Before text to image generation and Diffusion models hit the world by storm, Generative Adversarial Networks (GAN) were one of the most hyped generative models out there. The hype around GANs was not there without a reason. The quality of data generated by GANs was very impressive at that time. Generated images were crisp, didn’t lack high frequency details, and had a decent amount of variety (see StyleGan). Moreover, later people discovered methods enabling the usage of GANs as powerful image editing tools (see StyleGan Inversion). The generation could be conditioned well enough to produce certain desired results and much more (see pix2pix). To this date, GANs can hold their own in restricted domain scenarios and have their positives compared to the latest powerful Diffusion models.

Many real-world applications require processing and understanding of graph-structured data. For instance, one may want to process World Wide Web data, social networks, citation networks, scene graphs, human bodies, and many more. What unites these kinds of data is the availability of neighbourhood information amongst entities. This information can serve as a very strong end informative inductive bias. Thus, it is worthwhile exploring neural architectures that can embed this information in the processing. This is where Graph Neural Networks play their part in the abundance of neural architectures.