How does it feel to pass through a neural network?

An exploration of convolutional hallucinations.



Step 1: Glitched face using the StyleGAN FFHQ latent space


Step 2: Salient-object detection with U^2-Net 

My face is plunged into the latent space of a model trained on faces (StyleGAN FFHQ). A random masked is applied between my original face and my hallucinated face (see Step 1). Then, salient-object detection is performed in order to keep the main features. The result is a black and white image (see Step 2). Finally, 3D photo inpainting is used to bring the image to life (see below).