Paper Link : https://arxiv.org/pdf/1609.04802.pdf

References :

1. https://github.com/sgrvinod/Deep-Tutorials-for-PyTorch
2. https://www.kaggle.com/code/balraj98/single-image-super-resolution-gan-srgan-pytorch/notebook
3. https://www.kaggle.com/code/luizclaudioandrade/srgan-pytorch-lightning
4. https://pytorch.org/tutorials/beginner/
5. https://github.com/aladdinpersson/Machine-Learning-Collection/tree/master/ML/Pytorch/GANs
6. https://www.youtube.com/playlist?list=PLhhyoLH6IjfwIp8bZnzX8QR30TRcHO8Va
7. https://paperswithcode.com/

Google Colab link for direct use of the notebook : https://tinyurl.com/tbp9tevr

Concepts involved : Residual Connections, Generative Adversarial Networks(GANs), Perceptual Loss

Note : This Model was trained on a small subset of CelebA Dataset(you can get the whole dataset from here: https://www.kaggle.com/datasets/jessicali9530/celeba-dataset) due to computational constraints.

A GAN type model is trained to convert low resolution images into high resolution. The high resolution images are used as taregts to train the Generator and as inputs for the Discriminator of SRGAN Model respectively. Low resolution versions of the high resolution images are produced by bicubic downsampling, and they are input to the Generator of SRGAN Model.

I trained on only 1000 images and you can get that custom dataset here : https://tinyurl.com/2xfe8u6j

Use 'making_lr_img.py' to create your own custom dataset of Low Resolution images. I foud it easy as compared to do it in transforms.Compose() of torchvision, but it's upto you.

Use 'pytorch_prediction.py' to get predicted image from the trained model. Don't forget to add the weights of your trained model in torch.load() and the weights must be of generator only (we don't require discriminator here, it's role is over as it was only required to improve the predictions of generator model so that generator can produce something desirable and not just random noise).

Use 'after_predictions.py' to visualize the predicted/generated image.

Use 'gradio_interface.py' to get a nice UI Interface for the model. After cloning the repo into your local machine or copying the files, run : python3 pytorch_prediction.py to see the nice UI on your local host.

I retrained the model with around 10,000 images of celebA dataset and got quite satisfying results. I've added the outputs in the zip file, left most one is the real hr image, middle one is lr image and rightmost one is generated by our model.

GRADIO INTERFACE LOOKS SOMETHING LIKE THIS: