A content filter that knows its anatomy

What it does

NudeNet is a pre-trained image detector that spots nudity and partial nudity across 18 anatomical categories — from exposed genitalia to covered armpits. It wraps YOLOv8 models in a Python library that includes a built-in censor helper, and the project maintains a browser demo that runs inference locally via ONNX without uploading images. A Dockerized REST endpoint is also provided for server-side use.

The interesting bit

Most safety classifiers reduce an image to a single “unsafe” score; NudeNet instead returns bounding boxes for specific body parts, letting you censor only what you need to. That granularity is rare in open-source moderation tools.

Key highlights

• Built on Ultralytics YOLOv8n and YOLOv8m, with both 320×320 and 640×640 resolution checkpoints available in ONNX and PyTorch formats.
• The lightweight 320n model ships inside the Python package by default; the heavier 640m model can be swapped in for higher-resolution work.
• Detects 18 distinct labels, distinguishing between exposed and covered variants of anatomy, faces, feet, and armpits.
• Includes a browser demo that processes images client-side, plus a Docker container exposing a REST API for backend integration.

Caveats

• The sole maintainer admits they have become busy and are actively seeking replacements, so the project’s future velocity depends on new contributors.

Verdict

Worth evaluating if you need self-hosted content moderation with granular control over what gets flagged and why. Skip it if you require a mature, actively maintained ecosystem or enterprise SLAs.