MIRAGE

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

Bin Ren^1,2  ·  Yawei Li³  ·  Xu Zheng⁴  ·  Yuqian Fu⁵  ·  Danda Pani Paudel⁵
Hong Liu^6†  ·  Ming-Hsuan Yang⁷  ·  Luc Van Gool⁵  ·  Nicu Sebe²

¹MBZUAI, UAE   ²University of Trento, Italy   ³ETH Zürich, Switzerland   ⁴HKUST (GZ), China
⁵INSAIT Sofia University, Bulgaria   ⁶Peking University, China   ⁷UC Merced, USA
^†Corresponding author

Teaser: (a)–(d) Visual comparison for Denoising, Deraining, Composited Degradations (low-light, haze, and snow), and underwater image enhancement. (e) Average PSNR/SSIM across 4 all-in-one and 1 zero-shot settings.

• 🌐 Project page release
• 🖼️ Main visual results release
• 05/2026 🔖 Checkpoints released
• 05/2026 💻 Code released
• 01/2026 🍺 MIRAGE accepted at ICLR 2026!

TODO — architecture diagram and method overview coming soon.

conda create -n mirage python=3.9 -y
conda activate mirage

pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 \
    --index-url https://download.pytorch.org/whl/cu129

pip install -r requirements.txt

# Check CUDA availability
nvidia-smi
nvcc --version

# On cluster systems, load via environment modules:
module avail cuda
module load cuda/12.9

# Or set CUDA path manually:
export CUDA_HOME=/usr/local/cuda
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH

We follow dataset preparation from prior works:

• 3-Degradation: PromptIR (NeurIPS 2023)
• 5-Degradation: AdaIR (ICLR 2025)

⚠️ Important: Follow original dataset licenses. Provided datasets are for academic research only.

.../datasets/Train/
├── Deblur/
│   ├── blur/
│   └── sharp/
├── Dehaze/
│   ├── train/
│   └── test/
├── Denoise/
│   └── *.bmp / *.jpg
├── Derain/
│   ├── gt/
│   └── rainy/
└── Enhance/
    ├── gt/
    └── low/

.../datasets/test/
├── deblur/
│   └── gopro/
│       ├── input/
│       └── target/
├── dehaze/
│   ├── input/
│   └── target/
├── denoise/
│   ├── bsd68/
│   └── urban100/
├── derain/
│   └── Rain100L/
│       ├── input/
│       └── target/
└── enhance/
    └── lol/
        ├── input/
        └── target/

TODO — download links coming soon.

1. Download the checkpoint and place it under ./train_ckpt/3deg_[tiny|small]/
2. Set the correct project directory and test path in test_3deg_[tiny|small].sh
3. Run:

sh test_3deg_tiny.sh    # Tiny  (6M)
sh test_3deg_small.sh   # Small (10M)

1. Download the checkpoint and place it under ./train_ckpt/5deg_[tiny|small]/
2. Set the correct project directory and test path in test_5deg_[tiny|small].sh
3. Run:

sh test_5deg_tiny.sh    # Tiny  (6M)
sh test_5deg_small.sh   # Small (10M)

TODO

TODO

sh train_3deg_tiny.sh    # Tiny  (6M)
sh train_3deg_small.sh   # Small (10M)

sh train_5deg_tiny.sh    # Tiny  (6M)
sh train_5deg_small.sh   # Small (10M)

TODO

TODO

If you find this work useful, please consider citing:

@inproceedings{ren2026efficient,
  title     = {Efficient Degradation-agnostic Image Restoration via Channel-Wise
               Functional Decomposition and Manifold Regularization},
  author    = {Bin Ren and Yawei Li and Xu Zheng and Yuqian Fu and
               Danda Pani Paudel and Hong Liu and Ming-Hsuan Yang and
               Luc Van Gool and Nicu Sebe},
  booktitle = {The Fourteenth International Conference on Learning Representations (ICLR)},
  year      = {2026}
}

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

The codebase builds on excellent prior work:

• PromptIR (NeurIPS 2023)
• AirNet (CVPR 2022)