The objective of this project was to generate high resolution photos that looked cool; something you would use for your wallpaper. I created this during my senior year of high school for Advanced Studies. I started this project with almost zero knowledge of machine learning, and I learned so much about python, machine learning, Keras, Tensorflow, model architectures, and convolutional neural networks.

I had about 6 months to work on this project before I had to present it to a group of people in the industry, so I tried my hardest to create something that was simple to set up and reproduce. Unfortunately, I spent too much time training, and I didn't really refine the model's architecture, so I wasn't able to create good results.

There are 2 parts to deep-wallpaper: a image generator, and an image upscaler. The generator is an ACGAN, with one input for entropy, and another for image tags, and outputs an RGB image size 128x72. I wanted to be able to specify what kind of image I wanted to generate. The upscaler is a DCNN, taking a 32x32 image and upscaling it to 64x64. To train it, I created a webscraper to download high resolution images, then I created a couple of tools to help tag the images.

The source code is available on my GitHub, and model checkpoints are available upon request.