I’d like to share this mini Dash app (~120 lines) I’ve been working on. It first uses Tensorflow Hub to load the ESRGAN model, and let you upload an image and enhancing it with 4x upsampling.
- Colab Demo (simplest and fastest way to try this app, as well as modifying it).
- Source Code (if you want to run it locally with a Nvidia GPU).
ESRGAN is a type of Generative Adversarial Network trained to upscale smaller images by “faking” the new pixels such that another neural network would not be able to distinguish such images from the real ones. If you are not familiar with GANs, check out this article we wrote recently about them.
Here’s some supplementary details from the readme.
For optimal speed, we recommend running this app using a Nvidia GPU with sufficient memory (e.g. V100). To try this app, please use Google Colab with a GPU accelerator (access the notebook by clicking the badge above).
This works pretty well for very small images, such as texture from older games, or scenes of nature. However, it might not work so well with human faces. We also ask you to be careful about potentially harmful biases when using this model with pictures of people, considering that this model was not thoroughly validated in terms of biases.