[ICCV 2025 Highlight]⭐ISP2HRNet

PyTorch implementation of paper "ISP2HRNet: Learning to Reconstruct High Resolution Image from Irregularly Sampled Pixels via Hierarchical Gradient Learning".

Environment

We conduct our experiments on NVIDIA 3090 GPUs in an environment configured as follows:

• Python 3.9.0
• Pytorch 1.13.1 + cu116

Readers do not need to replicate our setup exactly.

Train

python train_liif.py --config configs/train-irregular/train_edsr-baseline-liif_irregular.yaml --name save_name --gpu gpu_id

Remember to modify the root_path in train_edsr-baseline-liif_irregular.yaml. All training parameters can be reconfigured in this file.

Reproducing Experiments

• EX1🌱Reconstruction from Incomplete Image with Random Missing Pixels

. test_ex1.sh

modify --root in the shell.

• EX2🌱Super Resolution from Incomplete Image with Random Missing Pixels

. test_ex2.sh

modify root_path_1 and root_path_2 in each config file (configs/test-irregular/test-*-*-rrs+sr.yaml).

• EX3🌱Reconstruction from Irregularly Sampled Pixels

. test_ex3.sh

modify --root in the shell.

• EX4🌱Super Resolution from a Fixed Number of Irregularly Sampled Pixels

. test_ex4.sh

modify root_path in the config file (configs/test-irregular/test-div2k-2_draw_arbi.yaml).

Acknowledgements

This code is built upon LIIF. We thank the authors for sharing the codes.