Neural Photo Editing with Introspective Adversarial Networks

摘要: The increasingly photorealistic sample quality of generative image models suggests their feasibility in applications beyond generation. We present the Neural Photo Editor, an interface that leverages power neural networks to make large, semantically coherent changes existing images. To tackle challenge achieving accurate reconstructions without loss feature quality, we introduce Introspective Adversarial Network, a novel hybridization VAE and GAN. Our model efficiently captures long-range dependencies through use computational block based on weight-shared dilated convolutions, improves generalization performance with Orthogonal Regularization, weight regularization method. validate our contributions CelebA, SVHN, CIFAR-100, produce samples high visual fidelity.