README.md

---
language:
- en
license: apache-2.0
tags:
- image-dehazing
- onnx
- tensorrt
- openvino
- mnn
---

# PGL-Net: Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling

Welcome to the official model repository for **PGL-Net**. This repository hosts the pre-trained weights and highly optimized deployment files for our efficient real-world image dehazing network.

## 📝 Project Overview

**PGL-Net (Physics-Inspired Global-Local Decoupling Network)** is a lightweight architecture that embeds physical inductive biases via operator-level emulation to address the challenging ill-posed problem of single image dehazing. 

Unlike traditional methods that suffer from inaccurate parameter estimation or deep learning approaches that act as heavy "black boxes," PGL-Net explicitly decouples the dehazing process into two stages:
- **Global Distribution Rectification:** The Physics-Inspired Affine Fusion (PAF) module implicitly models transmission and airlight subtraction to rectify feature distributions globally.
- **Local Structural Refinement:** The Degradation-Aware Modulation (DAM) block adaptively restores locally variant details.

As a result, PGL-Net achieves superior restoration quality comparable to heavy Transformer models, but with only ~3% of the parameters.

- **Paper:** [Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling](https://arxiv.org/abs/2606.25732) [[Code]](https://github.com/sc-30-bit/PGL-Net)
- **Task:** Real-World Image Dehazing
- **Model Variants:** PGL-Net-T (Tiny), PGL-Net-S (Small)

## 🚀 Deployment-Ready Weights

To facilitate real-world applications and industrial deployment, we provide extensive exported weights across multiple inference backends. All files are organized and ready for immediate download.

### Available Formats
- **PyTorch (`.pk`)**: Standard weights for research and fine-tuning.
- **ONNX (`.onnx`)**: Cross-platform deployment.
- **TensorRT (`.engine`)**: Ultra-low latency inference on NVIDIA GPUs.
- **OpenVINO (`.bin` / `.xml`)**: Optimized for Intel CPUs/GPUs.
- **MNN (`.mnn`)**: Lightweight deployment for mobile and edge devices.

### Supported Datasets (Pre-trained Domains)
The weights are provided for models trained on standard real-world dehazing benchmarks:
- `RUDB`
- `RRSHID` (Remote Sensing Dehazing)
- `RW2AH`

*(Files are named following the pattern: `{dataset}_pglnet_{size}.{format}`)*

## 💻 How to Download and Use

You can easily download specific deployment files using the `huggingface_hub` Python library.

```python
from huggingface_hub import hf_hub_download

# Example: Download the ONNX weights for PGL-Net-T trained on RW2AH
file_path = hf_hub_download(
    repo_id="klay11/PGL-Net", 
    filename="rw2ah_pglnet_t.onnx"
)
print(f"Model downloaded to: {file_path}")
```

## 📖 Citation

If you find our work or these weights useful in your research or deployment, please consider citing our paper:

```bibtex
@article{qu2026efficient,
  title   = {Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling},
  author  = {Qu, Yifei and Li, Ru and Chen, Junjie and Wu, Jinyuan},
  journal = {arXiv preprint arXiv:2606.25732},
  year    = {2026}
}
```