<p align="center"> <img src="assets/icon.png" alt="icon" width="200px"/> </p>

(CVPR 2025) Adversarial Diffusion Compression for Real-World Image Super-Resolution [PyTorch]

Bin Chen^1,3,*
| Gehui Li^1,*
| Rongyuan Wu^2,3,*
| Xindong Zhang³
| Jie Chen^1,†
| Jian Zhang^1,†
| Lei Zhang^2,3

¹ School of Electronic and Computer Engineering, Peking University

² The Hong Kong Polytechnic University, ³ OPPO Research Institute

^* Equal Contribution. ^† Corresponding Authors.

⭐ If AdcSR is helpful to you, please star this repo. Thanks! 🤗

📝 Overview

Highlights

• Adversarial Diffusion Compression (ADC). We remove and prune redundant modules from the one-step diffusion network OSEDiff and apply adversarial distillation to retain generative capabilities despite reduced capacity.
• Real-Time Stable Diffusion-Based Image Super-Resolution. AdcSR super-resolves a 128×128 image to 512×512 in just 0.03s 🚀 on an A100 GPU.
• Competitive Visual Quality. Despite 74% fewer parameters 📉 than OSEDiff, AdcSR achieves competitive image quality across multiple benchmarks.

Framework

1. Structural Compression
   • Removable modules (VAE encoder, text prompt extractor, cross-attention, time embeddings) are eliminated.
   • Prunable modules (UNet, VAE decoder) are channel-pruned to optimize efficiency while preserving performance.

<p align="center"> <img src="assets/teaser.png" alt="teaser" width="55%"/> </p>

2. Two-Stage Training
   1. Pretraining a Pruned VAE Decoder to maintain its ability to decode latent representations.
   2. Adversarial Distillation to align compressed network features with the teacher model (e.g., OSEDiff) and ground truth images.

<p align="center"> <img src="assets/method.png" alt="method" /> </p>

😍 Visual Results

<img src="assets/demo1.png" height="240px"/> <img src="assets/demo2.png" height="240px"/> <img src="assets/demo3.png" height="240px"/>

<img src="assets/demo4.png" height="242px"/> <img src="assets/demo5.png" height="242px"/> <img src="assets/demo6.png" height="242px"/>

<img src="assets/demo7.png" height="319px"/> <img src="assets/demo8.png" height="319px"/>

https://github.com/user-attachments/assets/1211cefa-8704-47f5-82cd-ec4ef084b9ec

<img src="assets/comp.png" alt="comp" width="840px" />

⚙ Installation

shellCopy
git clone https://github.com/Guaishou74851/AdcSR.git
cd AdcSR
conda create -n AdcSR python=3.10
conda activate AdcSR
pip install --upgrade pip
pip install -r requirements.txt
chmod +x train.sh train_debug.sh test_debug.sh evaluate_debug.sh

⚡ Test

1. Download test datasets (DIV2K-Val.zip, DRealSR.zip, RealSR.zip) from Hugging Face or PKU Disk.
2. Unzip them into ./testset/, ensuring the structure:

   ./testset/DIV2K-Val/LR/xxx.png
   ./testset/DIV2K-Val/HR/xxx.png
   ./testset/DRealSR/LR/xxx.png
   ./testset/DRealSR/HR/xxx.png
   ./testset/RealSR/LR/xxx.png
   ./testset/RealSR/HR/xxx.png
   

3. Download model weights (net_params_200.pkl) from the same link and place it in ./weight/.
4. Run the test script (or modify and execute ./test_debug.sh for convenience):

   bashCopy
   python test.py --LR_dir=path_to_LR_images --SR_dir=path_to_SR_images

   The results will be saved in path_to_SR_images.
5. Test Your Own Images:
   • Place your low-resolution (LR) images into ./testset/xxx/.
   • Run the command with --LR_dir=./testset/xxx/ --SR_dir=./yyy/, and the model will perform x4 super-resolution.

🍭 Evaluation

Run the evaluation script (or modify and execute ./evaluate_debug.sh for convenience):

bashCopy
python evaluate.py --HR_dir=path_to_HR_images --SR_dir=path_to_SR_images

🔥 Train

This repo provides code for Stage 2 training (adversarial distillation). For Stage 1 (pretraining the channel-pruned VAE decoder), refer to our paper and use the code of Latent Diffusion Models repo.

1. Download pretrained model weights (DAPE.pth, halfDecoder.ckpt, osediff.pkl, ram_swin_large_14m.pth) from Hugging Face or PKU Disk, and place them in ./weight/pretrained/.
2. Download the LSDIR dataset and store it in your preferred location.
3. Modify the dataset path in config.yml:

   yamlCopy
   dataroot_gt: path_to_HR_images_of_LSDIR

4. Run the training script (or modify and execute ./train.sh or ./train_debug.sh):

   bashCopy
   CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.run --nproc_per_node=8 --master_port=23333 train.py

   The trained model will be saved in ./weight/.

🥰 Acknowledgement

This project is built upon the codes of Latent Diffusion Models, Diffusers, BasicSR, and OSEDiff. We sincerely thank the authors of these repos for their significant contributions.

🎓 Citation

If you find our work helpful, please consider citing:

latexCopy
@inproceedings{chen2025adversarial,
  title={Adversarial Diffusion Compression for Real-World Image Super-Resolution},
  author={Chen, Bin and Li, Gehui and Wu, Rongyuan and Zhang, Xindong and Chen, Jie and Zhang, Jian and Zhang, Lei},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2025}
}