Abstract
A rhythm action game is a music-based video game in which the player ischallenged to issue commands at the right timings during a music session. Thetimings are rendered in the chart, which consists of visual symbols, callednotes, flying through the screen. KLab Inc., a Japan-based video gamedeveloper, has operated rhythm action games including a title for the "LoveLive!" franchise, which became a hit across Asia and beyond. Before this work,the company generated the charts manually, which resulted in a costly businessoperation. This paper presents how KLab applied a deep generative model forsynthesizing charts, and shows how it has improved the chart productionprocess, reducing the business cost by half. Existing generative modelsgenerated poor quality charts for easier difficulty modes. We report how weovercame this challenge through a multi-scaling model dedicated to rhythmactions, by considering beats among other things. Our model, named Gen\'eLive!,is evaluated using production datasets at KLab as well as open datasets.