Welcome! This is the official implementation of the paper "Semantic-Guided Diffusion Model for Single-Step Image Super-Resolution".

• Python 3.10, Pytorch 2.1.2

conda env create -n SamSR python=3.10
conda activate SamSR
pip install -r requirements.txt

from script_segment import generate_masks_from_batch,load_sam_model, generate_weighted_noise, normalize_noise_to_unit_variance
sam_model = load_sam_model()            
with th.no_grad():
  masks_tensor = generate_masks_from_batch(sam_model, x_lr)
  noise_weighted = generate_weighted_noise(masks_tensor)
  sam_noise = normalize_noise_to_unit_variance(noise_weighted) #change your gaussian noise to sam noise

python3 inference.py -i [image folder/image path] -o [result folder] --ckpt weights/SamSR_v1.pth --scale 4 --one_step

Download the necessary models. You can download them into the files of weights from the link below:

• SamSR_v1
• SamSR_v2
• sam

• Real data for image super-resolution: RealSet65 | RealSR
• Test the model

# Results on RealSet65
python inference.py -i testdata/RealSet65 -o results/SamSR/RealSet65 --scale 4 --ckpt weights/SamSR_v1.pth --one_step
# Results on RealSR
python inference.py -i testdata/RealSet65 -o results/SamSR/RealSR --scale 4 --ckpt weights/SamSR_v1.pth --one_step

If you are running on a GPU with limited memory, you could reduce the patch size by setting --chop_size 256 to avoid out of memory. However, this will slightly degrade the performance.

# Results on RealSet65
python inference.py -i testdata/RealSet65 -o results/SamSR/RealSet65 --scale 4 --ckpt weights/SamSR_v1.pth --one_step --chop_size 256 --task SamSR
# Results on RealSR
python inference.py -i testdata/RealSR -o results/SamSR/RealSR --scale 4 --ckpt weights/SamSR_v1.pth --one_step --chop_size 256 --task SamSR

• Download the image ImageNet-Test (Link) to the testdata folder.
• Unzip the downloaded dataset.
• Test the model

python inference.py -i testdata/imagenet256/lq/ -o results/SamSR/imagenet  -r testdata/imagenet256/gt/ --scale 4 --ckpt weights/SamSR_v2.pth --one_step

1. Download the necessary models. You can download them into the files of weights from the link below:
   • ResShift
   • Autoencoder
2. Adjust the data path in the config file. Specifically, correct and complete paths in files of traindata
3. Adjust batchsize according your GPUS.
   • configs.train.batch: [training batchsize, validation btatchsize]
   • configs.train.microbatch: total batchsize = microbatch * #GPUS * num_grad_accumulation

python3 main_distill.py --cfg_path configs/SamSR.yaml --save_dir logs/SamSR

This project is based on ResShift and SinSR.

Please cite our paper if you find our work useful. Thanks!

@article{liu2025semantic,
  title={Semantic-guided diffusion model for single-step image super-resolution},
  author={Liu, Zihang and Zhang, Zhenyu and Tang, Hao},
  journal={arXiv preprint arXiv:2505.07071},
  year={2025}
}

If you have any questions, please feel free to contact me via liuzihang@bit.edu.cn.