<h3>Quick automatic manga coloring (Preview)</h3>
<p>This LoRA colorizes black-and-white manga pages quickly. It is a preview release trained on a small dataset with a low step count — expect some rough edges (see Known Issues). Versions for other base models will be added to this page later.</p>

<p>📥 Workflow: the complete ComfyUI workflow is downloadable right from this page — check the Download menu for <code>manga_colorize_krea2.json</code>. It is also available as a standalone workflow item: <a href="https://civarchive.com/models/2980606" rel="ugc">Manga Colorize Krea 2 Workflow</a>. All preview images embed the workflow as well — drop any of them onto the ComfyUI canvas.</p>

<h3>Strengths</h3>
<ul>
<li>Designed for Krea 2 turbo-style inference: low censorship, fast generation, and you can stack style LoRAs on top to adjust the coloring style.</li>
<li>Fine-grained coloring, even for small areas inside a panel.</li>
<li>Colors are not dulled or desaturated by the original shadows and screentones.</li>
</ul>

<h3>Known issues</h3>
<ul>
<li>Character consistency can only be steered through prompts.</li>
<li>Slight pixel drift/offset can occur.</li>
<li>Occasionally produces meaningless mosaic artifacts.</li>
</ul>

<h3>Usage</h3>
<ul>
<li>Recommended LoRA strength: 0.7 – 1.0.</li>
<li>Trigger phrase: <code>漫画上色</code> ("manga coloring" in Chinese). Krea 2 understands Chinese prompts — append character / scene descriptions after the trigger.</li>
<li>UNet: the official original Krea 2 model works fine.</li>
<li>Required custom nodes: <a href="https://github.com/sh570655308/Comfyui-RayNodes" rel="ugc">Comfyui-RayNodes</a>.</li>
<li>Optional: the Krea 2 lineart-edit model (<a href="https://huggingface.co/Cierpliwy/krea2_lineart_edit/tree/main" rel="ugc">Cierpliwy/krea2_lineart_edit</a>) can be added at low strength (≤ 0.5) when the pixel offset gets large.</li>
<li>Keep the source image sharp enough for correct text extraction — a short side of 1000 px or more is usually sufficient.</li>
<li>If certain details keep coloring poorly, add explicit prompt words to draw attention to them — correct prompts help a lot.</li>
</ul>
<p>Preview images carry the embedded ComfyUI workflow — drop an image onto the ComfyUI canvas to load it.</p>

<h3>Roadmap</h3>
<p>A larger dataset is being prepared for Krea 2 KV-cache training and a Qwen 2.1 based version.</p>

<h3>License</h3>
<p>Non-commercial use only. Commercial use is strictly prohibited. This includes selling the model or outputs, paid services built on it, and ad-funded re-uploads / mirrors of this model (including machine-translated copies).</p>

<hr />

<h3>快速漫画上色（Preview）</h3>
<p>本 LoRA 用于快速给黑白漫画上色。这是一个 preview 版本，基于小样本、低步数训练，仍存在一些问题（见「已知问题」）。之后其他底模的版本也会陆续发布在本页面。</p>

<p>📥 工作流：完整 ComfyUI 工作流可直接在本页「下载」菜单获取（<code>manga_colorize_krea2.json</code>）；也可以用独立的工作流页面：<a href="https://civarchive.com/models/2980606" rel="ugc">Manga Colorize Krea 2 Workflow</a>。所有样例图片同样内嵌工作流，拖进 ComfyUI 画布即可加载。</p>

<h3>优点</h3>
<ul>
<li>基于 Krea 2 turbo 推理：审核宽松、生成快速，并且可以叠加画风 LoRA 调整上色风格。</li>
<li>漫画中小面积的场景也能够精细上色。</li>
<li>不会被原始阴影和网点降低饱和度。</li>
</ul>

<h3>已知问题</h3>
<ul>
<li>人物一致性仅能靠提示词控制。</li>
<li>存在像素偏移。</li>
<li>偶尔会出现没有意义的马赛克。</li>
</ul>

<h3>使用方法</h3>
<ul>
<li>推荐 LoRA 强度：0.7 – 1.0。</li>
<li>触发词：<code>漫画上色</code>，后面可以接中文的人物／场景描述（Krea 2 支持中文提示词）。</li>
<li>UNet 使用 Krea 官方原始模型即可。</li>
<li>必须节点：<a href="https://github.com/sh570655308/Comfyui-RayNodes" rel="ugc">Comfyui-RayNodes</a>。</li>
<li>可选模型：Krea 2 的 lineart 模型，像素偏移较大时可以低强度（不超过 0.5）加入：<a href="https://huggingface.co/Cierpliwy/krea2_lineart_edit/tree/main" rel="ugc">Cierpliwy/krea2_lineart_edit</a>。</li>
<li>尽量保证原始图像的清晰度以便于文字正确提取，短边 1000 像素以上的分辨率通常是足够的。</li>
<li>如果发现部分细节上色总是不尽如人意，可以尝试增加提示词提升注意力，正确的提示词通常会很有帮助。</li>
</ul>
<p>样例图片均内嵌 ComfyUI 工作流，把图片拖到 ComfyUI 画布上即可直接加载。</p>

<h3>后续计划</h3>
<p>目前正在准备数据集，以便开展 Krea 2 KV cache 的训练和 Qwen 2.1 的训练。</p>

<h3>许可</h3>
<p>本模型严禁商用。包括但不限于：出售模型或其产出、基于本模型的付费服务、以及以营利为目的的转载／镜像（包括机翻打广告的搬运）。欢迎中文留言交流。</p>


Manga Colorize

<p>EN: This version blends two training snapshots (weight ratio 1.0 : 0.1) for a much lower pixel offset — in most cases no lineart correction pass is needed. Usage is unchanged: trigger <code>漫画上色</code>, strength 0.7–1.0. The preview images embed the workflow used for this version (drag onto the ComfyUI canvas).</p>
<p>中文：本版本以 1.0 : 0.1 的比例混合了两个训练快照，像素偏移大幅降低，多数情况无需再叠加 lineart 修正。用法不变：触发词 <code>漫画上色</code>，强度 0.7–1.0。预览图内嵌的工作流即为本版本所出，拖进 ComfyUI 画布即可加载。</p>

Files with SHA256: aba549bc3de01c37ef08daa737e0788dcd5dbd585e0481691fee6b682e29ebcd