Optimizing Unsupervised Generative Adversarial Networks via Latent Space Regularizations

A system for training a generator G of Generative Adversarial Network (GAN), comprising:

• a processor;
• and
• a memory, the processor is configured to executed instructions stored in the memory to:
• generate, by the generator G and in response to receiving a first input Z, a first output G(Z), wherein the first output G(Z) being a first ambient space representation of the first input Z;
• generate, by an encoder E of the GAN and in response to receiving the first output G(Z) as input, a second output E(G(Z)), wherein the second output E(G(Z)) being a first latent space representation of the first output G(Z);
• generate, by the generator G and in response to receiving the second output E(G(Z)) as input, a third output G(E(G(Z))), wherein the third output G(E(G(Z))) being a second ambient space representation of the second output E(G(Z));
• generate, by the encoder E and in response to receiving the third output G(E(G(Z))) as input, a fourth output E(G(E(G(Z)))), wherein the fourth output E(G(E(G(Z)))) being a second latent space representation of the third output G(E(G(Z)));
• train the encoder E to minimize a difference between the second output E(G(Z)) and the fourth output E(G(E(G(Z))));
• and
• use the second output E(G(Z)) and fourth output E(G(E(G(Z)))) to constrain a training of the generator G.

A method for style transfer, comprising:

• receiving, by a generator G, an input corresponding to a source image including a first participant;
• and
• outputting, from the generator G, an output image, wherein the first participant is modified in the output image, wherein the generator G is trained using a Generative Adversarial Network (GAN) by steps comprising:
• generating, by the generator G and in response to receiving a first input Z, a first output G(Z), wherein the first output G(Z) being a first ambient space representation of the first input Z;
• generating, by an encoder E of the GAN and in response to receiving the first output G(Z) as input, a second output E(G(Z)), wherein the second output E(G(Z)) being a first latent space representation of the first output G(Z);
• generating, by the generator G and in response to receiving the second output E(G(Z)) as input, a third output G(E(G(Z))), wherein the third output G(E(G(Z))) being a second ambient space representation of the second output E(G(Z));
• generating, by the encoder E and in response to receiving the third output G(E(G(Z))) as input, a fourth output E(G(E(G(Z)))), wherein the fourth output E(G(E(G(Z)))) being a second latent space representation of the third output G(E(G(Z)));
• training the encoder E to minimize a difference between the second output E(G(Z)) and the fourth output E(G(E(G(Z))));
• and
• using the second output E(G(Z)) and fourth output E(G(E(G(Z)))) to constrain a training of the generator G.