Home

Awesome

Swift-SRGAN - Rethinking Super-Resolution for real-time inference

This repository is the official implementation of the paper "Swift-SRGAN - Rethinking Super-Resolution for real-time inference" https://arxiv.org/abs/2111.14320

Architecture

<p align="center"> <img src="https://github.com/Koushik0901/Swift-SRGAN/blob/master/image-samples/SwiftSRGAN-architecture.png" width="850" height="450" /> </p>

Super-Resolution Examples

<p align="center"> <b><i>All images on the left side are the original high resolution images and images on the right side are the 4x super-resolution output from our model.</i></b> <p align="center"> <img src="https://github.com/Koushik0901/Swift-SRGAN/blob/master/image-samples/4x_samples/baryon.png" width="800" height="400" /> </p> <p align="center"> <img src="https://github.com/Koushik0901/Swift-SRGAN/blob/master/image-samples/4x_samples/dwight.png" width="800" height="400" /> </p> <p align="center"> <img src="https://github.com/Koushik0901/Swift-SRGAN/blob/master/image-samples/4x_samples/steve.png" width="800" height="450" /> </p>

Pre-trained Models

Check the releases tab for pre-trained 4x and 2x upsampling generator models

Training

  1. install requirements with: pip install -r requirements.txt

  2. Train the model by executing:

    cd swift-srgan
python train.py --upscale_factor 4 --crop_size 96 --num_epochs 100

  3. To convert the generator model to torchscript, run

python optimize-graph.py --ckpt_path ./checkpoints/netG_4x_epoch100.pth.tar --save_path ./checkpoints/optimized_model.pt --device cuda

Please cite our article

 @article{krishnan2021swiftsrgan,
 title={SwiftSRGAN--Rethinking Super-Resolution for Efficient and Real-time Inference},
 author={Krishnan, Koushik Sivarama and Krishnan, Karthik Sivarama},
 journal={arXiv preprint arXiv:2111.14320},
 year={2021}
 }