SimFlowSR: Self-similarity Aggregation over Consistent Information Flow for Single Image Super-Resolution

Chia-Ming Lee^1,2, Chih-Chung Hsu^1,2

¹National Yang Ming Chiao Tung University, ²National Cheng Kung University

TL;DR: SimFlowSR integrates CEB (Contextual Encoding Branch) for consistent information flow with GAB (Geometric Aggregation Branch) for parameter-free self-similarity aggregation, achieving efficient high-fidelity super-resolution.

• 🎯 Dual-Branch Design: CEB stabilizes activation dynamics; GAB enhances high-frequency details
• ⚡ High Efficiency: Up to 37% fewer parameters with superior performance
• 🔧 Plug-and-Play: Seamlessly integrates into various backbones (SwinIR, MambaIR, RWKVIR)
• 🎨 Parameter-Free: GAB uses geometric transformations (dihedral group D₄) without learnable parameters

Benchmark results on standard datasets (×4 super-resolution):

SimFlowSR employs a modular dual-branch architecture:

• CEB (Contextual Encoding Branch): Maintains consistent information flow via dense-residual connections
• GAB (Geometric Aggregation Branch): Aggregates self-similar features via parameter-free D₄ transformations

git clone https://github.com/ming053l/SimFlowSR.git
cd SimFlowSR

# Create conda environment
conda create -n simflowsr python=3.8 -y
conda activate simflowsr

# Install PyTorch (adjust CUDA version as needed)
conda install pytorch==1.12.1 torchvision==0.13.1 cudatoolkit=11.6 -c pytorch -c conda-forge

# Install dependencies
pip install -r requirements.txt
python setup.py develop

# Test SimFlowSR-SwinIR
python simflowsr/test.py -opt options/test/SimFlowSR_SwinIR_test.yml

# Test SimFlowSR-MambaIR
python simflowsr/test.py -opt options/test/SimFlowSR_MambaIR_test.yml

# Train SimFlowSR-SwinIR (multi-GPU)
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch \
    --nproc_per_node=4 --master_port=4321 \
    simflowsr/train.py \
    -opt options/train/train_SimFlowSR_SwinIR.yml \
    --launcher pytorch

# Train SimFlowSR-MambaIR (multi-GPU)
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch \
    --nproc_per_node=4 --master_port=4321 \
    simflowsr/train.py \
    -opt options/train/train_SimFlowSR_MambaIR.yml \
    --launcher pytorch

• Training: DF2K (DIV2K + Flickr2K)
• Testing: Set5, Set14, BSD100, Urban100, Manga109

Download links and preprocessing scripts are provided in datasets/README.md.

SimFlowSR exhibits the most compressed activation dynamic range

SimFlowSR achieves the most stable feature propagation. Conventional methods (SwinIR-RSTB, HAT-RHAG) show dramatic activation fluctuations across depths, indicating unstable information flow. DRCT (RDG) improves stability through dense connections, but SimFlowSR further compresses the dynamic range with tighter activation clustering—validating that CEB + GAB maintains stable representations while preserving fine details.

LAM showing superior spatial aggregation with higher Diffusion Index

ERF demonstrating substantially broader spatial coverage

SimFlowSR captures long-range correspondences effectively. LAM visualization shows significantly higher Diffusion Index, while ERF demonstrates broader spatial coverage across all backbones—confirming GAB's multi-scale D₄ transformations consistently enhance spatial modeling capability.

If you find our work helpful, please consider citing:

@article{lee2025simflowsr,
  title={SimFlowSR: Self-similarity Aggregation over Consistent Information Flow for Single Image Super-Resolution},
  author={Lee, Chia-Ming and Hsu, Chih-Chung},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2025}
}

Our work builds upon excellent open-source projects:

• SwinIR
• MambaIR
• DRCT
• BasicSR

We are grateful for their outstanding contributions to the community.

For questions or discussions, please:

• Open an issue on GitHub
• Email: ming053l@gmail.com

This project is released under the MIT License. See LICENSE for details.