StarGAN Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation
Published:
scalable i2i translation on more than two domains
- single generator learns mapping between all domains
- inputs image and domain information
- model
- discriminator
- GAN loss, domain classification loss
- WGAN with gp, patchGAN
- GAN loss, domain classification loss
- generator
- GAN loss, reconstruction loss
- discriminator
- label construction
- c = [c_1, …, c_n, m], m being mask, eg. [00101 00000 10]