Figure 1.

Manipulating facial expressions in images of people with genetic conditions via HyperStyle and StyleGAN2. We started with our real images, and manually classified these facial expressions into smile versus no smile. We finetuned HyperStyle, where the goal is to find an image embedding that can reconstruct the original input image (via StyleGAN2). Next, we manipulated the embedding of the reconstructed image by using SVM as shown in InterfaceGAN (Pisner and Schnyer, 2020). Finally, we passed the newly edited embedding into StyleGAN2 to create a new image with a new facial expression that still closely resembles the original input. On the far right, the examples illustrate the conditions we analyzed. All appropriate permissions have been obtained to reuse images and the citations for each are provided in Supplementary Table 5.