: AdamW</li><li>Scheduler: Cosine, with 5% warmup</li><li>Released checkpoint: Step <code>2,000</code></li><li>Precision: BF16</li><li>Resolution: Aspect-ratio buckets at approximately 1 megapixel</li><li>GPU: NVIDIA RTX 3090</li></ul><h3>简介</h3>这是基于 Anima Base 1.0 训练的 <a target="_blank" rel="ugc" href="https://booth.pm/ja/items/8562330">Kaguya</a>（輝夜，かぐや）角色 LoRA。Kaguya 是由 <a target="_blank" rel="ugc" href="https://paryi.booth.pm/">IKUSIA</a> 制作的 VRChat 虚拟形象。<h3>使用</h3>推荐 LoRA 权重： <code>1.0</code> 激活词： <code>kaguya (vrchat)</code>在 ComfyUI 中需要转义括号：<pre><code>kaguya \(vrchat\)</code></pre>常服：<pre><code>black serafuku, black sailor collar, red bow, bell, black cardigan, long sleeves, black pleated skirt, black pantyhose, black footwear, black beret, multiple tails</code></pre>激活词单独使用即可带出角色人设。如需完整的原设造型，请附带上述常服 prompt。<h3>训练参数</h3><ul><li>训练工具： <code>sorryhyun/anima_lora v1.17.1</code></li><li>数据集： 15 张带 caption 的图片</li><li>LoRA rank / alpha： <code>32 / 128</code></li><li>学习率： <code>2e-5</code></li><li>优化器： AdamW</li><li>学习率调度： Cosine，5% warmup</li><li>发布权重： 第 <code>2,000</code> 步</li><li>训练精度： BF16</li><li>训练分辨率： 约 1 M 像素的宽高比分桶</li><li>显卡： NVIDIA RTX 3090</li></ul>

Files with SHA256: 38b5992b7702716aafcc72e034fd3cebd14cb2800307fc5712f06b1f75712fb1