MIRAGE

Efficient Degradation-agnostic Image Restoration via Channel-Wise Functional Decomposition and Manifold Regularization

The official PyTorch Implementation of AnyIR for All-in-One Image Restoration

Bin Ren ^1,2$^\star$, Yawei Li⁵, Xu Zheng⁴, Yuqian Fu⁵, Danda Pani Paudel⁵, Hong Liu⁶$^\dagger$, Ming-Hsuan Yang ⁷, Luc Van Gool ⁵, and Nicu Sebe ²

$\star$: This work was partially conducted during the visiting stay at INSAIT.
$\dagger$: Corresponding author

¹ Mohamed bin Zayed University of Artificial Intelligence, UAE,
² University of Trento, Italy,
³ ETH Zürich, Switzerland,
⁴ HKUST (Guangzhou), China,
⁵ INSAIT Sofia University, "St. Kliment Ohridski", Bulgaria,
⁶ Peking University, China,
⁷ University of California, Merced, USA

Latest

• Projectpage release.
• Ckpts release.
• Main visual results release.
• Code release.
• 01/2026: 🍺🎉 Our MIRAGE is accepted by ICLR2026!

Method

Installation

1) Environment

micromamba create -n mirage python=3.9 -y
micromamba activate mirage
# or
conda create -n mirage python=3.9 -y
conda activate mirage

2) Dependencies

# NOTE: file in this repo is currently named "requiements.txt"
pip install -r requiements.txt

3) CUDA (if needed on your cluster)

export LD_LIBRARY_PATH=/opt/modules/nvidia-cuda-11.8.0/lib64:$LD_LIBRARY_PATH
export PATH=/opt/modules/nvidia-cuda-11.8.0/bin:$PATH

Datasets Preparation:

TODO

Checkpoints Downloads:

TODO

Visual Results Downloads:

TODO

Citation

If you find this project useful, please cite:

TODO

Acknowledgements

This work was partially supported by the FIS project GUIDANCE (Debugging Computer Vision Models via Controlled Cross-modal Generation) (No. FIS2023-03251).

The code base is built on top of excellent prior work, including:

• PromptIR
• AirNet