IRF-Net: A Retinex-Guided Framework with Auxiliary Local Spectral Response Reconstruction for Generalizable AI-Generated Image Detection

Dengtai Tan*, Jing Wang*, Shujie Yang, Deyi Yang
Gansu University of Political Science and Law, Lanzhou, China
* Equal contribution

---

Overview

IRF-Net is a generated image detection framework that integrates:

1. Retinex-based illumination–reflectance decomposition — reduces illumination interference and produces a DCT-domain multi-channel reflectance representation (48 channels from 4×4 local DCT blocks).
2. Auxiliary Local Spectral Response (LSR) reconstruction — the decoder reconstructs LSR maps derived from block-wise DCT coefficients, guiding the shared encoder to preserve fine-grained spectral response information.
3. Joint encoder–decoder architecture — classification and LSR reconstruction are jointly optimized with (L_{total} = L_{cls} + \lambda_{rec} L_{LSR}) ((\lambda_{rec}=0.01)).

The reconstructed LSR maps provide observable auxiliary evidence for comparing response distributions between real and generated images.

---

Results

Cross-model evaluation on ForenSynths (Acc / AP)

Method	ProGAN	StyleGAN	StyleGAN2	BigGAN	CycleGAN	StarGAN	GauGAN	Deepfake	Mean
CNNDetect	91.4/99.4	63.8/91.4	76.4/97.5	52.9/73.3	72.7/88.6	63.8/90.8	63.9/92.2	51.7/62.3	67.1/86.9
NPR	99.8/100.0	96.3/99.8	97.3/100.0	87.5/94.5	95.0/99.5	99.7/100.0	86.6/88.8	77.4/86.2	92.5/96.1
IRF-Net	99.2/100.0	97.6/100.0	98.5/100.0	87.9/95.8	93.8/97.4	99.9/100.0	88.4/93.3	79.0/89.5	94.3/97.6

Cross-model evaluation on GenImage (Acc / AP)

Setting	Midjourney	SDv1.4	SDv1.5	ADM	Glide	Wukong	VQDM	BigGAN	Mean
IRF-Net (256×256 crop)	84.5/94.9	82.1/92.3	81.4/91.9	85.0/93.9	82.9/94.8	77.7/88.7	79.7/86.0	83.4/91.2	82.1/91.7
IRF-Net (full)	97.3/99.7	91.9/98.5	91.9/98.3	93.8/99.2	90.9/98.2	89.9/97.9	87.6/97.4	89.7/97.5	91.6/98.3

Bias-controlled evaluation on CSAIID (Acc %)

Method	Outdoor	People	Animals	Vehicles	Food	Complex	Avg
NPR	71.64	73.54	66.68	71.09	70.89	74.15	71.33
IRF-Net	96.40	96.90	96.00	97.80	95.90	95.80	96.47

---

Installation

git clone https://github.com/wangjing222-hue/IRF-Net.git
cd IRF-Net
pip install -r requirements.txt

---

Datasets

Dataset	Description	Link
ForenSynths	Training: ProGAN; Test: 8 unseen GANs	GitHub
GenImage	8 diffusion-model generators	GitHub
CSAIID	Cross-semantic benchmark	Cai et al., ICCV Workshops 2025

Organize each dataset as follows:

dataset_root/
    0_real/
        image001.jpg
        ...
    1_fake/
        image001.jpg
        ...

---

Training

python train.py \
    --train_dir /path/to/ForenSynths/train \
    --test_dir  /path/to/ForenSynths/test \
    --epochs 20 \
    --batch_size 32 \
    --lr 3e-4 \
    --recon_weight 0.01 \
    --output_dir ./checkpoints

Key arguments:

Argument	Default	Description
--train_dir	—	Path to training dataset
--test_dir	—	Path to test dataset
--epochs	20	Number of training epochs
--batch_size	32	Batch size
--lr	3e-4	Learning rate (AdamW, weight decay 1e-4)
--recon_weight	0.01	Weight λ_rec for LSR reconstruction loss
--output_dir	./checkpoints	Directory to save checkpoints

---

Evaluation

python test.py \
    --model_path checkpoints/best_model.pth \
    --test_dir   /path/to/test \
    --batch_size 32

Output includes:

• Accuracy and Average Precision (AP) for classification
• MSE, PSNR, SSIM for LSR map reconstruction quality

For GenImage full-resolution evaluation, process images at their original resolution without fixed-size cropping.

---

Project Structure

IRF-Net/
├── README.md
├── requirements.txt
├── train.py                # Training entry point
├── test.py                 # Evaluation entry point
├── models/
│   └── densenet.py         # DenseResNet encoder, LSR decoder, ECA, DenseBlock
├── utils/
│   ├── dct.py              # FastDCT, ConvDct_torch
│   ├── retinex.py          # DCT-domain RetinexProcessor
│   └── lsr_map.py          # LSRMapLayer (target LSR map computation)
└── datasets/
    └── dataset.py          # ForenSynthsDataset

---

Citation

If you find this work useful, please cite:

@article{tan2025irfnet,
  title   = {IRF-Net: A Retinex-Guided Framework with Auxiliary Local Spectral
             Response Reconstruction for Generalizable AI-Generated Image Detection},
  author  = {Tan, Dengtai and Wang, Jing and Yang, Shujie and Yang, Deyi},
  journal = {Engineering Applications of Artificial Intelligence},
  year    = {2025}
}

---

License

This project is released under the MIT License.