<h3 id="introduction">Introduction</h3>This LoRA is trained on <a rel="ugc" href="https://ponderogen.booth.pm/items/6106863">Shinano</a> (しなの, 信浓), a VRChat avatar created by <a rel="ugc" href="https://ponderogen.booth.pm/">Ponderogen</a> (ぽんでろ).<h3 id="usage">Usage</h3>Recommended LoRA strength: <code>0.8-1.0</code> Trigger word: <code>shinano (vrchat)</code>In ComfyUI, escape the parentheses:<pre><code>shinano \(vrchat\)</code></pre>Basic appearance:<pre><code>grey hair, blue eyes, cat ears, side braid, black cross hairclips, cat tail</code></pre>Default outfit:<pre><code>black sleeveless high-neck blouse, black ribbon bow at the neck, grey pleated skirt, black pantyhose, brown ankle boots, white off-shoulder cable-knit cardigan, wide folded white ribbed collar, vertical twisted cable-knit panels, ribbed cuffs and hem, gold buttons, white front belt tab with a rectangular gold buckle, two long flat tan straps hanging below the skirt</code></pre>Default outfit (without cardigan):<pre><code>black sleeveless high-neck blouse, black ribbon bow at the neck, grey pleated skirt, grey belt, black pantyhose, brown ankle boots, bare shoulders</code></pre><h3 id="training-details">Training Details</h3><ul><li>Training tool: <code>sorryhyun/anima_lora v1.17.1</code></li><li>Dataset: 11 manually captioned files, including 7 selected source images and 4 detail crops</li><li>LoRA rank / alpha: 64 / 256</li><li>Learning rate: <code>2e-5</code></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>Note: v0.9 is a beta version. To recover her appearance and settings precisely, it still carries 3D modeling vibes. I'm collecting more illustration images and hopefully it would get better in the next version.<h3>简介</h3>这个 LoRA 基于 <a target="_blank" rel="ugc" href="https://ponderogen.booth.pm/items/6106863">Shinano</a>（しなの、信浓）进行训练。Shinano 是由 <a target="_blank" rel="ugc" href="https://ponderogen.booth.pm/">Ponderogen</a>（ぽんでろ）制作的 VRChat 虚拟形象。<h3>使用</h3>推荐 LoRA 权重： <code>0.8–1.0</code> 激活词： <code>shinano (vrchat)</code>在 ComfyUI 中，请对括号进行转义：<pre><code>shinano \(vrchat\)</code></pre>基础外观：<pre><code>grey hair, blue eyes, cat ears, side braid, black cross hairclips, cat tail</code></pre>默认服装：<pre><code>black sleeveless high-neck blouse, black ribbon bow at the neck, grey pleated skirt, black pantyhose, brown ankle boots, white off-shoulder cable-knit cardigan, wide folded white ribbed collar, vertical twisted cable-knit panels, ribbed cuffs and hem, gold buttons, white front belt tab with a rectangular gold buckle, two long flat tan straps hanging below the skirt</code></pre>默认服装（不穿外套）：<pre><code>black sleeveless high-neck blouse, black ribbon bow at the neck, grey pleated skirt, grey belt, black pantyhose, brown ankle boots, bare shoulders</code></pre><h3>训练参数</h3><ul><li>训练工具： <code>sorryhyun/anima_lora v1.17.1</code></li><li>训练数据： 11 个手工标注文件，包括 7 张精选源图和 4 张服装细节裁图</li><li>LoRA rank / alpha： 64 / 256</li><li>学习率： <code>2e-5</code></li><li>优化器： AdamW</li><li>学习率调度： 余弦衰减，预热比例为 5%</li><li>训练量： 900 steps</li><li>训练精度： BF16</li><li>训练分辨率： 按长宽比分桶，约 1 百万像素</li><li>训练显卡： NVIDIA GeForce RTX 3090</li></ul>注意: v0.9 是一个测试版本。为了精确还原她的外观，仍然有比较强的 3D 味。我正在收集更多的平面插画图片，希望能在下一个版本变得更好。

Files with SHA256: c5fe50c30fc43a479ed7980263489eada3abd945bbd286a18a4d15f0d80b222a