# CleanFD Backup

**CleanFD** is a comprehensive modularized library for benchmarking AI-generated image detection methods. This backup includes the complete project with all detectors, datasets, pretrained models, and evaluation scripts.

## Download & Extract

```bash
# huggingface_hub 설치 (필요시)
pip install huggingface_hub

# 다운로드
huggingface-cli download leekwoon/260204_cfd_backup --repo-type dataset --local-dir ./cfd_backup

# 무결성 확인 (선택사항)
cd cleanfd_backup
md5sum -c checksums.md5

# 파일 합치기 및 압축 해제
cat data.tar.gz.part_* | tar -xzvf -
```

## Directory Structure

```
cleanfd/
├── cleanfd/                # Core library modules
│   ├── detector/           # Detector implementations (23 detectors)
│   ├── dataset/            # Dataset utilities
│   ├── nn_classifier/      # Neural network modules
│   └── utils/              # Utilities
├── configs/                # Hydra configuration files (23 configs)
├── data/                   # Dataset directory
│   ├── train/              # Training data (real + fake)
│   ├── val/                # Validation data
│   ├── test/               # Test data
│   └── test_processed/     # Processed test data (JPEG, WebP, resize)
├── pretrained/             # Pretrained model weights
├── pipelines/              # Training/evaluation pipelines
├── FerretNet/              # FerretNet implementation
├── icml26_paper/           # ICML 2026 paper materials
├── notebooks/              # Jupyter notebooks
├── results/                # Experiment results
├── scripts/                # Utility scripts
└── requirements.txt        # Python dependencies
```

## Supported Detectors (23 methods)

### Frequency/Reconstruction-based
- **DRCT**: Dual-Residual ConvNeXt Transformer
- **RINE**: Reconstruction-based detection
- **SAFE**: Spectral Analysis for Fake Evidence
- **Aeroblade**: Aerospace-inspired detector

### Gradient-based
- **Corvi / Corvi+ / Corvi Mask Gated / Corvi Inpaint**
- **Rajan / Rajan+ / Rajan Mask Gated / Rajan Inpaint**

### CLIP-based
- **AIDE**: CLIP-based detector
- **C2P-CLIP**: Contrastive-to-Positive CLIP
- **CLIPDet**: CLIP-based detection
- **UFD**: Universal Fake Detector

### Novel Methods
- **FerretNet**: Feature-extraction based detector
- **CoDE**: Code-based detector
- **LaDeDa**: Latent-based detector
- **BFree**: Boundary-free detector
- **DDA**: Domain Discriminative Attention
- **NPR**: Neural Pattern Recognition
- **WarpAD**: Warping-based Anomaly Detection

## Dataset Information

### Training Data
- **Real Images**: COCO, LSUN
- **Fake Images**: 
  - Aligned inversions (Rajan method)
  - Latent Diffusion Model (LDM) generations
  - Stable Diffusion v2.1 inpainted images
  - LSUN inpainted images

### Test Data
- **Real Images**: RedCaps dataset
- **Fake Images** (12 generators):
  - Stable Diffusion (SD)
  - Midjourney
  - Kandinsky
  - Playground
  - PixelArt
  - LCM
  - Flux
  - Wuerstchen
  - Amused
  - Chameleon
  - Loki
  - WildRF

### Data Processing
- **test**: Original test images
- **test_processed**: Images with perturbations (JPEG, WebP, resize)

## Key Features

- **Unified Interface**: Consistent API for all detectors
- **Hydra Configuration**: Flexible experiment management
- **Comprehensive Evaluation**: Multiple metrics (AP, ACC)
- **Perturbation Testing**: Robustness evaluation with various image processing
- **Modular Design**: Easy to add new detectors and datasets
- **Reproducible**: Configuration-based experiments

## Usage Example

```bash
# Train a detector
python pipelines/corvi.py

# Evaluate on test set
python scripts/eval.py --method corvi --test_dir test

# Analyze results
python scripts/analyze_results.py --methods corvi rajan --test_dirs test test_processed
```

## Citation

This library aggregates multiple AI-generated image detection methods for research purposes. Please cite the original papers when using specific detectors.