VecGAN: Image-to-Image Translation with Interpretable Latent Directions

AbstractWe propose VecGAN, an image-to-image translation framework for facial attribute editing with interpretable latent directions. Facial task faces the challenges of precise controllable strength and preservation other attributes image. For this goal, we design by space factorization each attribute, learn a linear direction that is orthogonal to others. The component change, scalar value. In our framework, can be either sampled or encoded from reference image projection. Our work inspired works fixed pretrained GANs. However, while those models cannot trained end-to-end struggle edit images precisely, VecGAN successful at preserving extensive experiments show achieves significant improvements over state-of-the-arts both local global edits.KeywordsImage translationGenerative adversarial networksLatent manipulationFace