Back to list
Comprehensive Open-Source Tool Remove-AI-Watermarks Enables Stripping of Visible and Invisible Identifiers from Major AI Models
Open SourceAI WatermarkingImage ProcessingGenerative AI

Comprehensive Open-Source Tool Remove-AI-Watermarks Enables Stripping of Visible and Invisible Identifiers from Major AI Models

A new open-source project titled "Remove-AI-Watermarks" has been released, providing a unified solution for removing both visible and invisible AI-generated markers from images. The tool targets outputs from industry-leading models including Google Gemini, OpenAI's DALL-E 3, Stable Diffusion, Adobe Firefly, and Midjourney. By utilizing techniques such as reverse alpha blending and diffusion-based regeneration, the library can eliminate SynthID, StableSignature, and TreeRing watermarks. Furthermore, it strips C2PA Content Credentials and EXIF/XMP metadata that trigger "Made with AI" labels on social media platforms like Instagram and Facebook. The project also introduces features like an "Analog Humanizer" and "Smart Face Protection" to maintain image quality while bypassing AI classifiers, offering both a command-line interface and a web-based service at raiw.cc.

Hacker News
Share

Key Takeaways

  • Multi-Model Support: The tool effectively removes watermarks and metadata from Google Gemini (including Nano Banana and Gemini 3 Pro), ChatGPT/DALL-E 3, Stable Diffusion, Adobe Firefly, and Midjourney.
  • Visible and Invisible Removal: Employs reverse alpha blending for visible logos (like Gemini's sparkle) and diffusion-based regeneration for invisible watermarks such as SynthID and StableSignature.
  • Metadata Stripping: Targets C2PA provenance manifests, EXIF/XMP labels, and PNG text chunks to prevent automated "Made with AI" labeling on social media platforms.
  • Advanced Image Preservation: Features "Smart Face Protection" to prevent distortion during the AI cleaning process and an "Analog Humanizer" to bypass AI image classifiers.
  • Accessibility: Offers a batch processing library for local use and a free online web service via raiw.cc.

In-Depth Analysis

Technical Mechanisms for Watermark Erasure

The "Remove-AI-Watermarks" project introduces a multi-tiered approach to cleaning AI-generated imagery, categorized by the type of watermark encountered. For visible watermarks, such as the Google Gemini "sparkle" logo, the tool utilizes a fast, offline, and deterministic method known as reverse alpha blending. This technique effectively reverses the transparency-based overlay to restore the original pixel data beneath the logo.

For more sophisticated, invisible watermarks—including Google's SynthID (versions 1 and 2), StableSignature, and TreeRing—the tool employs diffusion-based regeneration. This process involves using AI diffusion models to subtly re-process the watermarked areas, effectively breaking the steganographic patterns or Discrete Wavelet Transform (DWT) markers that allow detectors to identify AI origin. The tool includes a three-stage Normalized Cross-Correlation (NCC) watermark detection system with confidence scoring to identify these hidden markers before removal.

Metadata and the Battle for Content Provenance

A significant portion of the tool's functionality is dedicated to stripping metadata that triggers automated labeling on social media platforms. Platforms such as Instagram, Facebook, and X (Twitter) increasingly rely on C2PA (Coalition for Content Provenance and Authenticity) manifests and specific XMP DigitalSourceType labels like "Made with AI" to inform users about the nature of the content.

Remove-AI-Watermarks targets a wide array of metadata formats, including EXIF, PNG text chunks, and C2PA provenance manifests across various file types such as PNG, JPEG, AVIF, HEIF, and JPEG-XL. By stripping these manifests and labels, the tool allows users to bypass the automated classification systems implemented by major tech companies. This includes the removal of prompt data, model versions, and seed information typically embedded in Midjourney or Stable Diffusion outputs.

Specialized Processing: Humanization and Face Protection

Beyond simple watermark removal, the library addresses the challenges of image quality and classifier detection. The "Analog Humanizer" feature adds film grain and chromatic aberration to images. This is specifically designed to bypass AI image classifiers that look for the characteristic smoothness or mathematical perfection of synthetic images, making them appear more like traditional analog photography.

To address the potential for AI-based regeneration to distort human features, the tool includes "Smart Face Protection." This feature automatically extracts human faces from the image and blends them back in after the watermark removal process. This ensures that while the background or metadata is cleaned of AI markers, the integrity and recognizability of human subjects are preserved, preventing the common "AI distortion" effect.

