Neural Latent Traversal with Semantic Constraints

Majid Abdolshah, Hung Le, Thommen Karimpanal George, Vuong Le, Sunil Gupta, Santu Rana, Svetha Venkatesh

Abstract: Whilst Generative Adversarial Networks (GANs) generate visually appealing high resolution images, the latent representations (or codes) of these models do not allow controllable changes on the semantic attributes of the generated images. Recent approaches proposed to learn linear models to relate the latent codes with the attributes to enable adjustment of the attributes. However, as the latent spaces of GANs are learnt in an unsupervised manner and are semantically entangled, the linear models are not always effective. In this study, we learn multi-stage neural transformations of latent spaces of pre-trained GANs that enable more accurate modeling of the relation between the latent codes and the semantic attributes. To ensure identity preservation of images, we propose a sparsity constraint on the latent space transformations that is guided by the mutual information between the latent and the semantic space. We demonstrate our method on two face datasets (FFHQ and CelebA-HQ) and show that it outperforms current state-of-the-art baselines based on FID score and other numerical metrics.