Creating data visualizations is a lossy process, a user’s code and data is compiled in the form of a static image. These images are the default form of data visualizations, and are ubiquitous across both social media and academia.
As visualization researchers, this lossy compilation into images makes it very challenging to experiment and tweak with design decisions of the original author.
In this work, we present a novel deep-learning based algorithm that can decompile an image visualization, allowing researchers to easily edit and inspect its visual design. Our method and domain-specific language is more general than previous approaches and is extensible to more types of data visualizations without the need for expert-tuned heuristics.