Official implementation of our CVPR Workshop NTIRE 2025 accepted paper:

XYScanNet: A State Space Model for Single Image Deblurring
Hanzhou Liu, Chengkai Liu, Jiacong Xu, Peng Jiang, Mi Lu
arXiv 2412.10338

Deep state-space models (SSMs), like recent Mamba architectures, are emerging as a promising alternative to CNN and Transformer networks. Existing Mamba-based restoration methods process visual data by leveraging a flatten-and-scan strategy that converts image patches into a 1D sequence before scanning. However, this scanning paradigm ignores local pixel dependencies and introduces spatial misalignment by positioning distant pixels incorrectly adjacent, which reduces local noise-awareness and degrades image sharpness in low-level vision tasks.

To overcome these issues, we propose a novel slice-and-scan strategy that alternates scanning along intra- and inter-slices. We further design a new Vision State Space Module (VSSM) for image deblurring and tackle the inefficiency challenges of the current Mamba-based vision module. Building upon this, we develop XYScanNet, an SSM architecture integrated with a lightweight feature fusion module for enhanced image deblurring. XYScanNet maintains competitive distortion metrics and significantly improves perceptual performance. Experimental results show that XYScanNet enhances KID by 17% compared to the nearest competitor.

• Accepted at CVPR 2025 - NTIRE Workshop
• Novel slice-and-scan scanning for spatially aware state-space modeling
• Lightweight Vision State Space Module (VSSM) designed for efficiency
• Competitive results on GoPro, HIDE, RealBlur_J/R, and RWBI datasets

conda create -n XYScanNet python=3.8 -y
conda activate XYScanNet
conda install pytorch==2.0.0 torchvision==0.15.0 pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -r requirements.txt

Download "GoPro" dataset into './datasets' for example: './datasets/GoPro'. Note: we say the blur images is A and the sharp images is B, e.g., ./GOPRO/test/sharp <-> .GOPRO/test/testB.

Download "VGG19 Pretrained Weights" into './models', which is used to calculate ContrastLoss.

We train our XYScanNetPlus in two stages:

• We pre-train XYScanNetPlus for 4000 epochs with patch size 252x252
  
• Run the following command

python train_XYScanNet_stage1.py --job_name xyscannetp_gopro

• After 4000 epochs, we keep training XYScanNetPlus for 4000 epochs with patch size 320x320
  
• Run the following command

python train_XYScanNet_stage2.py --job_name xyscannetp_gopro

📄 Training Log: GoPro (can be viewed with linux cat commands)

I forgot the training details on RealBlur datasets. As much as I can remember, I used the gopro-trained model (final .pth in stage 1) and fine-tuned it based on the provided config file.

📄 Training Log: RealBlur_J 📄 Training Log: RealBlur_R

For testing on GoPro dataset

• Create folders with a structure results/xyscannetp_gopro/models, where models is the folder that saves the network weights.
• Download "GoPro" full dataset or test set into './datasets' (For example: './datasets/GoPro/test')
  
• Change the dataset path in predict_GoPro_test_results.py line 34.
• Run the following command

python predict_GoPro_test_results.py --job_name xyscannetp_gopro

For testing on HIDE dataset

• Create folders with a structure results/xyscannetp_hide/models, where models is the folder that saves the network weights.
• Download "HIDE" into './datasets'
  
• Run the following command

python predict_HIDE_results.py --job_name xyscannetp_gopro

For testing on RealBlur test sets

• Create folders with a structure results/xyscannetp_realj(or realr)/models, where models is the folder that saves the network weights.
• Download "RealBlur_J" and "RealBlur_R" into './datasets'
  
• Run the following command

python predict_RealBlur_J_test_results.py --job_name xyscannetp_realj

python predict_RealBlur_R_test_results.py --job_name xyscannetp_realr

The following table lists our released pretrained models and qualitative visual examples. All models are available from Google Drive.

PS: We have trained multiple versions of XYScanNetPlus. We are still double-checking whether the released models and images match with the paper results.

• Release Pretrained Models
• Release Test Images
• Release Training Logs
• Release Environment Setup

If you find our work useful, please consider citing us:

@InProceedings{Liu_2025_CVPR,
    author    = {Liu, Hanzhou and Liu, Chengkai and Xu, Jiacong and Jiang, Peng and Lu, Mi},
    title     = {XYScanNet: A State Space Model for Single Image Deblurring},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops},
    month     = {June},
    year      = {2025},
    pages     = {779-789}
}