<h3 id="introduction">Introduction</h3>This is a SeiroA style LoRA trained on Anima Base 1.0. <a target="_blank" rel="ugc" href="https://x.com/seiro_A">SeiroA</a> is an artist who primarily creates adult manga.I trained this LoRA at a friend's request. Along the way, I also learned some data-curation strategies for noisy training material such as manga, so I may release more models based on other manga artists whose work I personally like!<h3 id="usage">Usage</h3>Recommended LoRA strength: <code>1.0</code> Trigger word: <code>@seiroa</code><h3 id="training-details">Training Details</h3><ul><li>Base model: Anima Base 1.0</li><li>Trainer: <code>sorryhyun/anima_lora</code></li><li>Dataset: 285 character crops from manga</li><li>LoRA rank / alpha: <code>32</code> / <code>128</code></li><li>Learning rate: <code>2e-5</code></li><li>Optimizer: AdamW</li><li>Scheduler: Cosine, with 5% warmup</li><li>Batch size: <code>1</code> per GPU</li><li>Training duration: 20 epochs / 1,900 steps</li><li>Precision: BF16</li><li>Resolution: Aspect-ratio buckets at approximately 1 megapixel</li><li>GPU: 3 × NVIDIA RTX 3090</li></ul><h3>简介</h3>这是基于 Anima Base 1.0 训练的 SeiroA 画风 LoRA。<a target="_blank" rel="ugc" href="https://x.com/seiro_A">SeiroA</a> 是一位主要绘制成人漫画的画师。这个 LoRA 应朋友的请求训练。我在训练中也学习了一些数据清洗的策略（针对漫画之类的噪音较大的训练数据），所以我接下来可能还会发布一些其他我觉得较好的漫画画师的模型作品！<h3>使用</h3>推荐 LoRA 强度： <code>1.0</code> 激活词： <code>@seiroa</code><h3>训练参数</h3><ul><li>基础模型： Anima Base 1.0</li><li>训练工具： <code>sorryhyun/anima_lora</code></li><li>数据集： 285 张人物裁剪（来自漫画）</li><li>LoRA rank / alpha： <code>32</code> / <code>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>Batch size： 每卡 Batch 为 1</li><li>训练量： 20 epochs / 1,900 steps</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： 3 × NVIDIA RTX 3090</li></ul>

Files with SHA256: 9099697272a9f3b552048bd6d90db3a7c7c4e5c62d0827ecfd394535eb00813f