Single Image x4 Resolution (x4 each dimension, thus x16)

I used works by Dong, et al 2015

The model is simple, with 3 layers. Each is a conv2d layer, with Relu, except final layer is sigmoid.
First is 64x9x9, then 32x5x5, then 3x5x5. No padding.

I used perceptionloss with MobileNet. I chose it since it is small. In other words, I could gain performance with better model like Xception.

Trained using Animal Faces on Google Colab (the notebook is also included).
But I tested using my personal machine with Python 3.7, keras 2.1.3 and tensorflow 2.1. Opencv-python and numpy is also needed, any version would suffice. I did edit a bit of code so that it could run on my personal laptop (for example changing the backend sigmoid to built-in activation for conv2d), but I did make sure that training and prediction code runs perfectly.

Usage is simple, look at prediction.py.

Ofc, this is no SOTA. It is just spare time project. What is cool is that the model is small, the weight, with the model is < 1mb. Ideal for niche scenario.
Output image might have artifacts, the usual convolution blocky stuff. Also 16 pixels from all size will be cropped. I am too lazy to add padding.

Here are some output, there is more in example_input. From the order of input, prediction, and truth.