Fig. 16:

Generating high-resolution high-fidelity multi-domain and multimodal images via sampling from the latent space of StyleGAN2-FFHQ [9]. In each example, upper left shows the image generated by originally encoded $W^{+}$. The rest three RGB images are the corresponding multimodal generations obtained by replacing the 5-th to 18-th (left) and 7-th to 18-th (right) $W^{+}$ of StyleGAN2 [9] with random variables.