Emotion-Based Style Transfer On Visual Art Using Gram Matrices

Despite the importance of emotions in the creation and enjoyment of art, the focus of most autonomous art-generating models has not been on eliciting emotions. This project presents a process to transfer the style of an emotion, specifically sadness, excitement, or disgust, onto a seed image. Emotion-labeled art was processed through the VGG-16 network. Resulting layer activations were utilized to calculate average Gram matrices for each emotion, followed by a style loss function for each style transfer. The generated artworks had clear distinctions across emotions, demonstrating the existence of unique patterns that an artificial intelligence can learn to create emotion-evoking art.