demo/README.md

AI-Generated Image Detector Demo

A simple Streamlit web application for detecting AI-generated images using the ARNIQA perceptual feature extractor.

Features

• Upload images (PNG, JPG, JPEG)
• Detect if image is real or AI-generated
• Display confidence scores and probability breakdown
• Simple, clean interface
• CPU-friendly (no GPU required)

Requirements

• Python 3.8+
• Trained ARNIQA classifier checkpoint at: checkpoints/GenImage/extensive/MarginContrastiveLoss_CrossEntropy/arniqa/best_model.ckpt

Setup

1. Install dependencies:

   cd demo
   pip install -r requirements.txt
   

2. Verify checkpoint exists: Make sure the trained model checkpoint is at the correct location relative to the project root:

   ../checkpoints/GenImage/extensive/MarginContrastiveLoss_CrossEntropy/arniqa/best_model.ckpt
   

Running the Demo

From the demo directory:

streamlit run app.py

The app will open in your browser at http://localhost:8501

How It Works

1. Feature Extraction: Uses ARNIQA (loaded via torch.hub) to extract perceptual features from images
2. Classification: A trained 2-layer MLP classifier processes the features
3. Prediction: Outputs probability of image being real vs AI-generated

Model Architecture

• Feature Extractor: ARNIQA encoder (4096-dim features)
• Classifier: 2-layer MLP (4096 → 1024 → 2)
• Training Dataset: GenImage dataset
• Threshold: 0.5 for binary classification

Troubleshooting

Model not loading

• Ensure checkpoint file exists at the specified path
• Check that you're running from the demo/ directory
• Verify Python path includes parent directory

Out of memory

• The app runs on CPU by default
• Close other applications to free up RAM
• Image is automatically resized to 224x224 for processing

ARNIQA not loading

• First run downloads ARNIQA model from torch.hub
• Requires internet connection for initial download
• Model is cached locally after first download

Deployment (Future)

This demo can be deployed to:

• Oracle Cloud (CPU instances)
• Streamlit Cloud
• Docker containers

Docker support coming soon!

Credits

Based on research in detecting AI-generated images using perceptual features from Image Quality Assessment models.