UHDDIP: Ultra-High-Definition Image Restoration: New Benchmarks and A Dual Interaction Prior-Driven Solution [IEEE TCSVT2025]

Updates

[2025-09-18] Our paper is accepted to IEEE TCSVT !

Our proposed UHD-Snow dataset and UHD-Rain dataset

Download Link	Description
Google Drive (https://drive.google.com/drive/u/1/folders/1LaQvEBdjH5MwTwkfCZh3UJUGl8mYfvke?hl=zh_CN))	A total of 3000 pairs for training and 200 pairs for testing.

UHDDIP Framework

Requirements

• CUDA 10.1 (or later)
• Python 3.9 (or later)
• Pytorch 1.8.1 (or later)
• Torchvision 0.19
• OpenCV 4.7.0
• tensorboard, skimage, scipy, lmdb, tqdm, yaml, einops, natsort

Data preparation

1. Please download the following datasets:

Task	Training dataset	Testing dataset
UHD LLIE	UHD-LL (2000)	UHD-LL (150)
UHD Desnowing	UHD-Snow (3000)	UHD-Snow (200)
UHD Deraining	UHD-Rain (3000)	UHD-Rain (200)
UHD Dehazing	UHD-Haze (2290)	UHD-Haze (231)
UHD Deblurring	UHD-Blur (1964)	UHD-Blur (300)

2. To obtain the normal prior images, all training data and testing data are processed through Omnidata, and the obtained normal prior images are placed under the same path as the corresponding input image and GT image above.

Training

1. To train UHD LLIE model, modify the path where the UHD-LL dataset (including input, gt, normal) is located in the /src/Options/train_UHDDIP_LLIE.yml, then run

   cd UHDDIP
   python -m torch.distributed.launch --nproc_per_node=2 --master_port=4321 basicsr/train.py -opt src/Options/train_UHDDIP_LLIE.yml  --launcher pytorch
   

2. To train UHD desnowing/deraining/dehazing model, modify the path where the UHD-Snow/UHD-Rain/UHD-Haze dataset (including input, gt, normal) is located in the /src/Options/train_UHDDIP.yml, then run

   cd UHDDIP
   python -m torch.distributed.launch --nproc_per_node=2 --master_port=4321 basicsr/train.py -opt src/Options/train_UHDDIP.yml  --launcher pytorch
   

3. To train UHD deblurring model, modify the path where the UHD-Blur dataset (including input, gt, normal) is located in the /src/Options/train_UHDDIP_deblur.yml, then run

   cd UHDDIP
   python -m torch.distributed.launch --nproc_per_node=2 --master_port=4321 basicsr/train.py -opt src/Options/train_UHDDIP_deblur.yml  --launcher pytorch
   

Testing

1. Download all pre-trained models and place them in ./pretrained_models/

2. Testing

   cd src
   python test_uhd.py
   

3. Calculating PSNR/SSIM/LPIPS scores, run

   python calculate_psnr_ssim.py
   

Pre-trained models and visual results

Task	Model	Visual Results
UHD LLIE	Download	Download
UHD Desnowing	Download	Download
UHD Deraining	Download	Download
UHD Dehazing	Download	Download
UHD Deblurring	Download	Download

UHD Image LLIE (click to expand)

UHD Image Dehazing (click to expand)

UHD Image Deblurring (click to expand)

UHD Image Desnowing (click to expand)

UHD Image Deraining (click to expand)