leekwoon/kd_backup

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

# huggingface_hub 설치 (필요시)
pip install huggingface_hub

# 다운로드
huggingface-cli download leekwoon/cleanfd_backup --repo-type dataset --local-dir ./cleanfd_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
│   │   ├── aeroblade_detector.py
│   │   ├── aide_detector.py
│   │   ├── bfree_detector.py
│   │   ├── c2pclip_detector.py
│   │   ├── clipdet_detector.py
│   │   ├── corvi_detector.py
│   │   ├── corvi_plus_detector.py
│   │   ├── corvi_mask_gated_detector.py
│   │   ├── dda_detector.py
│   │   ├── drct_detector.py
│   │   ├── npr_detector.py
│   │   ├── rajan_detector.py
│   │   ├── rajan_plus_detector.py
│   │   ├── rajan_mask_gated_detector.py
│   │   ├── rine_detector.py
│   │   ├── safe_detector.py
│   │   ├── ufd_detector.py
│   │   └── warpad_detector.py
│   ├── dataset/            # Dataset utilities
│   │   ├── corvi_dataset.py
│   │   ├── rajan_dataset.py
│   │   ├── inpaint_dataset.py
│   │   ├── folder_dataset.py
│   │   └── transforms.py
│   ├── nn_classifier/      # Neural network modules
│   │   ├── resnet.py
│   │   └── sae.py
│   └── utils/              # Utilities
│       └── early_stopping.py
├── configs/                # Hydra configuration files
│   ├── aeroblade.yaml
│   ├── aide.yaml
│   ├── bfree.yaml
│   ├── c2pclip.yaml
│   ├── clipdet.yaml
│   ├── corvi.yaml
│   ├── corvi_plus.yaml
│   ├── corvi_mask_gated.yaml
│   ├── dda.yaml
│   ├── drct.yaml
│   ├── npr.yaml
│   ├── rajan.yaml
│   ├── rajan_plus.yaml
│   ├── rajan_mask_gated.yaml
│   ├── rine.yaml
│   ├── safe.yaml
│   ├── ufd.yaml
│   └── warpad.yaml
├── data/                   # Dataset directory
│   ├── train/              # Training data
│   │   ├── real/           # Real images
│   │   │   ├── coco/       # COCO dataset
│   │   │   └── lsun/       # LSUN dataset
│   │   └── fake/           # AI-generated images
│   │       ├── aligned/    # Aligned inversions
│   │       ├── ldm/        # Latent Diffusion Model
│   │       ├── sd21_inpainted_*/
│   │       └── lsun_inpaint_*/
│   ├── val/                # Validation data
│   │   ├── real/
│   │   └── fake/
│   ├── test/               # Test data
│   │   ├── real/
│   │   │   └── redcaps/    # RedCaps dataset
│   │   └── fake/           # Various generators
│   └── test_processed/     # Processed test data
│       ├── real/
│       └── fake/
├── pretrained/             # Pretrained model weights
│   ├── DRCT-2M/
│   ├── GenImage/
│   ├── aide/
│   ├── bfree/
│   ├── c2pclip/
│   ├── clipdet/
│   ├── dda/
│   ├── npr/
│   ├── rine/
│   ├── safe/
│   └── ufd/
├── pipelines/              # Training/evaluation pipelines
│   ├── aeroblade.py
│   ├── aide.py
│   ├── bfree.py
│   ├── c2pclip.py
│   ├── clipdet.py
│   ├── corvi.py
│   ├── corvi_plus.py
│   ├── corvi_mask_gated.py
│   ├── dda.py
│   ├── drct.py
│   ├── npr.py
│   ├── rajan.py
│   ├── rajan_plus.py
│   ├── rajan_mask_gated.py
│   ├── rine.py
│   ├── safe.py
│   ├── ufd.py
│   └── warpad.py
├── results/                # Experiment results
│   ├── corvi/
│   ├── corvi_plus/
│   ├── corvi_mask_gated_*/
│   ├── rajan/
│   ├── rajan_plus/
│   ├── rajan_mask_gated_*/
│   └── ...
├── scripts/                # Utility scripts
│   ├── eval.py
│   ├── analyze_results.py
│   ├── analyze_results_v2.py
│   └── analyze_results.sh
├── notebooks/              # Jupyter notebooks
│   ├── figures/            # Figure generation
│   ├── captum/             # Interpretability
│   └── ...
├── requirements.txt        # Python dependencies
└── setup.py               # Package setup

Supported Detectors

Frequency-based Methods

• DRCT: Dual-Residual ConvNeXt Transformer
• RINE: Reconstruction-based detection
• SAFE: Spectral Analysis for Fake Evidence

Gradient-based Methods

• Corvi: Core visual features detector
• Corvi+: Enhanced Corvi with additional processing
• Rajan: Gradient-based detector
• Rajan+: Enhanced Rajan detector

Masked/Gated Methods

• Corvi Mask Gated: Corvi with attention gating
• Rajan Mask Gated: Rajan with attention gating

CLIP-based Methods

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

Other Methods

• Aeroblade: Aerospace-inspired 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

# 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.