Lyric-Based Image Generation for Individual Songs with Text2Image Model
Shoichi Sasaki, Hiroko Nishida, Ken-ichi Sawai, Taketoshi Ushiama · 2024
This paper proposes a method for generating images to represent impressions of a song automatically. Thumbnail images associated with songs in music distribution services contribute greatly to the selection of songs because they allow users to visually grasp the characteristics of songs in a short period of time. On the other hand, when the same image is assigned to all of the songs in the same album, they may not adequately represent the contents of each song. In order to solve this problem, we propose a method for automatically generating images that represent the impression of a song using song lyrics information and the Text2Image model. Specifically, by analyzing the meaning of the lyrics, prompts suitable for image generation are created, and these prompts are input to the Text2Image model to obtain images. We then evaluate the validity of the generated images through user studies.