pairgen ✂️

pairgen is a fast, reliable, and dependency-light CLI tool designed to generate High-Resolution (HR) and Low-Resolution (LR) image pairs for Super-Resolution Machine Learning and Deep Learning datasets.

Whether you are preparing standard academic benchmarks (like Set5, Set14, DIV2K) or generating complex Real-World SR validation sets, pairgen handles exact interpolation, patch extraction, and advanced degradations out of the box.

✨ Features

• 🔬 Academic Reproducibility: Includes a pure NumPy implementation of MATLAB's imresize (bicubic interpolation with antialiasing), which is strictly required for evaluating standard SR models.
• 🌪️ BSRGAN-like Degradation Pipeline: Simulates real-world image degradation by applying random Gaussian/Sinc blur, Gaussian/Poisson noise, and JPEG compression in a randomized order (Random Shuffle Strategy).
• ✂️ Smart Patch Extraction: Easily extract multiple random crops from large images to maximize your dataset's utility.
• ⚡ Multiprocessing: Built with ProcessPoolExecutor to utilize all CPU cores, effortlessly processing thousands of high-resolution images.
• 🪶 Lightweight: No heavy dependencies like OpenCV or SciPy. All complex FFT math and kernels are implemented using numpy and Pillow.

🎯 Motivation

While working on Image Super-Resolution projects, preparing testing and validation datasets is always a bottleneck. Academic benchmarks require exact MATLAB-like downsampling, while Real-World SR evaluation requires complex degradation pipelines.

Applying degradations on-the-fly during training is a standard practice, but for Validation and Test sets, you need strictly fixed, pre-generated LR images to reliably compare models and calculate PSNR/SSIM metrics. pairgen was created to standardize this offline generation process into one CLI command, perfectly complementing tools like manigen.

📦 Installation

You can install pairgen directly from PyPI using pip:

pip install pairgen

Or, if you use uv (recommended for CLI tools):

uv tool install pairgen

🚀 Quick Start

Generate a standard benchmark dataset with MATLAB bicubic downsampling (x4 scale):

pairgen -i data/Set14 -o data/Set14_pairs -s 4

💡 Advanced Usage Examples

1. Generating Real-World SR Validation Sets

Create a complex, degraded test set by enabling the BSRGAN-like pipeline (blur, noise, and JPEG compression applied in random order):

pairgen -i data/validation -o data/validation_degraded -s 4 --blur --noise --jpeg

2. Extracting Multiple Patches

If you have 2K/4K images and want to extract 50 random 256x256 patches from each image to build a rich dataset:

pairgen -i data/DIV2K_train -o data/DIV2K_patches -s 4 -p 256 -np 50

3. Combining with manigen

pairgen natively supports reading file manifests (.txt files containing lists of paths). You can use manigen to index your dataset and split it, then pass the manifest to pairgen:

pairgen -i train_manifest.txt -o data/train_set -s 4

🛠️ CLI Reference

🤝 Contributing

1. Clone the repository

git clone https://github.com/ash1ra/pairgen
cd pairgen

2. Install dependencies using uv

uv sync
# On Windows
.venv\Scripts\activate
# on Unix or MacOS
source .venv/bin/activate

3. Format and lint the code

uv run ruff format .
uv run ruff check .

4. Run the tests

uv run pytest tests/ -v

5. Submit a pull request

If you'd like to contribute, please fork the repository and open a pull request to the main branch.