Emotion-conditional Image Generation Reflecting Semantic Alignment with Text-to-Image Models

Emotion plays a crucial role in human interactions. In the field of emotion analysis with machine learning, Emotion-conditional Image Generation (EIG) aims to create images that reflect specific emotions by leveraging image generation models. This paper introduces a novel framework that can align semantic information for specific emotions of the generated images. By incorporating object categories that evoke specific emotion into prompts, the proposed method can generate images that represent a variety of emotion-related objects. Our framework is evaluated by emotion classification of visual emotion datasets, demonstrating the improvements in both object diversity and emotion reflection.