metaclean-vlm 0.1.0

Remove metadata from image datasets before VLM training or ingestion to reduce hidden prompt injection risk.

metaclean-vlm

Remove metadata from image datasets before VLM training or ingestion to reduce hidden prompt injection risk.

Installation

pip install metaclean-vlm

Usage

from metaclean_vlm import clean_images

report = clean_images("raw_images", "clean_images")
print(report)

Bulk clean a dataset

from metaclean_vlm import clean_dataset

report = clean_dataset("dataset/raw", "dataset/clean")
print(report)

clean_dataset recursively processes supported images and preserves folder structure.

Output

[
    {
        "input": "raw_images/example.jpg",
        "output": "clean_images/example.jpg",
        "format": "JPEG",
        "metadata_found": True,
        "metadata_keys": ["exif", "icc_profile"],
        "ok": True,
        "error": None,
    }
]

Clean one image

from metaclean_vlm import clean_image

result = clean_image("image.jpg", "image.clean.jpg")
print(result)

Inspect metadata

from metaclean_vlm import inspect_metadata

metadata = inspect_metadata("image.jpg")
print(metadata)

Overview

metaclean-vlm is a tiny Python utility for removing metadata from image datasets before they are used in VLM pipelines. It is aimed at reducing the risk of hidden prompt injection or unwanted instructions stored in image metadata.

It is useful when building:

• VLM training datasets
• multimodal AI pipelines
• image ingestion systems
• dataset cleaning workflows
• AI safety preprocessing tools

Features

• Removes common image metadata
• Cleans batches and full image datasets
• Supports JPEG, PNG, WEBP, TIFF, and BMP
• Preserves folder structure
• Returns a simple cleaning report
• Uses Pillow
• Simple API

Limitations

metaclean-vlm removes metadata by re-encoding image pixels without metadata fields. It is not a complete security scanner and does not protect against steganography, visible prompt injection, OCR-based attacks, adversarial images, poisoned pixels, or malicious image content. Use it as one dataset hygiene layer, not as your only security control.

Issues

Report issues at: https://github.com/edujbarrios/metaclean-vlm

Author

Eduardo J. Barrios
[e-mail]

License

Mozilla Public License 2.0