newd

Overview

newd analyzes images and identifies specific NSFW body parts with high accuracy. It can also optionally censor detected areas.

Installation

pip install newd

Usage

Basic Detection

from newd import detect

# Standard detection with default settings
results = detect('path/to/image.jpg')
print(results)

Advanced Options

# Faster detection with slightly reduced accuracy
results = detect('image.jpg', mode="fast")

# Adjust detection sensitivity
results = detect('image.jpg', min_prob=0.3)  # Lower threshold catches more potential matches

# Combine options
results = detect('image.jpg', mode="fast", min_prob=0.3)

Compatible Input Types

The detect() function accepts:

• String file paths
• Images loaded with OpenCV (cv2)
• Images loaded with PIL/Pillow

Output Format

Detection results are returned as a list of dictionaries:

[
  {
    'box': [x1, y1, x2, y2],  # Bounding box coordinates (top-left, bottom-right)
    'score': 0.825,           # Confidence score (0-1)
    'label': 'EXPOSED_BREAST_F'  # Classification label
  },
  # Additional detections...
]

First-Time Use

When importing newd for the first time, it will download a 139MB model file to your home directory (~/.newd/). This happens only once.

Performance Notes

• Standard mode: Best accuracy, normal processing speed
• Fast mode: ~3x faster processing with slightly reduced accuracy

Censoring / Redacting Detected Regions

newd.censor() masks detected NSFW regions with solid black rectangles. Use it when you need to create a safe-for-work version of an image.

from newd import censor

# Censor all detected areas and write the result
censored_img = censor(
    'image.jpg',
    out_path='image_censored.jpg'  # file will be written to disk
)

# Only censor specific labels (e.g. exposed anus & male genitals)
selected_parts = ['EXPOSED_ANUS_F', 'EXPOSED_GENITALIA_M']
censored_img = censor(
    'image.jpg',
    out_path='image_censored.jpg',
    parts_to_blur=selected_parts
)

Function parameters:

Parameter	Type	Description
img_path	str / Path	Source image or path.
out_path	str / Path, optional	Destination path; if omitted you can still obtain the result via the return value when visualize=True.
visualize	bool, default False	If True, the censored numpy.ndarray image is returned for display (cv2.imshow, etc.).
parts_to_blur	List[str], optional	Restrict censoring to given label names. When empty, all detected labels are censored.

If neither out_path nor visualize=True is supplied, the function exits early because there is nowhere to deliver the censored image.