Abstract
GenéLive! is a deep generative model designed to automate the creation of "charts" for rhythm action games. Developed by researchers at KLab Inc. and Kyushu University, the model incorporates musical structures through a "beat guide" and a multi-scale convolutional architecture. Successfully deployed in the production of Love Live! School Idol Festival All Stars, GenéLive! has halved the human labor required for chart generation while maintaining high artistic standards.
Introduction
In rhythm games, "charts" are the sequences of notes that players interact with. Creating these charts is traditionally a labor-intensive process, requiring artists to manually align notes with a song's rhythm and melody. This manual workflow often takes up to 40 hours per song, creating a significant bottleneck in game production. Previous automated solutions often failed to maintain long-term musical structure or produce high-quality low-difficulty charts.
Methodology
GenéLive! improves upon standard CNN-BiLSTM architectures by introducing two key components:
- The Beat Guide: A trinary input array derived from song metadata (BPM and time signature) that indicates bar starts and regular beats. This provides the model with a structural "hint" to align notes correctly.
- Multi-Scale Conv-Stack: To address the different time scales of music, the model uses four parallel convolutional stacks. This allows the system to simultaneously process short-term percussive events and long-term melodic repetitions.
The model also employs "Beat Masking" during training—randomly dropping beats from the guide—to ensure it learns to interpret the audio signal rather than relying solely on metadata.
Results and Business Impact
The model was tested on private datasets from Love Live! and Utano Princesama, as well as public benchmarks. GenéLive! achieved state-of-the-art performance, notably outperforming previous models like Dance Dance Convolution (DDC) in "Beginner" difficulty modes ($F_1$-score of 0.866 vs 0.783).
In a production environment at KLab Inc., the model reduced the time required to generate a first-draft chart from 40 hours to 20 hours. Qualitative evaluations showed that professional artists accepted nearly all generated notes in sample sequences, requiring only minimal adjustments to meet final quality standards.
Conclusion
GenéLive! demonstrates the effectiveness of incorporating explicit musical priors into deep learning models for creative content generation. By reducing the manual workload of chart creation by half, it provides a scalable solution for the rhythm game industry while preserving the artistic quality expected by players.
Summary updated