[ECCV 2026] CUST : Clustered Unit-level Similarity Transformer for Lightweight Image Super-Resolution

Author : Jeongsoo Kim

Our project has been accepted as a poster presentation at ECCV 2026. You can see our paper at here(huggingface) or here(arXiv).

Requirements

# Install Packages
pip install -r requirements.txt
pip install matplotlib

# Install BasicSR
python3 setup.py develop

Dataset

We use DIV2K as Training dataset. You can download the dataset at https://github.com/dslisleedh/Download_df2k/blob/main/download_df2k.sh and prepare other test datasets at https://github.com/XPixelGroup/BasicSR/blob/master/docs/DatasetPreparation.md#Common-Image-SR-Datasets

And also, you'd better extract subimages using

python3 scripts/data_preparation/extract_subimages.py

By running the code above, you may get subimages of training datasets.

Pretrained Models

Pre-trained models can be downloaded from experiments/pretrained_model.

Training and Test

You can train our CUST following commands below

python3 basicsr/train.py -opt options/train/CUST/cust_base(plus, small)_x2(3,4).yml

Test

You can test our CUST following commands below

python3 basicsr/test.py -opt options/test/CUST_base(small)/test_base(small)_benchmark_x2(3, 4).yml

Results

Result Table with #Param and #FLOPs

Result Table with GPU Consumption and AVG Inference Time

Qualtitative Results

Inference Results

We will provide visual results of CUST_Base soon. If you want to see only architecture, please refer to CUST_arch.py.