First Describe, Then Depict: Generating Covers for Music and Books via Extracting Keywords: This paper presents two methods to generate high resolution uncopyrighted book covers or music album covers.

Abstract

In this paper, we consider the two algorithms of generating artwork covers based on texts or audio file features. The resulting image is combined from existing images labelled with keywords after applying filter-based image harmonization. To achieve realistic composition, we train GAN to predict an appropriate filter or apply emotion-based Neural Style Transfer. The quality of generated book covers and music album covers was evaluated by assessors. According to their assessment, the suggested algorithms appeared to produce a better result compared to the existing solutions. The suggested methods also achieve printing quality and require less time for computations, moreover, generated images can be used without copyright infringement.