[Neurocomputing 2025] LUTFormer: Lookup Table Transformer for Image Enhancement.

Jinwon Ko, Keunsoo Ko, Hanul Kim and Chang-Su Kim.

Official code for "LUTFormer: Lookup Table Transformer for Image Enhancement" in Neurocomputing 2025. [paper]

Dataset

The FiveK, PPR10K, UIEB, and EUVP datasets are used for experiments.
The AdaInt project also provides an instruction for generating a 480p version of FiveK to accelerate in the training process.

Installation

Create conda environment:

$ conda create -n LUTFormer python=3.9 anaconda
$ conda activate LUTFormer
$ conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
$ pip install opencv-python-headless==4.10.0.82

Train

To train LUTFormer,

1. Edit the configuration file:
   • Open root/LUTFormer_code/config.py.
   • Set run_mode to 'train'. Also, set the task_name, dataset_name, and expert.
   • (Optional) If you want to visualize the results, set viz to True.
2. Run with

$ cd root/LUTFormer_code/
$ python main.py

Test

To evaluate your trained LUTFormer model,

1. Edit the configuation file:
   • Open root/LUTFormer_code/config.py.
   • Set run_mode to 'test' and viz to True.
   • Specify the values for task_name, dataset_name, and expert.
2. Run with

$ cd root/LUTFormer_code/
$ python main.py

3. Calculate the score using Matlab code
   • FiveK

     (matlab) > ./fivek_calculate_metrics.m [evaluate image dir] [GT dir]

   • PPR10K

     (matlab) > ./ppr10k_calculate_metrics.m [evaluate image dir] [GT dir] [mask dir]

*** Note ***
If you want to get the performance of the paper, set run_mode to 'test_paper'.
Pretrained models are available in root/LUTFormer_code/pretrained. They can also be downloaded from here.

Demo

You can run a demo with pretrained models to enhance your own images.

1. Prepare your input images
   • Place your images in the directory specified by --input_dir (default: root/LUTFormer_code/demo_img/input)
   • The enhanced results will be saved to --output_dir (default: root/LUTFormer_code/demo_img/result)
2. (Optional) Check configuration:
   • You can also override settings via command-line arguments, including:
     • --yaml_path (default: root/LUTFormer_code/configs/Retouching_FiveK.yaml)
     • --pretrained_path (default: root/LUTFormer_code/pretrained/Retouching_FiveK_expertC.pth)
     • --task_name, --dataset_name, --expert
3. Run demo with

$ cd root/LUTFormer_code/
$ python demo.py --input_dir ./demo_img/input --out_dir ./demo_img/result

Results

1. Photo retouching on FiveK dataset

2. Photo retouching on PPR10K dataset

3. Tone mapping on FiveK dataset

4. Underwater image enhancement on UIEB dataset