GitHub - HaomingCai/CUGAN: [ Official ] - Toward Interactive Modulation for Photo-Realistic Image Restoration. CVPRW 2021 NTIRE. (original) (raw)

Toward Interactive Modulation for Photo-Realistic Image Restoration. (Paper Link)

By Haoming Cai*, Jingwen He*, Yu Qiao, and Chao Dong in CVPRW, NTIRE workshop 2021.

Two-dimension Modulation

Real-World Modulation

Dependencies and Installation

• pip install -r requirements.txt
• Python 3 (Recommend to use Anaconda)
• PyTorch >= 1.3
• NVIDIA GPU + CUDA

How to Test

• Prepare the test dataset
  1. Download LIVE1 dataset and CBSD68 dataset from Google Drive
  2. Generate LQ images with different combinations of degradations using matlab codes/data_scripts/generate_2D_val.m.
• Download the pretrained model
  1. Download pretrained CUGAN from Google Drive
  2. Modify the pretrain_model_G in configuration file options/test/xxxxxx.yml.
• Test CUGAN with range of restoration strength
  1. Modify the configuration file options/test/modulation_CUGAN.yml. ❗️Importantly, cond_init, range_mode, range_stride are crucial in this testing mode.
  2. Run command:
     cd codes
     python test-cugan_range-cond.py -opt options/test/modulation_CUGAN.yml
• Test CUGAN with specific restoration strength
  1. Modify the configuration file options/test/test_CUGAN.yml. ❗️Importantly, cond is crucial in this testing mode.
  2. Run command:
     cd codes
     python test-cugan_specific-cond.py -opt options/test/test_CUGAN.yml

How to Train

• Cooming Soon

Acknowledgement

• This code is based on mmsr, BasicSR and CResMD.