建议强度：0.8左右，不要太高 <h3>训练信息</h3>本模型是一款面向 Krea 2 的画风 LoRA，旨在学习并复现画师 miwano_rag 的视觉风格。<h3>数据集</h3><ul><li>来源：Danbooru 中 <code>miwano_rag</code> 画师标签下的公开图片</li><li>内容：人工筛选的高质量插画</li><li>数量：共 341 张</li><li>筛选方式：人工检查并剔除低质量、重复以及风格偏差较大的图片</li><li>标注方式：使用 Gemini 3.7 Flash 自动生成自然语言描述</li></ul><h3>训练方法</h3>训练采用阶梯式分辨率训练（Staged Progressive Resolution Training）。模型首先在较低分辨率下学习整体画风、配色和构图特征，随后逐步提高训练分辨率并降低学习率，从而进一步学习线条、材质和局部细节。三个阶段均基于同一数据集连续训练：<ul><li>512px：10 epochs</li><li>768px：10 epochs</li><li>1024px：10 epochs</li><li>总计：30 epochs</li></ul><h3>训练配置</h3>阶段Batch size梯度累积有效 Batch size学习率约更新步数512px224<code>5e-4</code>853768px122<code>3e-4</code>1,7051024px111<code>1e-4</code>3,410<ul><li>基础模型：<code>krea/Krea-2-Raw</code></li><li>训练工具：Ostris AI-Toolkit</li><li>网络类型：LoRA</li><li>Rank / Alpha：32 / 32</li><li>优化器：AdamW 8-bit</li><li>权重精度：BF16</li><li>学习率调度器：Cosine，最低学习率比例为 0.1</li><li>Caption Dropout：0.05</li><li>训练硬件：单张 NVIDIA GeForce RTX 4090</li></ul><h3>使用建议</h3>本 LoRA 建议搭配 Krea 2 使用。经过实际测试，即使不在提示词中添加特定触发词，也能正常呈现目标画风。建议从 LoRA 权重 <code>0.8</code> 开始测试，并根据所需的画风强度、角色一致性和提示词遵循度进行调整。较高权重会强化画风特征，但也可能降低基础模型对提示词的响应能力。<hr /><h3 id="training-details">Training Details</h3>This is a style LoRA for Krea 2, trained to reproduce the visual characteristics associated with the artwork of miwano_rag.<h3 id="dataset">Dataset</h3><ul><li>Source: Public images collected from the <code>miwano_rag</code> artist tag on Danbooru</li><li>Content: Manually curated, high-quality illustrations</li><li>Size: 341 images</li><li>Curation: Low-quality, duplicate, and stylistically inconsistent images were removed manually</li><li>Captioning: Natural-language captions generated with Gemini 3.7 Flash</li></ul><h3 id="training-method">Training Method</h3>The LoRA was trained using staged progressive resolution training.Lower-resolution training was used to establish the overall visual style, color palette, and compositional characteristics. The training resolution was then progressively increased while the learning rate was reduced, allowing the model to refine its linework, materials, and local details.All stages were trained sequentially on the same dataset:<ul><li>512px: 10 epochs</li><li>768px: 10 epochs</li><li>1024px: 10 epochs</li><li>Total: 30 epochs</li></ul><h3 id="training-configuration">Training Configuration</h3>StageBatch sizeGradient accumulationEffective batch sizeLearning rateApprox. updates512px224<code>5e-4</code>853768px122<code>3e-4</code>1,7051024px111<code>1e-4</code>3,410<ul><li>Base model: <code>krea/Krea-2-Raw</code></li><li>Training framework: Ostris AI-Toolkit</li><li>Network type: LoRA</li><li>Rank / Alpha: 32 / 32</li><li>Optimizer: AdamW 8-bit</li><li>Weight precision: BF16</li><li>Learning-rate scheduler: Cosine with a minimum learning-rate ratio of 0.1</li><li>Caption dropout: 0.05</li><li>Hardware: Single NVIDIA GeForce RTX 4090</li></ul><h3 id="usage-recommendations">Usage Recommendations</h3>This LoRA is intended for use with Krea 2. Testing indicates that no specific trigger phrase is required to activate the target style.A LoRA weight of <code>0.8</code> is recommended as the starting point. The weight can be adjusted according to the desired style strength, character consistency, and prompt adherence. Higher weights may strengthen the learned style but can also reduce the base model’s responsiveness to the prompt.

Files with SHA256: b05e1775c1076d89c0d1010f39b741b2ee8459f80273f6f266e4ec86d02643ee