GitHub - Koushik0901/Swift-SRGAN: This repository is the official implementation of the paper "Swift-SRGAN - Rethinking Super-Resolution for real-time inference" https://arxiv.org/abs/2111.14320 (original) (raw)

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

Super-Resolution Examples

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.

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} }