NAPS: Non-adversarial polynomial synthesis

Grigorios G. Chrysos, Yannis Panagakis

Abstract: Highlights • We address image generation task by training a decoder-only generator without auxiliary modules. • We design a polynomial generator using tensorial factors to capture high-order correlations of the input variables. • Two variants, which are implemented and compared, demonstrate the way to design new polynomial architectures. • Our experiments in both digit and face generation demonstrate the efficacy of the proposed polynomial generator. • Our generator can synthesize images with significantly less parameters than the corresponding baseline. Abstract Generative Adversarial Nets (GANs) are currently the dominant model for high fidelity image synthesis. GANs suffer from two major drawbacks: complicated dynamics and the requirement for an auxiliary network for training (discriminator). However, if we train a decoder-only network we circumvent both drawbacks. To achieve that, the decoder should capture high-order correlations that exist between the variables. We demonstrate this is possible by designing a high-order polynomial generator using tensorial factors. We implement two variants of the model, which we call NAPS. We experiment with both MNIST and CelebA and showcase that our model captures the data distribution and synthesizes new images with significantly less parameters than the corresponding baseline.