<h2 id="introduction">Introduction</h2>This concept/style LoRA captures animegao kigurumi photography, including rigid anime masks, synthetic wigs, realistic costumes, and human body proportions.Someone asked me if I could train the same style on some models that are better at realism, so here it is!If you're looking for one that is better at anime characters, please check out the Anima model version: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2922637/style-animegao-kigurumi-or-or">https://civarchive.com/models/2922637/style-animegao-kigurumi-or-or</a><h2 id="usage">Usage</h2>Recommended LoRA strength: 0.8-1.0 (a bit overtrained, so a lower strength is recommended)Trigger word:<pre><code>animekig</code></pre>Use Krea 2 Turbo and describe the appearance, clothing, pose, and environment in natural language. Recommended settings are: 8 steps, CFG 1, Euler / simple, ModelSamplingFlux shift 1.15. For example:<pre><code>animekig. A full-body studio photograph of a subject with a pink wig and blue eyes, wearing a maid dress and white thigh-high stockings.</code></pre><pre><code>animekig. A photograph of two subjects wearing colorful idol costumes, posing together on a brightly lit stage.</code></pre><h2 id="training-details">Training Details</h2><ul><li>Base model: Krea 2 RAW</li><li>Training tool: <code>kohya-ss/musubi-tuner</code></li><li>Dataset: 176 curated photographs and 14 detail crops</li><li>LoRA rank / alpha: 32 / 32</li><li>Learning rate: <code>1e-4</code> initially; <code>5e-5</code> during targeted refinement</li><li>Optimizer: AdamW8bit</li><li>Learning rate schedule: Constant, without warmup</li><li>Training duration: 2,000 initial steps; a separate 1,200-step refinement initialized from the initial step-750 checkpoint</li><li>Released checkpoint: Refinement step 800, initialized from initial step 750</li><li>Precision: BF16 mixed precision with a scaled FP8 base model; saved LoRA weights are FP32</li><li>Resolution: Aspect-ratio buckets at 1024 for full images and 512 for detail crops</li><li>GPU: NVIDIA GeForce RTX 3090</li></ul><h2>简介</h2>这是一个学习 Animegao Kigurumi 摄影风格的概念/风格 LoRA，涵盖硬质动漫头壳、假发、真实服装和真人身体比例。有人问我能否在一些更擅长真实感的模型上训练相同的风格，所以这就来了！如果你在寻找一个更擅长卡通形象的，请查看 Anima 模型版本：<a target="_blank" rel="ugc" href="https://civarchive.com/models/2922637/style-animegao-kigurumi-or-or">https://civarchive.com/models/2922637/style-animegao-kigurumi-or-or</a><h2>使用</h2>推荐权重：0.8-1.0（有点过训练了，所以建议使用较低的权重）激活词：<pre><code>animekig</code></pre>搭配 Krea 2 Turbo，在激活词后用自然语言描述外观、服装、姿势和环境。推荐参数：8 步、CFG 1、Euler / simple、ModelSamplingFlux shift 1.15。例如：<pre><code>animekig. A full-body studio photograph of a subject with a pink wig and blue eyes, wearing a maid dress and white thigh-high stockings.</code></pre><pre><code>animekig. A photograph of two subjects wearing colorful idol costumes, posing together on a brightly lit stage.</code></pre><h2>训练参数</h2><ul><li>基础模型：Krea 2 RAW</li><li>训练工具：<code>kohya-ss/musubi-tuner</code></li><li>训练数据：176 张筛选照片，以及 14 张局部裁切</li><li>LoRA rank / alpha：32 / 32</li><li>学习率：初始训练 <code>1e-4</code>；针对性训练 <code>5e-5</code></li><li>优化器：AdamW8bit</li><li>学习率调度：恒定学习率，无预热</li><li>训练量：初始训练 2,000 步；另从初始 step750 初始化，进行 1,200 步针对性训练</li><li>发布版本：从初始 step750 初始化后的针对性训练 step800</li><li>训练精度：BF16 混合精度，底模使用 scaled FP8；保存的 LoRA 权重为 FP32</li><li>训练分辨率：按长宽比分桶，全图 1024、局部 512</li><li>训练显卡：NVIDIA GeForce RTX 3090</li></ul>

Files with SHA256: 9eea370c4c9a0bc81971dbf1de8ddb5da9ec8374e251c1ac3b1273bb6554a440