---
license: mit
tags:
  - image-classification
  - ai-detection
  - deepfake-detection
  - clip
datasets:
  - Rajarshi-Roy-research/Defactify_Image_Dataset
  - NasrinImp/COCO_AI
---

# VerifAI Model Weights

Pre-trained model weights for [VerifAI](https://github.com/canemirbora4/VerifAI) - AI-Generated Media Detector.

AI-Generated Media Detector — Detect AI-generated images and videos with calibrated confidence scores, localized evidence, and robustness evaluation.

## Files

| File | Size | Description |
|------|------|-------------|
| `modern_ai_detector.pt` | 2.0 MB | CLIP ViT-L/14 classification head |
| `frequency_classifier.joblib` | 774 KB | Gradient Boosting frequency classifier |

## Performance

| Dataset | Real Acc | Fake Acc | Overall |
|---------|----------|----------|---------|
| Defactify (SD3, SDXL, DALL-E 3, MidJourney) | 97.0% | 100.0% | 98.5% |
| COCO_AI (SD, FLUX, Ideogram) | 92.5% | 98.5% | 95.5% |
| **Average** | **94.8%** | **99.2%** | **97.0%** |

## Usage

```python
from verifai import VerifAI

# Weights are downloaded automatically
detector = VerifAI()
result = detector.detect("image.jpg")

print(f"Label: {result.label}")
print(f"Confidence: {result.confidence:.1%}")
```

## Architecture

**Fusion Detector** = CLIP (80%) + Frequency (20%)

- **CLIP Head**: 768 → 512 → 256 → 2 with LayerNorm, GELU, Dropout
- **Frequency Classifier**: GradientBoostingClassifier on FFT/DCT features

## Training

Trained on:
- [Defactify Dataset](https://huggingface.co/datasets/Rajarshi-Roy-research/Defactify_Image_Dataset) - Modern AI generators
- [VCT² COCO_AI](https://huggingface.co/datasets/NasrinImp/COCO_AI) - COCO + AI variants

## License

MIT License