Colorizing Mobile Websites

Today I learned, via this paper, that colorization tasks go beyond photos: people are training models to colorize mobile websites!

When I first heard about it, the task didn't seem that hard. But then I saw the summary table and it made me rethink.

The paper does a good job at explaining the problem a bit more and they also describe how they've built a model for this task. They also go the extra step by also providing a dataset and by introducing a new metric. Turns out that color distance is a unique thing to train for, so they introduce "Frechet Color Distane". They also investigate if the generated results adhere to web accessibility guidelines with regards to contrast violations via Google Lighthouse. This results in the following table:

I can see how these kinds of models might have some applications in automated website design, but the authors also mention some weaknesses.

There is still room for improvement regarding the dataset. The e-commerce mobile web pages used in our study have low relevance between individual product im- ages and color styles, making them unsuitable for evaluat- ing the harmony between images and colors. Also, it has not been tested whether our methods can handle more ele- ments, such as full-page view on a mobile-sized screen or on a laptop-sized screen.

Interesting though, I had not imagined that colorization would enter the realm of websites.