NP-DRAW: A Non-Parametric Structured Latent Variable Model for Image Generation

In this paper, we present a non-parametric structured latent variable model\nfor image generation, called NP-DRAW, which sequentially draws on a latent\ncanvas in a part-by-part fashion and then decodes the image from the canvas.\nOur key contributions are as follows. 1) We propose a non-parametric prior\ndistribution over the appearance of image parts so that the latent variable\n``what-to-draw'' per step becomes a categorical random variable. This improves\nthe expressiveness and greatly eases the learning compared to Gaussians used in\nthe literature. 2) We model the sequential dependency structure of parts via a\nTransformer, which is more powerful and easier to train compared to RNNs used\nin the literature. 3) We propose an effective heuristic parsing algorithm to\npre-train the prior. Experiments on MNIST, Omniglot, CIFAR-10, and CelebA show\nthat our method significantly outperforms previous structured image models like\nDRAW and AIR and is competitive to other generic generative models. Moreover,\nwe show that our model's inherent compositionality and interpretability bring\nsignificant benefits in the low-data learning regime and latent space editing.\nCode is available at https://github.com/ZENGXH/NPDRAW.\n