黑白漫画（含网点）语义上色工具箱。语义级自动上色 + 交互式精修，覆盖 Windows 桌面、浏览器与 Android 端侧三种使用方式，核心引擎为纯 Dart 确定性实现。
Semantic colorization for black-and-white manga (screentone-aware). Combines fully automatic ONNX semantic colorization with interactive refinement, available on Windows desktop, in the browser, and on-device (Android). The core engine is a deterministic pure-Dart implementation.
- Fully automatic — ONNX semantic colorization (SAM-guided generator) for hair / skin / eyes / background, on desktop, in the browser, and on-device in the Android app
- Color schemes — Standard / Vivid Cel / Soft Pastel / Impasto / Retro Showa; the default Standard finish neutralizes the model's paper-yellow cast and adaptively re-saturates muted output
- Hint points — click points on the canvas to pin region colors
- Reference image — transfer the overall tone of any colored image onto the page (Reinhard / histogram matching, strength control, paper-white protection; also available model-free in the Android app)
- Batch queue — up to 32 pages per run; a single failure never blocks the rest
- Long pages — automatic tiled inference above 1024px (256px overlap, linear feathering)
- Luminance fidelity — every mode writes only Lab a/b chroma; L always comes from the original page (sole exception: enclosed near-white foreground like white hair gets a small chroma-proportional L dip — at paper-white L there is no room for chroma to render)
- White-foreground rescue — white regions enclosed by ink (white hair, clothes, in-panel backgrounds) keep the model's intended colors; paper white connected to the page edge stays pure white
- GPU acceleration — DirectML first on Windows (any DX12 GPU), CUDA on NVIDIA, automatic CPU fallback
| Path | Purpose |
|---|---|
src-tauri/ |
Tauri 2 desktop shell (Windows) + Rust backend; bundles the Python sidecar |
tool/colorizer_service/ |
Python (uvicorn) colorization service — shared by the desktop sidecar and the browser workbench |
web/ |
Browser workbench frontend (static), served by the Python service |
mobile/ |
Flutter Android app — on-device ONNX colorization |
packages/manga_colorizer_core/ |
Pure-Dart colorization engine (hint-based chroma diffusion + luminance tint) |
packages/manga_colorizer_server/ |
Shelf-based HTTP backend over the core engine |
packages/manga_colorizer_cli/ |
CLI tools: batch colorize, tint, sample / comparison sheets |
models/ |
ONNX weights (gitignored; downloaded at runtime) |
Windows installer (recommended for most users)
- Download
Manga Colorizer_<version>_x64-setup.exefrom Releases and install. - On first launch the app downloads model weights (~287 MB, one-time, to
%LOCALAPPDATA%\manga-colorizer\weights). - Drag in manga pages → accept the disclaimer → start colorizing. Results appear in the Gallery tab.
Requirements: Windows 10 1809+ / 11 (WebView2 bundled by the installer). GPU acceleration needs a DX12 GPU (DirectML); falls back to CPU automatically.
Python service + browser workbench Requires Python ≥ 3.10.
# 1) Dependencies (CPU baseline)
python -m pip install -r tool/colorizer_service/requirements.txt
# Windows GPU (DirectML, any DX12 GPU): pick ONE, do not mix with the CPU build
python -m pip install -r tool/colorizer_service/requirements-directml.txt
# 2) Download model weights (~300 MB, not shipped with the repo;
# for CN network replace the host with hf-mirror.com)
python -c "import urllib.request as u; [u.urlretrieve(f'https://huggingface.co/sharky172/manga-light-colorizer/resolve/main/models/{n}', f'models/manga-light-colorizer/{n}') for n in ['v6_generator.onnx','v6_sam_encoder.onnx']]"
# 3) Start the service (binds 127.0.0.1:8788)
python -m uvicorn tool.colorizer_service.service:app --port 8788
# 4) Open http://127.0.0.1:8788/ in your browserNVIDIA CUDA path: pip install -r tool/colorizer_service/requirements-gpu.txt, supply matching CUDA/cuDNN runtimes, then launch with COLORIZER_DEVICE=cuda. The service fails loudly on CUDA errors and never silently falls back.
