We introduce a novel global-to-local(G2L) approach for 3D model generation through building an adversarial network(GAN) assiciated with additional local discriminators to construct the overall structure of the shape while yielding more plausible part geometries.