nsfw-classification

NSFW Image Classifier

A CLIP-based image classification API that categorizes images into three classes: Safe, NSFW Mild, and NSFW Explicit.

Model Architecture

• Backbone: OpenAI CLIP ViT-Large/14 (openai/clip-vit-large-patch14)
• Classifier Head: Linear(1024, hidden_dim=1024) → ReLU → Dropout(0.5) → Linear(1024, 3)
• Training Strategy: Two-phase fine-tuning (frozen backbone → full fine-tuning)
• Dataset: 100K+ images with balanced class distribution
• Validation Accuracy: ~97%

API Endpoints

Health Check

GET /health

Returns model status and configuration.

Single Prediction

POST /predict
Content-Type: multipart/form-data

Form field: image (file)

Response:

{
  "filename": "image.jpg",
  "predicted_class": "Safe",
  "predicted_index": 0,
  "is_nsfw": false,
  "confidence": 0.95,
  "all_scores": {
    "Safe": 0.95,
    "NSFW Mild": 0.03,
    "NSFW Explicit": 0.02
  },
  "status": "success"
}

Batch Prediction

POST /predict/batch
Content-Type: multipart/form-data

Form field: images (multiple files)

Preprocessing

Images are preprocessed with CLIP-specific normalization:

• Resize to 224x224 (bicubic interpolation)
• Center crop
• Normalize with mean [0.481, 0.458, 0.408] and std [0.269, 0.261, 0.276]

Local Development

# Install dependencies
pip install -r requirements.txt

# Run the API
python app.py

The API runs on port 7860 by default. Override with the PORT environment variable.

Docker Deployment

docker build -t nsfw-classifier .
docker run -p 7860:7860 nsfw-classifier

The model is automatically downloaded from Hugging Face Hub on startup.

Example Usage

import requests

with open("image.jpg", "rb") as f:
    response = requests.post(
        "http://localhost:7860/predict",
        files={"image": f}
    )
print(response.json())

Configuration Reference

See Hugging Face Spaces Config Reference for deployment options.