[ICME'26] Revisiting Event Guided Deblurring with State Space Model

[News] You may also want to check our related works:

• Event Deblur Pro (2026.03) Code 🏆 3rd Place of 2nd Event-based Image Deblurring Challenge (NTIRE@CVPR'26)
• TRM-UNet (2026.01) Code ICASSP'26 (CCF-B)

TL;DR: New state-of-the-art performance on the GoPro dataset.

Installation

git clone https://github.com/NikonD850/REDGSSM.git
cd REDGSSM

conda create -n REDGSSM python=3.10 -y
conda activate REDGSSM
pip install torch==2.5.1+cu124 torchvision==0.20.1+cu124 --index-url https://download.pytorch.org/whl/cu124

git clone https://github.com/state-spaces/mamba.git
cd mamba
git checkout 8ffd905
python -m pip install . --no-build-isolation
cd ..

python -m pip install matplotlib scikit-image opencv-python yacs joblib natsort h5py tqdm timm thop

Training and Evaluation

The model is trained with 4 NVIDIA RTX 4090D.

The time for 1 epoch (1000 iterations) is within 75 minutes, including both training and validating.

For TRAINING SPEED UP, please follow our new repository.

Train

• Download the GoPro events train/test dataset (code: kmaz) to your data root (provided by AHDINet's authors)
• Change both training.yml and config.py to your settings.
• Train the model with default arguments by running

 nohup python main_train.py > REDGSSM-train.log 2>&1 &

Evaluation

• Download the GoPro events test dataset (code: kmaz) to your data root (provided by AHDINet's authors)
• Download the pretrained model to REGDSSM/models/REGDSSM/model_best.pth
• Change both testing.yml and config.py to your settings.
• Test the model with default arguments by running

  python main_test.py

Acknowledgement

Thanks to the inspirations and codes from AHDINet and EVSSM

Cite this work (BibTeX)

@INPROCEEDINGS{fan2026revisiting,
  author={Fan, Dawei and Ji, Fan and Tang, Xiongxin and Chu, Xiaofeng and Chen, Qiao and Yang, Hanxiang and Lin, Yijun and Xu, Fanjiang},
  booktitle={2026 IEEE International Conference on Multimedia and Expo (ICME)}, 
  title={Revisiting Event Guided Deblurring with State Space Model}, 
  year={2026},
  volume={},
  number={},
  pages={1-6},
  }