Unified Color Harmony Model

Abstract

Color harmony is an aesthetic sensation evoked by the balanced and coherent arrangement of the colors of visual elements. While traditional methods define harmonious subspaces from geometric relationships or numerical formulas, we employ a data-driven approach to create a unified model for evaluating and generating color combinations of arbitrary sizes. By treating color sequences as linguistic sentences, we construct a color combinations generator using SeqGAN, a generative model capable of learning discrete data through reinforcement learning. The resulting model produces color combinations as much preferred as those by the best models of each size and excels at penalizing color combinations from random sampling. The distribution of the generated colors has more diverse hues than the input data, in contrast to the NLP-based model that predominantly predicts achromatic colors due to exposure bias. The flexible structure of our model allows for simple extension to additional conditions such as group preference or emotional keywords.