Machine Learning Model to Evaluate the Appropriateness of Layout for Automatic Generation of Graphic Design Works

Generative networks (GANs) are commonly used for image generation to generate images with noise as input to the trained generator. However, it is difficult to generate appropriate graphic designs with GANs when the user specifies the material images and text to be used as input, as in the case of graphic design. In this study, we focus on the graphic design layout and propose a method for automatically generating graphic designs using adversarial GANs. We trained the generator and discriminator using GANs by converting the training data of graphic designs into images that represent their visual importance. The trained discriminators were then used to evaluate the automatically generated layouts using specified materials. The results of experiments with subjective evaluations using subjects indicates that the proposed method was effective.