There is a rising interest in animating realistic virtual agents for multiple purposes in diferent domains. Such a task requires systems capable of generating complex mental states on par with human emotional complexity. Considering the high representational capacity of Generative Adversarial Networks (GANs), it is only natural to consider them in such applications. In this work, we propose a conditional GAN model for generating sequences of facial expressions of categorical complex emotions. Trained on a scarce and highly imbalanced dataset, the proposed model is able to generate realistic variable-length sequences in a single inference step. These expressions of emotional states, of which there are 24 in total, follow the Facial Actions Coding System (FACS) formatting. In the absence of meaningful objective evaluation methods, we propose a deeplearning-based metric to assess the realism of generated Action Unit (AU) sequences: the Action Unit Fréchet Inception Distance (AUFID). Objective and subjective results validate the realism of our generated samples.