Industry Impact

The release of "Remove-AI-Watermarks" represents a significant development in the ongoing arms race between AI watermarking technologies and tools designed to circumvent them. As major AI providers like Google, OpenAI, and Adobe implement C2PA and SynthID to comply with safety standards and transparency goals, the availability of a "one-command" removal tool challenges the effectiveness of these industry-wide provenance standards.

This tool highlights the technical difficulty of maintaining permanent AI identifiers. By combining metadata stripping with pixel-level regeneration, it provides a blueprint for how AI-generated content can be decoupled from its origin. For the AI industry, this underscores the limitations of current watermarking techniques and may drive the development of more robust, tamper-resistant identification methods. Furthermore, it places additional pressure on social media platforms that rely on these markers for content moderation and user transparency.

Frequently Asked Questions

Question: Which AI models are currently supported by this tool?

The tool supports a wide range of models, including Google Gemini (Nano Banana and Gemini 3 Pro), OpenAI's DALL-E 3 and ChatGPT (including gpt-image-2), Stable Diffusion (via AUTOMATIC1111 and ComfyUI), Adobe Firefly, and Midjourney.

Question: How does the tool handle invisible watermarks like SynthID?

It uses a diffusion-based regeneration process to overwrite the imperceptible pixel patterns used by SynthID, StableSignature, and TreeRing. It also includes a detection system to identify these watermarks before the removal process begins.

Question: Can this tool prevent social media platforms from labeling my images as AI-generated?

Yes, the tool is specifically designed to strip C2PA manifests, EXIF/XMP "Made with AI" labels, and other metadata that trigger automated AI labeling on platforms like Instagram, Facebook, and X (Twitter).

Read original article

Related News

Anthropic Releases Public Repository for Agent Skills to Enhance Claude's Capabilities
Open Source

Anthropic Releases Public Repository for Agent Skills to Enhance Claude's Capabilities

Anthropic has officially launched a public GitHub repository dedicated to 'Agent Skills,' specifically featuring implementations designed for its Claude AI model. This repository serves as a practical extension of the Agent Skills standard, a framework aimed at regularizing how artificial intelligence agents interact with tools and execute complex tasks. By providing these implementations openly, Anthropic is facilitating a more standardized approach to agentic AI development. The repository also directs developers to agentskills.io for comprehensive information regarding the overarching standards. This move highlights Anthropic's commitment to fostering an interoperable ecosystem where AI agents can perform specialized functions with greater consistency and transparency, marking a significant step in the evolution of autonomous AI tool usage.

Addy Osmani Introduces Agent-Skills: A Framework for Production-Grade Engineering in AI Coding Agents
Open Source

Addy Osmani Introduces Agent-Skills: A Framework for Production-Grade Engineering in AI Coding Agents

Renowned engineer Addy Osmani has released a new project titled 'agent-skills,' aimed at elevating AI coding agents to production-grade standards. The project focuses on encoding essential engineering workflows, quality gates, and industry best practices directly into the operational logic of AI agents. By providing a structured approach to how AI interacts with codebases, agent-skills seeks to ensure that autonomous agents do not just generate code, but adhere to the rigorous standards required in professional software development environments. This initiative marks a significant step toward making AI-driven development more reliable, maintainable, and integrated into existing high-stakes production pipelines, addressing the gap between experimental AI outputs and enterprise-level engineering requirements.

Agency-Agents: A Comprehensive Open-Source Framework for Specialized AI Agency Workflows
Open Source

Agency-Agents: A Comprehensive Open-Source Framework for Specialized AI Agency Workflows

Agency-agents, a new project by developer msitarzewski, introduces a structured approach to AI automation by providing a complete 'AI agency' framework. Unlike generic AI tools, this project features a suite of specialized agents, ranging from 'frontend wizards' to 'Reddit community ninjas.' Each agent is designed with a distinct personality, specific operational processes, and a focus on producing mature, professional-grade deliverables. By incorporating unique roles such as 'whim injectors' for creativity and 'reality checkers' for validation, the framework aims to bridge the gap between abstract AI generation and practical, industry-standard output. This development signals a shift toward more nuanced, role-based AI systems that prioritize specialized expertise over general-purpose assistance.