[WIRN 2024] A Wavelet Diffusion GAN for Image Super-Resolution

If you find this project useful, consider giving it a star ⭐

---

📰 News

• [2026] Conference Proceeding available 🎉
• [2024] Paper accepted at WIRN 2024 🎉
• [2024] Code and experiments released
• [2024] Presentation available here

---

😮 Highlights

⚡ Fast Diffusion for Super-Resolution

We propose a Diffusion GAN framework that significantly reduces the number of diffusion steps, making diffusion-based super-resolution much faster and more practical.

🌊 Wavelet-Based Representation

By leveraging the Discrete Wavelet Transform (DWT), the model operates in a compressed frequency domain, reducing computational cost while preserving fine details.

🖼️ High-Fidelity Image Reconstruction

Our method achieves superior perceptual quality and reconstruction fidelity, outperforming strong baselines such as SR3, DiWa, and ESRGAN.

---

🚀 Main Results

Quantitative Comparison

Metric	ESRGAN	SR3	DiWa	Ours
PSNR ↑	21.13	14.65	13.68	23.38
SSIM ↑	0.59	0.42	0.13	0.68
LPIPS ↓	0.082	0.365	0.336	0.061
FID ↓	20.8	99.4	270	47.2

---

⏱️ Efficiency Comparison

Model	Runtime	Parameters
ESRGAN	0.04s	31M
SR3	60.3s	98M
DiWa	34.7s	92M
Ours	0.12s	57M

---

🛠️ Installation

conda create --name=wadigan python=3.10
conda activate wadigan

pip install -r requirements.txt

---

📂 Dataset Preparation

We train on CelebA-HQ (16×16 → 128×128).

Download dataset:

• https://www.kaggle.com/datasets/badasstechie/celebahq-resized-256x256

Prepare dataset:

python datasets_prep/prepare_data.py \
  --path [dataset root] \
  --out [output root] \
  --size 16,128 -l

⚠️ Important:

• Images must be sequentially numbered for LMDB conversion.

Structure:

data/
└── celebahq_16_128

---

▶️ How to Run

We provide a unified script:

bash run.sh <DATASET> <MODE> <#GPUS>

Example:

bash run.sh celebahq_16_128 train 1

Arguments:

• <DATASET>: celebahq_16_128
• <MODE>: train / test
• <#GPUS>: number of GPUs

⚠️ Use the same --exp name for training and testing.

---

🧪 Evaluation

Metrics are computed on the full test set (6000 samples).

Generate Samples

bash run.sh celebahq_16_128 test 1

FID

Add:

--compute_fid --real_img_dir /path/to/real/images

PSNR / SSIM / LPIPS

python benchmark/eval.py -p [result root]

---

📦 Pretrained Models

Place checkpoints in:

saved_info/srwavediff/<DATASET>/<EXP>

---

📖 Method Overview

Our framework combines:

• Diffusion models (reduced timesteps)
• GAN training paradigm
• Wavelet-domain processing

This enables:

• Faster inference
• Lower memory usage
• Improved texture reconstruction

---

📚 Acknowledgements

• WaveDiff: https://github.com/VinAIResearch/WaveDiff
• SR3: https://github.com/Janspiry/Image-Super-Resolution-via-Iterative-Refinement
• DiWa: https://github.com/Brian-Moser/diwa
• ESRGAN: https://arxiv.org/abs/1809.00219
• BasicSR: https://github.com/XPixelGroup/BasicSR
• pytorch_wavelets: https://github.com/fbcotter/pytorch_wavelets

---

📌 Citation

@Inbook{aloisi2026wavelet,
author="Aloisi, Lorenzo
and Sigillo, Luigi
and Uncini, Aurelio
and Comminiello, Danilo",
editor="Esposito, Anna
and Faundez-Zanuy, Marcos
and Morabito, Francesco Carlo
and Pasero, Eros
and Cordasco, Gennaro",
title="A Wavelet Diffusion GAN for Image Super-Resolution",
bookTitle="Neural Networks: Overview of Current Theories and Applications",
year="2026",
publisher="Springer Nature Singapore",
address="Singapore",
pages="425--435",
isbn="978-981-95-4072-3",
doi="10.1007/978-981-95-4072-3_36",
url="https://doi.org/10.1007/978-981-95-4072-3_36"
}

---

⭐ Star History