Remove-AI-Watermarks open source release: AI image watermark removal tool sparks security discussion

267 points HN hot post: The open source tool Remove-AI-Watermarks can remove visible and invisible watermarks from AI images such as Gemini and ChatGPT, and supports SynthID and C2PA metadata removal. It received 622 stars upon launch.

Core conclusion

On May 20, the open source tool Remove-AI-Watermarks was on the hot list on Hacker News with 267 points, and the GitHub Star exceeded 620+ that day. This Python command line tool can not only remove visible Sparkle watermarks from Gemini/Nano Banana, but also remove invisible frequency domain watermarks such as SynthID and C2PA/EXIF metadata tags. For AI content creators, this means:

Visible watermark removal: Google Gemini / Nano Banana’s Sparkle Logo, 0.05 seconds/image, no GPU required
Invisible watermark removal: SynthID v1/v2, StableSignature, TreeRing, regeneration through diffusion model
Metadata Cleanup: C2PA Content Credentials, EXIF "Made with AI" tags, PNG text chunks

Key Points

Release time: 2026-05-20 (HN hot list)
Stars for the day: 622 ⭐, 35 Forks
Core Positioning: Remove traceability marks from AI-generated images to maintain visual quality
Point of Controversy: SynthID’s 136-bit user ID cannot erase Google server records even if the watermark is cleared

Background: Two sides of AI watermark standardization

Just last week, OpenAI announced its adoption of Google’s SynthID watermarking system. AI content attribution is moving toward standardization—but it’s also giving rise to adversarial tools.

Remove-AI-Watermarks was created by developer wiltodelta, based on multiple open source projects (noai-watermark, GeminiWatermarkTool, CtrlRegen), and is open source under the MIT license. The tool achieves watermark removal through an inverse alpha blending algorithm (precisely extracting the Sparkle Logo overlay) and diffusion model regeneration technology (adding controlled noise and then denoising).

The tool clearly states that it does not support artist protective perturbations such as Nightshade/Glaze - "Removing these is an attack on the artist, not AI traceability".

Key Impact: Three Dimensions

Dimensions	Change	What it means to us	Recommended actions
AI traceability	SynthID watermarks can be removed programmatically	The reliability of the platform's "Made with AI" label has declined	Pay attention to subsequent platform detection strategy updates
Content security	AI images are less resistant to tampering	Enterprises need additional layers of content verification	Combining C2PA certificate chains and multi-dimensional inspection
Developer Ecosystem	622 ⭐ Confirms the market demand for countermeasures tools	Watermark technology will accelerate iteration	Pay attention to SynthID v3 and Server-side detection solutions
Legal Compliance	The COPIED Act has explicitly prohibited the removal of traceability information for deceptive purposes	Be cautious in commercial use	Clarify the purpose of use and only use it in reasonable circumstances

Technical principle: Three layers of watermarks are processed separately

1. Visible watermark (Gemini Sparkle Logo)

Google Gemini’s Sparkle Logo is overlaid with Alpha blending:

watermarked = α × logo + (1 − α) × original

Alpha ，

original = (watermarked − α × logo) / (1 − α)

NCC（），。。

2. （SynthID / StableSignature）

SynthID ，、 JPEG 。 SDXL

code

resize to 1024px → VAE → （）
→ （，50 0.05）→ VAE →

⚠️ ****SynthID v2 136-bit （/）。（ Gemini 、 Google ），Google 。。

3. （C2PA / EXIF / "Made with AI"）

EXIF prompt、seed、model hash
XMP DigitalSourceType（trainedAlgorithmicMedia， Instagram/Facebook/X AI ）
PNG text chunks（ComfyUI workflows、AUTOMATIC1111 ）
C2PA Content Credentials（）

4.

（YOLO），， AI 。 Analog Humanizer ，， AI 。

Remove-AI-Watermarks

****，——SynthID
**** C2PA/EXIF AI ，
**** SynthID
**** COPIED ，

terminal

#
pipx install git+https://github.com/wiltodelta/remove-ai-watermarks.git

# （ + + ）
remove-ai-watermarks all image.png -o clean.png

#
remove-ai-watermarks batch ./images/ --mode all

**** raiw.cc

（2026 5 ）


EU	AI Act 50(2)	2026.12.2
（）	COPIED		，
	2025		AI ，
	Online Safety Act		，

AI OpenAI、ChatGPT、Gemini、Stable Diffusion 。 Claude、DeepSeek content generated by the model.

Internal link guidance

Learn the principles behind SynthID watermarking technology: How to add quality gates to your AI automation workflow: A practical guide from output to trustworthy results
Real case: Independent developers use n8n+OpenClaw to build automated workflows and earn monthly income $5,000: Indie Developer: n8n + OpenClaw Automation Workflow Earning $5,000/mo