Transferring hairstyles between portrait images utilizing deep latent representations

A system comprising:

• one or more memory devices comprising:
• a source image depicting a first person with a first hairstyle;
• and
• a target image depicting a second person with a second hairstyle;
• one or more computing devices configured to cause the system to:
• project the source image into a source latent vector utilizing a face-generative neural network;
• project the target image into a target latent vector utilizing the face-generative neural network;
• generate a hairstyle-transfer latent vector by:
• determining activation values from the source latent vector and the target latent vector that control the first hairstyle or the second hairstyle;
• and
• selectively combining the source latent vector and the target latent vector based on the activation values that control the first hairstyle or the second hairstyle;
• and
• generate a transferred hairstyle image that depicts the first person from the source image with the second hairstyle from the target image by synthesizing the hairstyle-transfer latent vector utilizing the face-generative neural network.

A computer-implemented method comprising:

• projecting a source image into a source latent vector utilizing a face-generative neural network, the source image depicting a first person with a first hairstyle;
• projecting a target image into a target latent vector utilizing the face-generative neural network, the source image depicting a second person with a second hairstyle;
• performing a step for generating a hairstyle-transfer latent vector that encodes the first person from the source image with the second hairstyle from the target image;
• and
• generating a transferred hairstyle image that depicts the first person from the source image with the second hairstyle from the target image by synthesizing the hairstyle-transfer latent vector utilizing the face-generative neural network.