360 Secure SkillHub
360 安全技能中心

AI Watermark Margin

提供两个 Pillow 脚本：add_margin.py 在图片边缘加入留白或黑边，把生成器水印引到内容外；crop_margin.py 检测并裁掉纯色边，从而去除水印。仅适用于自己生成的 AI 图片，不处理内容内水印或第三方版权水印。

分类：效率
catsmansion
catsmansion
热度 3.03
最近更新 2026-10-30 09:33:46

安装

发给 Agent
帮我安装这个技能：npx secure-skills-cli add catsmansion/ai-watermark-margin
兼容性
通用兼容
联网要求
不联网
数据出网
数据不出本机
Metadata
name
ai-watermark-margin
description
Drop the in-image watermark/signature that an AI image editor stamps on YOUR OWN ai-generated or ai-edited pictures, deterministically — by adding a blank margin before the edit and cropping it off after. No inpainting, no ML model, no third-party copyright watermark removal. Use when: 'AI edit leaves a watermark', 'remove AI signature', 'AI 修图水印', '去AI生成签名', '图片下方有水印', 'generator watermark'. Two tiny Pillow scripts: add_margin.py pads the canvas with whitespace (or black) so the generator's mark lands outside the content; crop_margin.py detects that uniform border and crops it away. Includes an ethics/compliance note: this only removes the generator's own visual stamp; you still disclose AI use per platform rules.
metadata
{"displayName":"AI Watermark Margin Drop (AI水印留白裁剪)","slug":"ai-watermark-margin","category":"image-processing","agent_created":"true","description_zh":"确定性去掉 AI 修图/生成器在你【自己】的 AI 图片上盖的那行水印/签名——做法是修图前先加留白边，修完再把边裁掉。不靠 inpainting、不调模型、不去别人的版权水印。触发：AI 修图带水印、去AI生成签名、图片下方有水印、generator watermark。两个极简 Pillow 脚本：add_margin.py 给画布加留白（或黑边）让生成器把水印盖在边外；crop_margin.py 检测那圈纯色边并裁掉。含合规说明：本技法只去生成器自带的视觉签名，AI 使用声明你照常单独标。"}

AI Watermark Margin Drop

What this solves

AI image editors (image-to-image / inpainting tools, online generators) often stamp a small watermark, signature, or "made with X" text on the output — usually along an edge (most commonly the bottom). It looks cheap and you'd rather not have it baked into the file, even though you still disclose AI use separately.

This is a self-developed, deterministic technique (no model, no inpainting, no copyright-watermark removal):

Pad the canvas with a blank margin before sending the image to the AI; the generator drops its watermark into that margin; crop the margin off after the edit → the watermark is gone with zero reconstruction artifacts.

Workflow (3 steps)

  1. Pad. Run add_margin.py on your source image to add a ~10% blank (white) margin, typically at the bottom (where watermarks land). This is the image you hand to the AI.
  2. Edit. Send the padded image to the AI editor / generator. Its watermark/signature lands in the blank margin instead of on your content.
  3. Crop. Run crop_margin.py on the AI output. It auto-detects the uniform blank border and crops it away, taking the watermark with it.

Scripts

add_margin.py (the "pad" step)

python scripts/add_margin.py input.png output.png --percent 10 --side bottom --color white
  • --percent N : margin thickness as % of the relevant dimension (default 10).
  • --side : bottom (default), top, uniform, left, right.
  • --color : white (留白, default) or black (黑边). Hex like #000000 also accepted.

crop_margin.py (the "drop" step)

python scripts/crop_margin.py ai_output.png cleaned.png
# dry-run to see detected margins without writing:
python scripts/crop_margin.py ai_output.png --report
  • Auto-detects the uniform-color border (the margin you added) and crops all four sides.
  • --tol N : color tolerance for "is this border?" (default 12).
  • --min-frac F : fraction of a row/column that must be border color (default 0.90).
  • --max-percent P: safety cap — never crop more than P% per side (default 30).
  • --fallback N : if detected border < N% per side, crop a fixed N% bottom margin instead.
  • --report : print detected left/top/right/bottom crops, write nothing.

Ethics & compliance

  • ✅ This removes only the generator's own visual watermark/signature from images you generated or edited with AI.
  • ✅ It is not a tool for removing third-party copyright watermarks from stock/licensed media — don't do that.
  • ⚠️ You remain responsible for disclosing AI use per each platform's rules. This technique only cleans the in-file stamp for aesthetics; keep your "AI-generated" labeling separate and compliant.

When NOT to use

  • The watermark is baked into the content (not in a blank margin) → padding won't help; use a different method.
  • You need to keep the margin for composition → don't crop.
  • The image is someone else's copyrighted work with a rights watermark → do not use.

Tested backfires (don't retry)

  • Adding too little margin (<5%) lets the watermark overlap content → crop can't remove it cleanly. Use ≥8–10%.
  • Some generators recolor or fill the margin with content; if crop_margin.py reports ~0 border, bump --fallback or re-pad with a larger percent.