Facial Generation, Recognition, & Reconstruction

Using Nvidia’s StyleGAN2 code run in Google Colab, Figure 1 provides an uncurated look at nine randomly generated images. It is worth noting the outstanding true to life results. Row 1, column 3 has noticeably fallen victim to artifacts. Artifacts are unwanted objects or noise that are generated by mistake through the code. The technology is by no means flawless, but definitely on a fascinating path. Certain logical fallacies are understood by the code, for example, a man with long hair would not be generated using StyleGAN2. 

​

The code separates image variables into two groups: global and stochastic variables. Global variables are things like pose or color, which can easily be addressed. Stochastic variables present a different type of challenge, as they are measurable, but not predictable. These are instances like the distribution of hair on the subject's head, or the wrinkles in their clothing. These features are guided by a standard deviation curve that dictates the distribution of the stochastic variable. 

​

While training this code is very demanding in terms of computational power, the previously mentioned Google Colab accessible pretrained model circumvents this main barrier to entry. It is possible to generate faces of people who have never existed on your laptop or phone.

​

In theory, general adversarial networks could be able to account for the generation of facial features even with a facial mask, as various instances would be interpreted by the software and realised in the output. 

We will be testing the adverse impact of facial obstructions on general adversarial networks in the following weeks.

​

Recently, the technology was used to visualise the death toll of COVID-19 in the United States