Face Image Editing And Synthesis Based On The Latent Space Of Generator

Face image editing and synthesis have a wide range of applications in entertainment and social security.In view of the advantages of the generative adversarial network(GAN)model which can synthesize high-quality images,many researchers realized the generation of a target face by editing the latent code in recent years.For example,by interpreting the latent space of the generator,it can take the latent code as input of the generator to synthesize the modified target faces based on editing the latent code with a solved semantic control vector.For the problem that existing algorithms use a single control vector to edit the latent code will introduce changes of irrelevant attributes,and the difficulty of synthesizing a face consistent with the target face in the user’s mind from scratch,we propose the following algorithm: 1)A decoder is constructed to predict the independent control vectors of each face for a specific attribute,it can achieve a precise attribute editing effect by editing the latent code with independent control vectors.2)Based on the reference faces selected by the user from the database,fusing the features of the reference faces and synthesizing new faces.According to the user’s judgment that whether the generated face is similar to the target or not to modify the generated result,the target face is generated iteratively in the way of selection-fusionfeedback interaction when the user cannot specify the facial detail.The experimental results show that the computed adaptive attribute control vector can be used for the precise editing of face attributes,which keeping the background and other irrelevant attributes unchanged.In cooperation with the interactive face synthesis algorithm,the interactively synthesized faces are consistent with the target faces in terms of face shape,contour,emotion,and the five sense organs,that achieve overall similarity.