As a base of my prototype, I picked a Kenney asset pack. It looks cute. Includes most of the visuals I may need. But when I started to use it, I noticed a problem. Assets have lower resolution than needed.
While searching for a solution, I found an article related to this issue. According to that article, Waifu2x is a tool to magnify low-resolution non-vector images. I used it a couple of times, and the result was impressive.
As they say on their website, they use convolutional neural networks (CNN) for this magnification. The tool does its job very quickly. I am even jealous of this product. I can develop one with a CNN and an autoencoder. Most probably, it will be easy to create a dataset for it. (Greedy side project alert!)