Fig. 4 

Face images generated on the test dataset. In each panel from left to right are facial-part patches, our synthesized facial images and the original facial images (the ground truth)