Sketches are abstract representations of visual perception and visuospatial construction. In this work, we proposed a new framework, GAN-CNMP, that incorporates a novel adversarial loss on CNMP to increase sketch smoothness and consistency. Through the experiments, we show that our model can be trained with few unlabeled samples, can construct distributions automatically in the latent space, and produces better results than the base model in terms of shape consistency and smoothness.