Evolution of the intermediate G(w,n) results throughout the optimization procedure. The leftmost image shows the input I. The next examples show the intermediate results for G(w,n), where the numbers at the bottom correspond to the current optimization iteration of w and n, that is, w_{itr} and n_{itr}, respectively. The initial image (w_{itr}=0,n_{itr}=0) is always the same. During the w latent code optimization procedure, the face shape and target-attribute appearance is adjusted in accordance with the considered constraints. After convergence, the noise component is optimized to introduce realistic fine image details. The final optimization result I_{G} is then blended based on B to generate the output image I'. The attributes edited in the presented examples are Black hair (top) and Big nose (bottom).
