FADPNet: Frequency-Aware Dual-Path Network for Face Super-Resolution (TMM 2026)

We have trained and tested the codes on:

• Ubuntu 20.04
• CUDA 11.1
• Python 3.8+

Install required packages:

pip install -r requirements.txt

Note: If you need to train from scratch, please install Mamba-ssm:

pip install causal-conv1d>=1.1.0
pip install mamba-ssm

The commands used to train the released models are provided in script train.sh. Here are some training tips:

• You should download CelebA to train FADPNet. Please change the --dataroot to the path where your training images are stored.
• To train FADPNet, we simply crop out faces from CelebA without pre-alignment, because for ultra-low resolution face SR, it is difficult to pre-align the LR images.
• Please change the --name option for different experiments. Tensorboard records with the same name will be moved to check_points/log_archive, and the weight directory will only store weight history of the latest experiment with the same name.
• If there's not enough memory, you can turn down the --batch_size.
• --gpus specifies the number of GPUs used for training. The script will use GPUs with more available memory first. To specify the GPU index, uncomment the export CUDA_VISIBLE_DEVICES=.

# Train FADPNet (8x scale)
CUDA_VISIBLE_DEVICES=0,1 python train.py --gpus 2 --name fadpnet --model fadpnet \
    --lr 0.0002 --beta1 0.9 --beta2 0.99 --scale_factor 8 --load_size 128 \
    --dataroot /path/to/datasets/CelebA --dataset_name celeba --batch_size 16 --total_epochs 150 \
    --visual_freq 100 --print_freq 10 --save_latest_freq 500

# On CelebA Test set
python test.py --gpus 1 --model fadpnet --name fadpnet \
    --load_size 128 --dataset_name single --dataroot /path/to/datasets/test_datasets/CelebA1000/LR_x8_up/ \
    --pretrain_model_path ./pretrain_models/fadpnet/fadpnet_best.pth \
    --save_as_dir results_celeba/fadpnet

# On Helen Test set
python test.py --gpus 1 --model fadpnet --name fadpnet \
    --load_size 128 --dataset_name single --dataroot /path/to/datasets/test_datasets/Helen50/LR_x8_up/ \
    --pretrain_model_path ./pretrain_models/fadpnet/fadpnet_best.pth \
    --save_as_dir results_helen/fadpnet

We provide evaluation codes in script test.sh to calculate PSNR/SSIM/LPIPS/VIF/Params/FLOPs scores.

This code is built on WFEN and MambaIR. We thank the authors for sharing their codes.

If you find this work useful for your research, please cite:

@article{xu2026fadpnet,
  title={FADPNet: Frequency-Aware Dual-Path Network for Face Super-Resolution},
  author={Xu, Siyu and Li, Wenjie and Gao, Guangwei and Yang, Jian and Qi, Guo-Jun and Lin, Chia-Wen},
  journal={IEEE Transactions on Multimedia},
  year={2026}
}

If you have any questions, please email xusiyu200107@163.com.