passerer / hpinet

Hierarchical Pixel Integration for Lightweight Image Super-Resolution

From Coarse to Fine: Hierarchical Pixel Integration for Lightweight Image Super-Resolution

This is the Pytorch implementation of HPINet (AAAI2023 Acceptance).

[arXiv paper] [pretrained models] [visual results (code:z86e)]

Dependencies

• python == 3.6
• pytorch == 1.8.0
• scikit-image == 0.17.2
• einops == 0.3.2

Other versions of the packages may also work, but they are not tested.

Test Datasets

• Download the five test datasets (Set5, Set14, B100, Urban100, Manga109) from Google Drive
• Put them in benchmarks/ directory, or customize the path in the code

Run Test

python test.py --model L --scale 4

The pretrained models are provided in checkpoints/:

	x2	x3	x4
S			✔
M	✔	✔	✔
L			✔

Train Datasets

DIV2K homepage

DIV2K
├── DIV2K_train_HR
├── DIV2K_train_LR_bicubic
│   ├── X2
│   ├── X3
│   ├── X4

Run Train

Here is an example:

CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" python train.py --model M --root DIV2K/ --ext .png --scale 4