Recent deep generative models for static graphs have focused on areas such as molecular design. However, many real- world problems involve temporal graphs whose topology and attribute values evolve dynamically over time. Examples include theoretical models such as activity-driven networks and link-node memories, as well as important applications such as protein folding, human mobility networks, and social network growth. % This work proposes the “Temporal Graph Generative Adversarial Network” (TG-GAN) model for continuous-time temporally-bounded graph generation, which captures the deep generative process of temporal graphs through compositions of time-budgeted temporal walks, which are themselves composed of truncated temporal walks. Specifically, a novel temporal graph generator is proposed that jointly models truncated edge sequences, time budgets, and node attributes, incorporating novel activation functions that enforce temporal validity constraints under a recurrent architecture. In addition, a new temporal graph discriminator is proposed that combines time and node encoding operations over a recurrent architecture to distinguish generated sequences from real ones sampled by a newly-developed truncated temporal walk sampler. Extensive experiments on both synthetic and real-world datasets confirm that TG-GAN significantly outperforms five bench- marking methods in terms of efficiency and effectiveness.