You can deploy this model and start earning money today!

necrosyth/
nsfw-detection-with-yolo

cc-by-4.0

not-for-all-audiences

Views

No views yet

NSFW-YOLO: Real-Time NSFW Content Detection

Model Description

A fine-tuned YOLOv8 model for real-time detection and localization of NSFW (Not Safe For Work) content in images and video frames. Unlike classification-only approaches, this model outputs bounding boxes — identifying where in an image the flagged content appears, not just whether it exists. This enables granular content moderation, selective blurring, and region-level logging.

Built for deployment in content platforms, upload pipelines, and real-time video streams.

Architecture

Backbone: YOLOv8-Medium (YOLOv8m)

Base: Ultralytics YOLOv8m pretrained on COCO
Neck: PANet (Path Aggregation Network) for multi-scale feature fusion
Head: Decoupled detection head with 3 output scales
(80×80, 40×40, 20×20) for small, medium, and large region detection
Input resolution: 640×640 (dynamic tiling for high-res images)
Parameters: ~25.9M

Why YOLOv8m specifically?

YOLOv8n/s are too lightweight for nuanced body-region detection
YOLOv8l/x offer marginal gains at 2× the inference cost
YOLOv8m hits the sweet spot: 45+ FPS on a T4 GPU

Detection Classes

Class ID	Label	Description
0	`safe`	No sensitive content
1	`suggestive`	Revealing but not explicit (cleavage, underwear, etc.)
2	`explicit_partial`	Partial nudity (topless, etc.)
3	`explicit_full`	Full nudity
4	`sexual_act`	Depiction of sexual activity
5	`text_overlay`	NSFW text rendered on image

Multi-label per image is supported — a single image can contain both suggestive and text_overlay regions simultaneously.

necrosyth/nsfw-detection-with-yolo