Disentangled and Controllable Face Image Generation via 3D Imitative-Contrastive Learning

Figure 1: This paper presents DiscoFaceGAN that generates realistic face images of virtual people with independent latent variables of identity, expression, pose, and illumination. The latent space is interpretable and highly disentangled, which allows precise control of the targeted images (e.g., degree of each pose angle, lighting intensity and direction), as shown in the top row. The bottom row shows the generated images when we keep the identity and randomize other properties. The faces generated by our method are not any real person in the world.