Skip to content

Face synthesis

Face synthesis refers to the generation of photorealistic facial images and videos that do not correspond to real people, created using generative models such as Generative Adversarial Networks (GANs). Unlike face-swapping or expression transfer (which manipulate existing faces), face synthesis creates entirely new facial content from scratch.

Techniques

StyleGAN and variants: Modern state-of-the-art GAN architectures that enable fine-grained control over generated facial attributes through latent space manipulation. StyleGAN achieves an unsupervised separation of high-level facial attributes.

ProGAN: Progressive training of GANs allowing generation of high-resolution (1024×1024) photorealistic face images with improved stability and convergence.

Conditional GANs: GANs trained to generate faces conditioned on attribute vectors, enabling controlled synthesis of faces with specified characteristics.

Key papers