Automatic Layout Generation Based on GAN

Layout generation is to construct a proper combination by setting the location, size and other attributes of each layout element. Creative and visually attractive layouts are generally designed by excellent professional graphic designers. Yet, in the Internet era, it is challenging to keep up with the rapid growing requirement because the whole design process is heavily dependent on manual work and inefficient, and communication costs are expensive between Party A and designers. This paper mainly solves the problem of automatic layout generation in various graphic designs, including documents and images. Understanding the relationships between these graphic elements is necessary when designing a new layout or extending an existing layout. To achieve this, we propose a neural network model based on adversarial generative network and self-attention mechanism. It trains on real data samples, learns the distribution of samples, and is finally used for layout generation tasks. Based on LayoutGAN + +, we investigate the layout generation issue in graphic design, create a network structure that reflects its fundamentals, and theoretically analyze the model's generator, discriminator, and auxiliary decoder. Then, three data sets (Magazine, RICO, PubLayNet ) were tested on the model, and on the basis of this, potential model optimization directions were explored. Experimental results demonstrate that our model can generate meaningful layouts in various layout generation application scenarios, such as mobile applications, documents, and magazine layout generation tasks