Latent Image Animator: Learning to Animate Images via Latent Space Navigation

Due to the remarkable progress of deep generative models, animating images become increasingly efficient, whereas associated results have become realistic. Current animation-approaches commonly exploit structure extracted from driving videos. Such structure representation is in transferring motion from driving videos to still images., such approaches fail in case the source image and driving video large appearance variation. Moreover, the extraction of structure requires additional modules that endow the animation-model with complexity. Deviating from such models, we here introduce the Latent Animator (LIA), a self-supervised autoencoder that evades need for representation. LIA is streamlined to animate images by linear in the latent space. Specifically, motion in generated video is by linear displacement of codes in the latent space. Towards this, learn a set of orthogonal motion directions simultaneously, and use their combination, in order to represent any displacement in the latent space. quantitative and qualitative analysis suggests that our model and significantly outperforms state-of-art methods on VoxCeleb, and TED-talk datasets w.r.t. generated quality.