<h3 id="introduction">Introduction</h3>This LoRA is trained on <a target="_blank" rel="ugc" href="https://glitter-lens-03b.notion.site/17c0616c62d680bd982ecfcf19cb70df">Nekomiya Shiromi</a> (猫宮しろみ), an indie Virtual YouTuber designed by artist <a target="_blank" rel="ugc" href="https://www.pixiv.net/users/32213152">momomaron</a>.This character has some intricate outfit details (the fishbone hair ornament, hair clips of different shapes on the left, centre and right, the belt at her waist, and so on), and I cannot precisely reproduce them with Anima in v0.1 (though I have done my best)! You may want to generate several times to improve consistency. I'll try to improve them in the next version (I'm working on something "more interesting" recently, so please be patient ♥️)<h3 id="usage">Usage</h3>Recommended LoRA strength: <code>1.0</code>Trigger word: <code>nekomiya shiromi (vtuber)</code>In ComfyUI, escape the parentheses:<pre><code>nekomiya shiromi \(vtuber\)</code></pre>Beret outfit:<pre><code>nekomiya shiromi \(vtuber\), white hair, long hair, green eyes, cat ears, animal ear fluff, ahoge, cat tail
bandaid hair ornament, fish bone hair ornament, hairpin, beret, blue hat, hat, blue sailor collar, sailor collar, blue bow, bow, blue jacket, open jacket, open clothes, white shirt, sleeveless, sleeveless shirt, puffy long sleeves, sleeves past fingers, sleeves past wrists, frills, white skirt, thigh strap, tail ornament</code></pre>Maid outfit:<pre><code>nekomiya shiromi \(vtuber\), white hair, long hair, green eyes, cat ears, animal ear fluff, ahoge, cat tail
braid, x hair ornament, maid headdress, hair bell, jingle bell, black ribbon, hair ribbon, tassel, maid, chinese clothes, blue dress, frilled dress, frills, puffy sleeves, wide sleeves, frilled sleeves, juliet sleeves, apron, white apron, frilled apron, waist apron, belt, white thighhighs, frilled thighhighs, cat paw ornament with bow, black shoes, tail ornament</code></pre><h3 id="training-details">Training Details</h3><ul><li>Training tool: sorryhyun/anima_lora v1.17.1</li><li>Dataset: 15 manually captioned files</li><li>LoRA rank / alpha: 32 / 128</li><li>Learning rate: 2e-5</li><li>Optimizer: AdamW</li><li>Learning rate schedule: Cosine decay with 5% warmup</li><li>Training duration: 900 steps</li><li>Precision: BF16</li><li>Resolution: Aspect-ratio buckets, approximately 1 megapixel</li><li>GPU: NVIDIA GeForce RTX 3090</li></ul><h3>简介</h3>这个 LoRA 在独立 VTuber <a target="_blank" rel="ugc" href="https://glitter-lens-03b.notion.site/17c0616c62d680bd982ecfcf19cb70df">猫宮しろみ</a>（ねこみや しろみ）上训练，角色设计者为画师 <a target="_blank" rel="ugc" href="https://www.pixiv.net/users/32213152">momomaron</a>。这个角色有一些复杂的服饰细节（例如鱼骨头饰、左中右不同的发卡形状、腰上的皮带等），我无法很好地在 v0.1 版本中精确还原它们（不过我已经尽力做到）！你可能希望多生成几次来提高稳定性。我会在下一个版本尝试进一步缓解细节问题（我最近正在忙着研究一些更有趣的东西，所以请保持耐心❤️）。<h3>使用方法</h3>推荐 LoRA 权重： <code>1.0</code> 激活词： <code>karin (vrchat)</code>在 ComfyUI 中需要转义括号：<pre><code>nekomiya shiromi \(vtuber\)</code></pre>贝雷帽：<pre><code>nekomiya shiromi \(vtuber\), white hair, long hair, green eyes, cat ears, animal ear fluff, ahoge, cat tail
bandaid hair ornament, fish bone hair ornament, hairpin, beret, blue hat, hat, blue sailor collar, sailor collar, blue bow, bow, blue jacket, open jacket, open clothes, white shirt, sleeveless, sleeveless shirt, puffy long sleeves, sleeves past fingers, sleeves past wrists, frills, white skirt, thigh strap, tail ornament</code></pre>女仆：<pre><code>nekomiya shiromi \(vtuber\), white hair, long hair, green eyes, cat ears, animal ear fluff, ahoge, cat tail
braid, x hair ornament, maid headdress, hair bell, jingle bell, black ribbon, hair ribbon, tassel, maid, chinese clothes, blue dress, frilled dress, frills, puffy sleeves, wide sleeves, frilled sleeves, juliet sleeves, apron, white apron, frilled apron, waist apron, belt, white thighhighs, frilled thighhighs, cat paw ornament with bow, black shoes, tail ornament</code></pre><h3>训练详情</h3><ul><li>训练工具： sorryhyun/anima_lora v1.17.1</li><li>数据集： 15 个手工标注的文件</li><li>LoRA rank / alpha： 32 / 128</li><li>学习率： 2e-5</li><li>优化器： AdamW</li><li>学习率调度： 余弦衰减，预热阶段占 5%</li><li>训练时长： 900 步</li><li>精度： BF16</li><li>分辨率： 按宽高比分桶，约 100 万像素</li><li>显卡： NVIDIA GeForce RTX 3090</li></ul>

Files with SHA256: d220157425c4bd2e5b46eadee0a5391be869ef2c61275e675aa4ad444316441a