DRNet: All-in-One Image Restoration via Prior-Guided Dynamic Reparameterization

Ao Li, Xiaoning Liu, Sheng Li, Yapeng Du, Zhen Long, Lei Luo, Le Zhang, and Ce Zhu

IEEE Transactions on Multimedia (TMM), 2026

---

📌 Overview

DRNet is a novel all-in-one image restoration framework that handles multiple degradation types within a single unified model via Prior-Guided Dynamic Reparameterization. It efficiently addresses:

• Denoising (σ=15, 25, 50)
• Dehazing
• Deraining
• Deblurring
• Low-light Enhancement

---

🔧 Requirements

Environment Setup

# Create conda environment
conda create -n drnet python=3.8
conda activate drnet

# Install dependencies
pip install -r requirements.txt

📦 Pretrained Weights

Pretrained weights are available for download:

🔗 Download Pretrained Weights

Place the weights in the ckpt/ directory:

mkdir -p ckpt/5tasks ckpt/3tasks
# Download and move pretrained weights to corresponding directories

📂 Dataset Preparation

Organize your training data as follows:

data/
├── Denoise/       # Denoising images (noisy/clean pairs)
├── Dehaze/        # Dehazing images (hazy/clear pairs)
├── Derain/        # Deraining images (rainy/clean pairs)
├── Deblur/        # Deblurring images (blurred/sharp pairs)
└── Enhance/       # Low-light enhancement images (low-light/normal pairs)

🚀 Training

Training Command

CUDA_VISIBLE_DEVICES=0,1 python train.py \
    --epochs 150 \
    --batch_size 8 \
    --num_gpus 4 \
    --de_type denoise_15 denoise_25 denoise_50 dehaze derain deblur enhance \
    --ckpt_dir train_ckpt/DRNet \
    --num_workers 8

🔍 Inference

Denoising (5 tasks model)

Test on BSD68 dataset:

# Denoising σ=15
python inference.py --task "denoise" --ckpt_path ckpt/5tasks/DRNet_5tasks.pth --data_path test/denoise/bsd68/ --sigma 15

# Denoising σ=25
python inference.py --task "denoise" --ckpt_path ckpt/5tasks/DRNet_5tasks.pth --data_path test/denoise/bsd68/ --sigma 25

# Denoising σ=50
python inference.py --task "denoise" --ckpt_path ckpt/5tasks/DRNet_5tasks.pth --data_path test/denoise/bsd68/ --sigma 50

Deraining (5 tasks model)

Test on Rain100L dataset:

python inference.py --task "derain" --ckpt_path ckpt/5tasks/DRNet_5tasks.pth --data_path test/derain/Rain100L/

Dehazing (3 tasks model)

Test on SOTS dataset:

python inference.py --task "dehaze" --ckpt_path ckpt/3tasks/DRNet_3tasks.pth --data_path test/dehaze/SOTS/

📝 Citation

If you find DRNet useful for your research, please cite:

@article{li2026drnet,
  title={DRNet: All-in-One Image Restoration via Prior-Guided Dynamic Reparameterization},
  author={Li, Ao and Liu, Xiaoning and Li, Sheng and Du, Yapeng and Long, Zhen and Luo, Lei and Zhang, Le and Zhu, Ce},
  journal={arXiv preprint arXiv:2605.08627},
  year={2026}
}

🙏 Acknowledgement

This codebase is built upon PromptIR. We thank the authors for their awesome work.