Enjoy Your Editing: Controllable GANs for Image Editing via Latent Space Navigation

Controllable semantic image editing enables a user to change entire image\nattributes with a few clicks, e.g., gradually making a summer scene look like\nit was taken in winter. Classic approaches for this task use a Generative\nAdversarial Net (GAN) to learn a latent space and suitable latent-space\ntransformations. However, current approaches often suffer from attribute edits\nthat are entangled, global image identity changes, and diminished\nphoto-realism. To address these concerns, we learn multiple attribute\ntransformations simultaneously, integrate attribute regression into the\ntraining of transformation functions, and apply a content loss and an\nadversarial loss that encourages the maintenance of image identity and\nphoto-realism. We propose quantitative evaluation strategies for measuring\ncontrollable editing performance, unlike prior work, which primarily focuses on\nqualitative evaluation. Our model permits better control for both single- and\nmultiple-attribute editing while preserving image identity and realism during\ntransformation. We provide empirical results for both natural and synthetic\nimages, highlighting that our model achieves state-of-the-art performance for\ntargeted image manipulation.\n