Devil is in the Uniformity: Exploring Diverse Learners within Transformer for Image Restoration

Shihao Zhou, Dayu Li, Jinshan Pan, Juncheng Zhou, Jinglei Shi and Jufeng Yang

News

• Jul 19, 2025: Hugging Face Demo is available now, thanks contribution of Sen
• Jun 26, 2025: HINT has been accepted to ICCV 2025 🎉

---

Training

Derain

To train HINT on rain100L, you can run:

./train.sh Deraining/Options/Deraining_HINT_syn_rain100L.yml

Dehaze

To train HINT on SOTS, you can run:

./train.sh Dehaze/Options/RealDehazing_HINT.yml

Denoising

To train HINT on WB, you can run:

./train.sh Denoising/Options/GaussianColorDenoising_HINT.yml

Desnowing

To train HINT on snow100k, you can run:

./train.sh Desnowing/Options/Desnow_snow100k_HINT.yml

Enhancement

To train HINT on LOL_v2_real, you can run:

./train.sh Enhancement/Options/HINT_LOL_v2_real.yml

To train HINT on LOL_v2_synthetic, you can run:

./train.sh Enhancement/Options/HINT_LOL_v2_synthetic.yml

Evaluation

To evaluate HINT, you can refer commands in 'test.sh'

For evaluate on each dataset, you should uncomment corresponding line.

Results

Experiments are performed for different image processing tasks. Here is a summary table containing hyperlinks for easy navigation:

Benchmark	Pretrained model	Visual Results
Rain100L	(code:ngn8)	(code:bdpg)
SOTS	(code:64j8)	(code:dypf)
Snow100K	(code:q2cm)	(code:s7xx)
LOL-v2-Real	(code:6cux)	(code:5bxm)
LOL-v2-Syn	(code:7fi5)	(code:y9uq)
WB	(code:7fi5)	(code:ss8c)

Citation

If you find this project useful, please consider citing:

@inproceedings{zhou_ICCV25_HINT,
  title={Devil is in the Uniformity: Exploring Diverse Learners within Transformer for Image Restoration},
  author={Zhou, Shihao and Li, Dayu and Pan, Jinshan and Zhou, Juncheng and Shi, Jinglei and Yang, Jufeng},
  booktitle={ICCV},
  year={2025}
}

Acknowledgement

This code borrows heavily from Restormer.