Consistency Trajectory Matching for One-Step Generative Super-Resolution (ICCV 2025)

Weiyi You, Mingyang Zhang, Leheng Zhang, Xingyu Zhou, Kexuan Shi, Shuhang Gu

News

• 📄 [2025.03.27] Paper preprint released!
• 🏆 [2025.06.26] Our paper has been accepted to ICCV 2025!
• 💾 [2025.06.30] Codebase and model checkpoints are now available.

Environment

• Python 3.9
• PyTorch 2.0.1

Installation

git clone https://github.com/LabShuHangGU/CTMSR.git

conda create -n ctmsr python=3.9
conda activate ctmsr

pip install -r requirements.txt
python setup.py develop

Training

Data Preparation

• Download the training dataset ImageNet and put them in the folder ./datasets.

Training Commands

• Refer to the training configuration files in ./options/train folder for detailed settings.

# batch size = 4 (GPUs) × 8 (per GPU)

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --use-env --nproc_per_node=4 --master_port=1145  basicsr/train.py -opt options/train/ctmsr_train.yml --launcher pytorch

Testing

Data Preparation

• Download and generate the testing data (ImageNet-Test + RealSR + RealSet65) and put them in the folder ./datasets.

Pretrained Models

• Download the pretrained models and put them in the folder ./experiments/pretrained_models.

Testing Commands

• Refer to the testing configuration files in ./options/test folder for detailed settings.

CUDA_VISIBLE_DEVICES=0 python basicsr/test.py -opt options/test/ctmsr_test.yml

Citation

@article{you2025consistency,
  title={Consistency Trajectory Matching for One-Step Generative Super-Resolution},
  author={You, Weiyi and Zhang, Mingyang and Zhang, Leheng and Zhou, Xingyu and Shi, Kexuan and Gu, Shuhang},
  journal={arXiv preprint arXiv:2503.20349},
  year={2025}
}

Acknowledgements

This code is built on BasicSR and ResShift.