Quick automatic manga coloring (Preview)

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.

📥 Workflow: the complete ComfyUI workflow is downloadable right from this page — check the Download menu for manga_colorize_krea2.json. It is also available as a standalone workflow item: Manga Colorize Krea 2 Workflow. All preview images embed the workflow as well — drop any of them onto the ComfyUI canvas.

Strengths

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

Known issues

• Character consistency can only be steered through prompts.
• Slight pixel drift/offset can occur.
• Occasionally produces meaningless mosaic artifacts.

Usage

• Recommended LoRA strength: 0.7 – 1.0.
• Trigger phrase: 漫画上色 ("manga coloring" in Chinese). Krea 2 understands Chinese prompts — append character / scene descriptions after the trigger.
• UNet: the official original Krea 2 model works fine.
• Required custom nodes: Comfyui-RayNodes.
• Optional: the Krea 2 lineart-edit model (Cierpliwy/krea2_lineart_edit) can be added at low strength (≤ 0.5) when the pixel offset gets large.
• Keep the source image sharp enough for correct text extraction — a short side of 1000 px or more is usually sufficient.
• If certain details keep coloring poorly, add explicit prompt words to draw attention to them — correct prompts help a lot.

Preview images carry the embedded ComfyUI workflow — drop an image onto the ComfyUI canvas to load it.

Roadmap

A larger dataset is being prepared for Krea 2 KV-cache training and a Qwen 2.1 based version.

License

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).

---

快速漫画上色（Preview）

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。

📥 工作流：完整 ComfyUI 工作流可直接在本页「下载」菜单获取（manga_colorize_krea2.json）；也可以用独立的工作流页面：Manga Colorize Krea 2 Workflow。所有样例图片同样内嵌工作流，拖进 ComfyUI 画布即可加载。

优点

• 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。
• 漫画中小面积的场景也能够精细上色。
• 不会被原始阴影和网点降低饱和度。

已知问题

• 人物一致性仅能靠提示词控制。
• 存在像素偏移。
• 偶尔会出现没有意义的马赛克。

使用方法

• 推荐 LoRA 强度：0.7 – 1.0。
• 触发词：漫画上色，后面可以接中文的人物／场景描述（Krea 2 支持中文提示词）。
• UNet 使用 Krea 官方原始模型即可。
• 必须节点：Comfyui-RayNodes。
• 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：Cierpliwy/krea2_lineart_edit。
• 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。
• 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。

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。

后续计划

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。

许可

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。