University of Pittsburg researchers have come up with a conditioner variational autoencoder to generate unique faces for advertisement. This is a result of the further extension of their previous study which tried to understand using automated methods an advertisement. It tried to decipher the complicated visual rhetoric in advertisements.
In their latest project, the researchers have leveraged machine learning to create faces that are persuasive and would work well for various kinds of advertisements. To that end they banked upon conditional variational autoencoders, also known as “generative models” that can recreate synthetic data similar to that it is trained on.
An autoencoder takes an image and learns to represent it as a few numbers. The decoder, on the other hand, learns those numbers and recreates the original image from it. One can almost imagine it to be a large image represented by a few numbers.
However, imparting training to generative models to create computer vision can be a daunting task and needs massive sets of data, namely ads. To tackle the problem, the team of researchers used an autoencoder that needed lesser amount of data and could also deal with the substantial variance in advertisements.
The main aim of the advertising industry is to pitch their products to consumers through persuasive languages and visuals. Hence human faces are an important component of advertisements.
The findings of the team of researchers provide a means to generate tailor-made and targeted ads that are meant for specific sets of individuals. For example, they could create faces matching those of the viewer.