Graphics Tutorial: Deploy AI Watermark Cleaning Service on-premises to keep your documents and images clean图文教程:本地部署 AI 水印清理服务,让你的文档和图片干干净净
Then do it again: install Python and source code, → start the local cleaning service, install the skills → in Claude Code, and → use an AI-generated diagram and an AI-written document to check and clean up. It runs locally all the time, and the files don't leave your computer.
跟着做一遍:装好 Python 和源码 → 启动本地清理服务 → 在 Claude Code 里装技能 → 用一张 AI 生成图和一个 AI 写的文档实测检查与清理。全程本地运行,文件不出你的电脑。

Step 1: Prepare the environment第一步:准备环境
Python 3.10 or above is required (Windows users directly use the py command). The project uses only the standard library and does not require pip to install any dependencies. This is particularly friendly to older machines.
Download the source code package of this page and unzip it to enter the directory. You can also git clone https://github.com/guillaumemeyer/watermarks-remover.git to track updates.
需要 Python 3.10 以上(Windows 用户直接用 py 命令),项目只用标准库,不需要 pip 装任何依赖,这一点对老机器格外友好。
下载本页的源码包并解压,进入目录。想跟踪更新也可以 git clone https://github.com/guillaumemeyer/watermarks-remover.git。

Step 2: Start the local service第二步:启动本地服务
macOS/Linux runs make serve; Windows executes py service/scripts/server.py --host 127.0.0.1 --port 8765 without make.
The service is implemented only by the standard library, and the startup is on the second level. By default, 127.0.0.1: 8765 is monitored, and only native connections are accepted; if the port is changed, the watermarks_service_URL environment variable is set to tell the skill where to find it.
Verification: The browser opens http://127.0.0.1: 8765/health and returns normal as-a-service ready.
macOS/Linux 跑 make serve;Windows 没有 make 就执行 py service/scripts/server.py --host 127.0.0.1 --port 8765。
服务只有标准库实现,启动是秒级的。默认监听 127.0.0.1:8765,只接受本机连接;换端口就设 WATERMARKS_SERVICE_URL 环境变量告诉技能去哪找它。
验证:浏览器打开 http://127.0.0.1:8765/health,返回正常即服务就绪。

Step 3: Load your skills into your agent第三步:把技能装进你的代理
Claude Code users are the most convenient:/plugin marketplace add guillaumemeyer/watermarks-remover, then/plugin install watermarks-remover @ watermarks-remover, and later/plugin update in place.
Generic installer for other hosts (Codex, Cursor, Cowork): py install_skill.py --skill remove-ai-marks --target corresponds to the name, --list can see all supported targets.
The advanced gameplay is to turn on the automatic hook: set the watermarks_hook_mode to clean, and the agent will automatically clean the file every time it writes; keep the default check and report only no action, which is more stable.
Claude Code 用户最省事:/plugin marketplace add guillaumemeyer/watermarks-remover,然后 /plugin install watermarks-remover@watermarks-remover,以后 /plugin update 原地升级。
其他宿主(Codex、Cursor、Cowork)用通用安装器:py install_skill.py --skill remove-ai-marks --target 对应名字,--list 能看全部支持的目标。
进阶玩法是开自动钩子:把 WATERMARKS_HOOK_MODE 设为 clean,代理每次写文件都会自动清一遍;保持默认 check 则只报告不动手,更稳。

Step 4: Measuring a round of "inspection and → cleaning"第四步:实测一轮「检查 → 清理」
Documents written to AI: Say "Check draft.md for AI traceability marks" in Claude Code. The skill will adjust/inspect to list the suspicions of invisible Unicode and statistical watermarks. After confirmation, say "clean up" to generate a clean version.
Generate a diagram for AI: inspect_file.py photo.png under py service/scripts Check C2PA/EXIF/XMP first, clean_file.py photo.png -o photo.cleaned.png Output a striped copy.
Just want to clear the text, lazy running service: install the clean-user-facing-text skill (installer with the same name option), it comes with text cleaning logic, no service required.
对 AI 写的文档:在 Claude Code 里说「检查 draft.md 有没有 AI 溯源标记」,技能会调 /inspect 列出隐形 Unicode 和统计水印的嫌疑,确认后说「清理」即可生成干净版本。
对 AI 生成图:py service/scripts 下的 inspect_file.py photo.png 先查 C2PA/EXIF/XMP,clean_file.py photo.png -o photo.cleaned.png 输出剥净的副本。
只想清文本、懒得跑服务:装 clean-user-facing-text 技能(安装器同名选项),它自带文本清理逻辑,无需服务。

Problem Encountered遇到问题
Skill report connection failure: first confirm that the service is running and the port is not occupied; clean up: rewrite the statistical watermark-dependent model and try again with a stronger model. Please only process content that you own the copyright to. See the download button above for the companion source code package for this article.
技能报连接失败:先确认服务在跑、端口没被占用;清理不干净:统计型水印依赖模型重写,换更强的模型重试。请只处理你拥有版权的内容。本文配套源码包见上方下载按钮。
想马上用起来?去「AI工具」栏挑一个直接下载,或在「AI教程」里跟着图文步骤做一遍。
去 AI工具 → 看 AI教程 →阅读与点赞数据保存在你的浏览器本地,欢迎留下你的想法。
今日正能量学一点,用一点;今天种下的种子,会长成明天的能力。去免费下载专区 →广告
评论 文明发言,让讨论更有价值