Build the desktop app from source Requires Rust stable, Node.js (for the Tauri CLI), and Python ≥ 3.10.
# 1) Build the Python sidecar (onefile exe; Python 3.10–3.12 for the GPU path)
python -m venv .venv
.venv/Scripts/pip install -r tool/colorizer_service/requirements-directml.txt pyinstaller
.venv/Scripts/pyinstaller sidecar/manga-colorizer-sidecar.spec --noconfirm \
--distpath build/sidecar-dist --workpath build/sidecar-work
cp build/sidecar-dist/manga-colorizer-sidecar.exe \
src-tauri/binaries/manga-colorizer-sidecar-x86_64-pc-windows-msvc.exe
# 2) Build / run the desktop shell
npm install -g @tauri-apps/cli
tauri dev # run in development
tauri build # output: src-tauri/target/release/bundle/nsis/*.exeThe desktop window loads its UI from the bundled frontend (http://tauri.localhost); the sidecar serves the API on 127.0.0.1:8788, reached by the frontend via CORS. The sidecar binds loopback only.
Android app (on-device fully automatic)
Build from source under mobile/ (flutter build apk --debug, verified to build), or grab the official APK from Releases.
- Install and open the app, switch to the Auto tab.
- Weights are not bundled (CC BY-NC-SA); the app guides an in-app download on first use (~300 MB, resumable, SHA-256 verified).
- Pick images and start; results can be shared / saved in-app. The Hint tab needs no weights and works fully offline.
Requirement: Android arm64 physical device. On-device inference uses the ONNX Runtime CPU execution provider (no GPU/NNAPI path yet).
📖 Getting started · Architecture & API · Development guide · Licensing
Repository code is Apache-2.0. Model weights are CC BY-NC-SA 4.0 — not distributed with the repo; downloaded by the app or the user separately (see NOTICE and docs/licensing.en.md).
黑白漫画（含网点）语义上色工具箱：全自动 ONNX 语义上色 + 交互式精修，提供 Windows 桌面、浏览器、Android 端侧 三种使用方式，核心引擎为纯 Dart 确定性实现。
- 全自动 — ONNX 语义上色（SAM 引导 + generator），发/肤/瞳/背景自动配色，桌面、浏览器与 Android 应用端侧均可用
- 色彩方案 — 标准 / 鲜明赛璐璐 / 柔和淡彩 / 厚涂浓郁 / 昭和复古；默认「标准」收尾修复纸白发黄与整体发灰（纸白中和 + 自适应色度）
- 提示点 — 在画布上点击落点，指定区域颜色精修
- 参考图 — 根据人设彩图对齐上色：模型语义色 ⊕ 参考图色调迁移（Reinhard / 直方图两种算法 + 强度，纸白保护），Lab 上 5:5 融合；桌面与 Android 同算法（Android 端需模型权重）
- 批量队列 — 一次最多 32 张，逐个处理，单张失败不阻断
- 长图 — >1024px 自动分块推理（overlap 256 + 线性羽化融合），整页漫画无下采样损失
- 亮度保真 — 所有模式只写 Lab a/b 色度，L 通道始终回写原稿（唯一例外：白发等被墨线包围的封闭近白区按色度小幅压低 L——纸白亮度处色度没有渲染空间）
- 白色前景救援 — 被墨线包围的白色区域（白发、白衣、画格内背景）保留模型配色；连通页边的纸白保持纯白
- GPU 自动加速 — Windows 优先 DirectML（任意 DX12 显卡），NVIDIA 可用 CUDA，无 GPU 自动回退 CPU
| 路径 | 用途 |
|---|---|
src-tauri/ |
Tauri 2 桌面壳（Windows）+ Rust 后端；打包内嵌 Python sidecar |
tool/colorizer_service/ |
