Multi-Scale Adversarial Diffusion Network for Image Super-Resolution

This repository contains the official PyTorch implementation of MSADN.

🔧 Environment Requirements

• Python == 3.7.15
• PyTorch == 1.12.1
• CUDA == 11.3

📁 Dataset Preparation

We use the following datasets:

• FFHQ (1024x1024)
• CelebA-HQ (256x256)
• DIV2K

Preprocessing

1. FFHQ / CelebA-HQ:

   • For the CelebA-HQ dataset, 100 images were selected based on the evaluation method of GLEAN.
   • Resize images directly to 128×128 and 16×16.

2. DIV2K:

   • Crop training and validation images.
   • Then downsample to get 16×16 and 128×128 versions.
   • For details, please refer to the original paper.

🚀 Training

Pretrained Module

Before training, please refer to GLEAN to obtain the pre-trained style modulation model that you need to use.

Running

1. Set Parameters:

   • Modify parameters in train.py or use command-line arguments.

2. Launch Training:

python train.py

📥 Pretrained Weights

Pretrained model weights are available via Baidu Netdisk: 🔗 Download Link

📜 Citation

If you find our work helpful, please consider citing it:

@article{shi2025multi,
  title={Multi-scale adversarial diffusion network for image super-resolution},
  author={Shi, Yanli and Zhang, Xianhe and Jia, Yi and Zhao, Jinxing},
  journal={Scientific Reports},
  volume={15},
  number={1},
  pages={11690},
  year={2025},
  publisher={Nature Publishing Group UK London}
}