RestorMixer: An Efficient and Mixed Heterogeneous Model for Image Restoration

RestorMixer is a novel general-purpose image restoration (IR) model that combines the strengths of multiple architectural paradigms—Convolutional Neural Networks (CNNs), Transformers, and Mamba—to achieve state-of-the-art performance with high inference efficiency. This project is the official implementation of the paper, “An Efficient and Mixed Heterogeneous Model for Image Restoration”.

🌟 Why RestorMixer?

Image restoration is a critical task for recovering high-quality images from degraded ones, but traditional methods are often task-specific and costly to develop. While recent general-purpose models have shown promise, they often rely on a single or two combined architectures, which limits their ability to handle diverse degradation types effectively. RestorMixer addresses this by strategically fusing these architectures.

---

✨ Key Features

• Heterogeneous Architecture Fusion: RestorMixer is the first model to effectively integrate CNN, Mamba, and Transformer architectures into a single, cohesive framework for image restoration.
• Multi-Stage Encoder-Decoder: A hierarchical design processes features at different resolutions, from high-resolution local details to low-resolution global contexts.
• Superior Performance: Achieves leading performance on multiple IR tasks, including deraining, desnowing, and super-resolution, for both single and mixed degradations.
• High Inference Efficiency: The model is designed for speed, ensuring high performance without sacrificing computational efficiency.
• Versatility: This framework is also applicable to other vision tasks (using an encoder-decoder structure).

---

🚀 Getting Started

To get started with RestorMixer, follow the instructions below.

Installation

1. Clone the repository:

   git clone https://github.com/ClimBin/RestorMixer.git
   cd RestorMixer

---

📊 Results

Click to expand and view the performance of RestorMixer across various image restoration tasks.

View Example Results

Here are some visual comparisons demonstrating RestorMixer's effectiveness:

---

📄 Paper & Citation

If you use this code in your research, please cite our paper:

@article{gu2025efficient,
  title={An Efficient and Mixed Heterogeneous Model for Image Restoration},
  author={Gu, Yubin and Meng, Yuan and Zheng, Kaihang and Sun, Xiaoshuai and Ji, Jiayi and Ruan, Weijian and Cao, Liujuan and Ji, Rongrong},
  journal={arXiv preprint arXiv:2504.10967},
  year={2025}
}

---

🙏 Acknowledgements

We would like to thank all the researchers whose work has inspired this project and the open-source community for providing invaluable tools and resources. Thanks to VMamba, Restormer, and all other works inspiring us.

---