DiNAT-IR: Exploring Dilated Neighborhood Attention for High-Quality Image Restoration

Official PyTorch implementation of the paper:
DiNAT-IR: Exploring Dilated Neighborhood Attention for High-Quality Image Restoration
Hanzhou Liu, Binghan Li, Chengkai Liu, Mi Lu

Transformers, with their self-attention mechanisms for modeling long-range dependencies, have become a dominant paradigm in image restoration tasks. However, the high computational cost of self-attention limits scalability to high-resolution images, making efficiency-quality trade-offs a key research focus. To address this, Restormer employs channel-wise self-attention, which computes attention across channels instead of spatial dimensions. While effective, this approach may overlook localized artifacts that are crucial for high-quality image restoration. To bridge this gap, we explore Dilated Neighborhood Attention (DiNA) as a promising alternative, inspired by its success in high-level vision tasks. DiNA balances global context and local precision by integrating sliding-window attention with mixed dilation factors, effectively expanding the receptive field without excessive overhead. However, our preliminary experiments indicate that directly applying this global-local design to the classic deblurring task hinders accurate visual restoration, primarily due to the constrained global context understanding within local attention. To address this, we introduce a channel-aware module that complements local attention, effectively integrating global context without sacrificing pixel-level precision. The proposed DiNAT-IR, a Transformer-based architecture specifically designed for image restoration, achieves competitive results across multiple benchmarks, offering a high-quality solution for diverse low-level computer vision problems.

---

🔥 News

• 2025.08.02: ✅ Visual results released!
• 2025.08.02: ✅ Training and inference instructions updated!
• 2025.08.02: ✅ Code and pretrained models released!
• 2025.07.23: 📄 Paper available on arXiv

---

📦 Supported Tasks

We evaluate DiNAT-IR on multiple image restoration tasks:

• ✅ Motion Deblurring (e.g., GoPro)
• ✅ Dual-Pixel Defocus Deblurring (e.g., DPDD)
• ✅ Single Image Defocus Deblurring
• ✅ Image Denoising (e.g., SIDD)
• ✅ Image Deraining (e.g., Rain100H, Rain1400)

---

🧱 Installation

This implementation is based on BasicSR.

git clone https://github.com/HanzhouLiu/DiNAT-IR.git
cd DiNAT-IR

# Conda environment
conda create -n DiNAT-IR python=3.8
conda activate DiNAT-IR

# Install PyTorch & CUDA
conda install pytorch==2.0.0 torchvision==0.15.0 pytorch-cuda=11.8 -c pytorch -c nvidia

# Install natten (via Shi-Labs wheels)
pip3 install natten==0.14.6+torch200cu118 -f https://shi-labs.com/natten/wheels

# Install other dependencies
pip install -r requirements.txt

# Develop mode without CUDA extensions
python setup.py develop --no_cuda_ext

---

📊 Benchmark Results

We evaluate DiNAT-IR across several benchmarks. Full results are available on Hugging Face:

👉 📁 Hugging Face Repository

and pretrained models are in the /experiments/pretrained_weights folder.

Task	Dataset	Metric (↑)	DiNAT-IR	Restormer
Motion Deblurring	GoPro	PSNR	33.80	32.92
Motion Deblurring	HIDE	PSNR	31.57	31.22
Defocus Deblurring (DP)	DPDD	PSNR	27.05	26.66
Single Image Defocus	DPDD	PSNR	26.14	25.98
Image Denoising	SIDD	PSNR	39.89	40.02
Image Deraining	Rain100H	PSNR	31.26	31.46
Image Deraining	Rain100L	PSNR	38.93	38.99
Image Deraining	Test2800	PSNR	33.91	34.18
Image Deraining	Test1200	PSNR	32.31	33.19
Image Deraining	Test100	PSNR	31.22	32.00

📖 Training & Inference Instructions

For step-by-step guides on training and evaluating DiNAT-IR across different datasets, please refer to:

• 🔧 GoPro Reproduction Guide — Motion Deblurring
• 🔧 DPDD Reproduction Guide — Defocus Deblurring
• 🔧 SIDD Reproduction Guide — Image Denoising
• 🔧 Rain13K Reproduction Guide — Image Deraining

📜 Citation

If you find this work helpful, please consider citing:

@article{liu2025dinat,
  title={DiNAT-IR: Exploring Dilated Neighborhood Attention for High-Quality Image Restoration},
  author={Liu, Hanzhou and Li, Binghan and Liu, Chengkai and Lu, Mi},
  journal={arXiv preprint arXiv:2507.17892},
  year={2025}
}