Python（uvicorn）上色服务——桌面 sidecar 与浏览器工作台共用 |
web/ |
浏览器工作台前端（静态），由 Python 服务托管 |
mobile/ |
Flutter Android 应用——端侧 ONNX 上色 |
packages/manga_colorizer_core/ |
纯 Dart 上色引擎（提示点色度扩散 + 亮度染色） |
packages/manga_colorizer_server/ |
基于 Shelf 的 HTTP 后端，封装核心引擎 |
packages/manga_colorizer_cli/ |
命令行工具：批量上色、染色、取样 / 对比图 |
models/ |
ONNX 权重（gitignore；运行时下载） |
Windows 安装包（推荐普通用户）
- 从 Releases 下载
Manga Colorizer_<版本>_x64-setup.exe并安装。 - 首次启动会下载模型权重（约 287 MB，一次性，下载到
%LOCALAPPDATA%\manga-colorizer\weights）。 - 拖入漫画页 → 勾选免责声明 → 开始上色。成品在「图库」页。
系统要求：Windows 10 1809+ / 11（WebView2 由安装包自动装）。GPU 加速需 DX12 显卡（DirectML）；无 GPU 时自动使用 CPU。
Python 服务 + 浏览器工作台 要求 Python ≥ 3.10。
# 1) 依赖（CPU 基础路径）
python -m pip install -r tool/colorizer_service/requirements.txt
# Windows GPU（DirectML，任意 DX12 显卡）：二选一，不要与 CPU 版混装
python -m pip install -r tool/colorizer_service/requirements-directml.txt
# 2) 下载模型权重（约 300MB，不随仓库分发；国内可把域名换成 hf-mirror.com）
python -c "import urllib.request as u; [u.urlretrieve(f'https://huggingface.co/sharky172/manga-light-colorizer/resolve/main/models/{n}', f'models/manga-light-colorizer/{n}') for n in ['v6_generator.onnx','v6_sam_encoder.onnx']]"
# 3) 启动服务（绑定 127.0.0.1:8788）
python -m uvicorn tool.colorizer_service.service:app --port 8788
# 4) 浏览器打开 http://127.0.0.1:8788/NVIDIA CUDA 路径：pip install -r tool/colorizer_service/requirements-gpu.txt 并自备匹配的 CUDA/cuDNN 运行库，以 COLORIZER_DEVICE=cuda 启动；CUDA 不可用时服务明确报错，不会静默回退。
从源码构建桌面应用 要求：Rust stable、Node.js（仅用于安装 Tauri CLI）、Python ≥ 3.10。
# 1) 构建 Python sidecar（onefile exe；GPU 路径建议 Python 3.10–3.12）
python -m venv .venv
.venv/Scripts/pip install -r tool/colorizer_service/requirements-directml.txt pyinstaller
.venv/Scripts/pyinstaller sidecar/manga-colorizer-sidecar.spec --noconfirm \
--distpath build/sidecar-dist --workpath build/sidecar-work
cp build/sidecar-dist/manga-colorizer-sidecar.exe \
src-tauri/binaries/manga-colorizer-sidecar-x86_64-pc-windows-msvc.exe
# 2) 构建/调试桌面壳
npm install -g @tauri-apps/cli
tauri dev # 开发运行
tauri build # 产物：src-tauri/target/release/bundle/nsis/*.exe桌面窗口固定从打包内前端（http://tauri.localhost）加载界面，sidecar 在 127.0.0.1:8788 提供 API，前端通过 CORS 跨源访问。sidecar 仅绑定回环地址。
Android 应用（端侧全自动）
在 mobile/ 下用 Flutter 从源码构建（flutter build apk --debug，构建验证通过），或等 Releases 的正式 APK。
- 安装并打开应用，切到「全自动」页签。
- 权重不随 APK 分发（CC BY-NC-SA），首次使用由应用内引导下载（约 300 MB，支持断点续传与 SHA-256 校验）。
- 选图开始全自动上色，结果可在应用内分享/保存。「提示点上色」页签无需权重、完全离线可用。
设备要求：Android arm64 物理设备。端侧推理走 ONNX Runtime CPU 执行提供者（暂无 GPU/NNAPI 路径）。
📖 快速开始 · 架构与 API · 开发指南 · 许可与发行边界
仓库代码 Apache-2.0；模型权重 CC BY-NC-SA 4.0 （不随仓库分发，由应用或使用者自行下载，见 NOTICE 与 docs/licensing.zh.md）。