GANravel: User-Driven Direction Disentanglement in Generative Adversarial Networks

Figure 1: GANravel enables users to disentangle editing directions in generative adversarial networks (GAN) using global and local disentanglement approaches. (a) A direction is often entangled when created by selecting exemplary images from the gallery. (b) The weights of the exemplary images can be adjusted to disentangle global attributes such as age and gender. (c) The direction can be tested on the live-testing section using multiple test images. (d) The user can hover over an exemplary image to see its weight and go back and forth between weight adjustments and live-testing until global attributes are disentangled. (e) The user can use masks to disentangle local attributes such as glasses and closed mouth. (f) The masks can be combined to either preserve or discard a region of interest and they can be tested. (g) Resulting disentangled direction can be applied to other test images in the live-testing section. (h) The final disentangled direction can be saved and applied in other future images.