# wiltodelta/remove-ai-watermarks [Health: Active]

**Category:** 🎥 Multimedia Process  
**Repository:** https://github.com/wiltodelta/remove-ai-watermarks  
**GitHub Stars:** 6144  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/wiltodelta-remove-ai-watermarks

## Description
Identify and remove AI watermarks and provenance metadata (Gemini and Veo marks, SynthID, C2PA, EXIF) in images and video you generated yourself, plus provenance metadata in audio. Runs as uvx remove-ai-watermarks mcp.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "remove-ai-watermarks": {
    "command": "uvx",
    "args": ["remove-ai-watermarks"]
  }
}
```

## Documentation

## What wiltodelta/remove-ai-watermarks does

wiltodelta/remove-ai-watermarks provides MCP access to local media provenance and cleanup operations. Its scope includes images, video, and provenance metadata in audio. Image processing can inspect signals, classify some image sources, remove registered visible marks, erase a selected region, strip metadata, and regenerate pixels to disrupt supported invisible watermarks.

Video processing can inspect supported provenance, remove visible marks and AI metadata, process directories, and apply a calibrated profile for video SynthID removal. The package also handles metadata formats and container structures used by MP4, MOV, M4V, WebM, MKV, AVI, and FLV files.

The project is intended for lawful use on content the operator owns or generated. It is not positioned as a tool for removing watermarks protecting third-party paid content.

## How it works

Run the MCP entry point with:

```bash
uvx remove-ai-watermarks mcp
```

The underlying operations are organized around commands such as `identify`, `visible`, `erase`, `metadata`, `all`, and `batch`. Video adds corresponding `video identify`, `video all`, `video metadata`, `video visible`, `video invisible`, and `video batch` paths.

Metadata-only processing can inspect or remove C2PA, EXIF, XMP, IPTC, and related AI provenance fields. Image commands may overwrite the source when no output path is supplied, while video metadata removal writes a `_clean` output when no destination is given. Video metadata cleanup does not transcode video or audio streams; formats outside the MP4/MOV family are remuxed with stream copy for removal.

Invisible image removal uses diffusion regeneration and requires CUDA. Video SynthID removal uses a separate oracle-certified profile and is opt-in. Microsoft Paint and Photos InvisMark declarations are routed to pixel regeneration, while the `all` image path also removes a linked C2PA manifest when applicable.

## Setup and configuration

The default package installation supports metadata inspection and stripping:

```bash
uv tool install remove-ai-watermarks
```

Install feature extras for additional processing:

- `remove-ai-watermarks[visible]` for visible image marks and metadata
- `remove-ai-watermarks[video]` for visible video processing
- `remove-ai-watermarks[video,diffusion]` for video SynthID removal
- `remove-ai-watermarks[qwen-zimage]` for invisible image removal with CUDA
- `remove-ai-watermarks[all]` for the production feature set

Other extras cover image classification, source-export classification, torch-free detection, direct pixel disruption, HEIF, and specialized removal models. The MCP command can then be started with `uvx`, allowing an MCP client to launch it as a local process.

## Tools and capabilities

- Identify image and supported video provenance signals
- Remove visible image and registered video marks
- Strip C2PA, EXIF, XMP, IPTC, and related metadata
- Process image or video directories in batch modes
- Regenerate image pixels to disrupt selected invisible watermarks
- Apply calibrated video processing for SynthID removal
- Erase a user-selected image region
- Classify photographs or source exports through optional packages

## Limitations and notes

Invisible image removal requires a CUDA environment and the relevant optional dependencies. Video invisible-watermark removal is lossy and excluded from the default `video all` path unless explicitly enabled. There is no local SynthID pixel detector in the package; signed provenance is the supported identification route.

Classification is separate from provenance identification. The photograph classifier is an opt-in pixel-based AI-versus-camera signal, while source classification provides an abstaining OpenAI, Google, or unknown export signal after metadata removal.

The project is supplied for lawful use on owned content, and legal restrictions on removing AI labels or providing such tools may apply in some jurisdictions.

_Full upstream README: https://allmcps.com/mcp/wiltodelta-remove-ai-watermarks/readme_