KG-GAN: Knowledge-Guided Generative Adversarial Networks
This addresses a limitation in generative models for image synthesis, enabling more flexible and knowledge-driven generation, though it appears incremental as it builds on existing GAN frameworks.
The paper tackles the problem of GANs failing to generate unseen categories, such as roses of different colors from only red roses, by proposing KG-GAN, which fuses domain knowledge with GANs using two generators and a constraint function, resulting in effective generation of unseen flower categories from seen ones given textual descriptions.
Can generative adversarial networks (GANs) generate roses of various colors given only roses of red petals as input? The answer is negative, since GANs' discriminator would reject all roses of unseen petal colors. In this study, we propose knowledge-guided GAN (KG-GAN) to fuse domain knowledge with the GAN framework. KG-GAN trains two generators; one learns from data whereas the other learns from knowledge with a constraint function. Experimental results demonstrate the effectiveness of KG-GAN in generating unseen flower categories from seen categories given textual descriptions of the unseen ones.