LightBSR: Towards Lightweight Blind Super-Resolution via Discriminative Implicit Degradation Representation Learning

[arxiv]

Created by Jiang Yuan, Ji Ma, Bo Wang, Guanzhou Ke, Weiming Hu

This repository contains PyTorch implementation for LightBSR: Towards Lightweight Blind Super-Resolution via Discriminative Implicit Degradation Representation Learning (Accepted by ICCV 2025).

Environment

pip install -r requirements.txt

Train

1. Prepare data

1.1 Training data

1.1.1 Download the DIV2K dataset and the Flickr2K dataset.

1.1.2 Combine the HR images from these two datasets in ./datasets/DF2K/HR to build the DF2K dataset.

1.2 Testing data

Download benchmark datasets (e.g., Set5, Set14 and other test sets) and prepare HR/LR images in ./datasets/benchmark.

2. Train

2.1 Teacher model

2.1.1 Replace all code related to degradation settings in the specified files to ensure they match the training degradation settings (All locations requiring changes are marked with TODO) : ./model/teacher.py ./teacher_main.sh ./teacher_trainer.py

2.1.2 training: bash teacher_main.sh

2.2 Student model

2.2.1 Replace all code related to degradation settings in the specified files to ensure they match the training degradation settings (All locations requiring changes are marked with TODO) : ./student_main.sh ./student_option.py ./student_trainer.py

2.2.2 training: bash student_main.sh

3. Test model

Select the corresponding degradation condition parameters and perform testing: bash student_test.sh

Citation

@InProceedings{Yuan_2025_ICCV,
    author    = {Yuan, Jiang and Ma, Ji and Wang, Bo and Ke, Guanzhou and Hu, Weiming},
    title     = {LightBSR: Towards Lightweight Blind Super-Resolution via Discriminative Implicit Degradation Representation Learning},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {11927-11936}
}