A mobile-friendly visual content moderation model, based on the work of F. C. Akyon obtained by fine-tuning EfficientNet-b4 on nsfw images to detect nudity and sexual content in images or video frames with high accuracy.

For a Demo, see: [viddexa/moderators]

Performance Comparison

Image detection performance of this model on 5 classes described above compared with other models trained by our team provided below with average and per-class F-scores.

Model	F_macro	F_safe	F_porn	F_hentai	F_drawing	F_sexy	Params
nsfw-detection-2-nano	93.00%	96.82%	96.34%	93.43%	93.24%	85.15%	4M
nsfw-detection-2-mini	96.09%	98.59%	98.05%	96.06%	96.83%	90.92%	17M
nsfw-detection-1-mini	N/A	97.90%	N/A	N/A	N/A	N/A	17M
Azure AI	N/A	96.79%	N/A	N/A	N/A	N/A	N/A
Falconsai	N/A	89.52%	N/A	N/A	N/A	N/A	85M

nsfw-detection-2-mini and nsfw-detection-2-nano supports 5 classes: Normal, Hentai, Porn, Sexy, Drawing

Normal: Photos without nsfw content.
Porn: Photos with pornographic content.
Hentai: Drawing with sexual content.
Drawing: Comics, cartoons or drawings with no nsfw content.
Sexy: Photos with no explicit nudity but might contain some characteristics of nsfw content such as suggestive nudity or risqué clothing.

Usage with moderators library

Install with pip install moderators then run:

python

from moderators import AutoModerator

model = AutoModerator.from_pretrained("viddexa/nsfw-detection-2-mini")
results = model("<path-to-image-file>")

probs = {k: v for r in results for k, v in r.classifications.items()}
predicted_label = max(probs, key=probs.get) 

Usage with transformers library

Install with pip install transformers then run:

python

from transformers import (
    AutoImageProcessor,
    AutoModelForImageClassification)
from PIL import Image
import torch

img = Image.open("<path-to-image-file>")

processor = AutoImageProcessor.from_pretrained("viddexa/nsfw-detection-2-mini", use_fast = False)
model = AutoModelForImageClassification.from_pretrained("viddexa/nsfw-detection-2-mini")

with torch.no_grad():
    inputs = processor(images=img, return_tensors="pt")
    outputs = model(**inputs)
    logits = outputs.logits
    probs = torch.softmax(logits, dim=-1)
    pred_id = int(probs.argmax())
    print(model.config.id2label[pred_id])

Model Details

5-class nudity and sexual content classification model that classifies images as Normal, Porn, Hentai, Drawing or Sexy. Model is obtained by fine-tuning google's Efficientnet-b4.

Developed by: [Kerem Bozgan, Abdullah Kırman, Fatih Çağatay Akyön]
License: [apache-2.0]
Finetuned from model: [google/efficientnet-b4]

Intended Use and Limitations

This model is designed to detect explicit nudity alone. For instance, an image containing suggestive nudity or risqué clothing is labeled as safe if there is no explicit nudity in the image.

Appropriate Use Cases:

Filtering explicit images in user-generated content platforms
Assisting moderation workflows for social apps and community forums
Parental control or restricted content filtering
Content review pipelines for games, and streaming services

In these settings, the model acts as an initial screening tool to help reduce moderator workload. A human review step is recommended for final decisions.

This model is not intended for:

Moral or ethical judgment of individuals or their bodies
Surveillance or monitoring of people in real-world environments
Determining sexual intent
Legal classification of pornography vs. art
Use in contexts affecting employment, policing, social credit, or identity

BibTeX entry and citation info

bibtex

@article{akyon2023nudity,
  title={State-of-the-art in nudity classification: A comparative analysis},
  author={Akyon, Fatih Cagatay and Temizel, Alptekin},
  booktitle={2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW)},
  pages={1--5},
  year={2023},
  organization={IEEE}
}

Model Card Authors

Kerem Bozgan & Fatih Akyon