There isn’t a lot of content of a commercial nature in my day-to-day life, but there are some special examples slightly related to some content within the business. During the development of my original music rhythm game, Domain Echoing, I encountered a tricky design challenge. Level design for rhythm games takes a long time and there are usually hundreds of levels in a rhythm game. Most commercial rhythm games recruit dozens of members to complete the level of production. However, the game I developed did not have such a large team. So, how to improve production efficiency while maintaining the quality of the game’s levels became a major challenge. At the same time, the degree of freedom of the gameplay in my game is much higher than in other rhythm games, so the traditional manual creation of level editing is extremely time-consuming. I wanted to use technology to improve efficiency.
To solve this problem, earlier this year I decided to train a generative model. The model consists of two sub-modules: one categorizes the types of music clips and generates rhythm scores; the other generates the level (a JSON file) for specific Note combinations based on these scores generated and optimizes transitions between passages. Since I have played many rhythm games, I am familiar with many of their levels. However, due to copyright reasons, I could not use the levels of rhythm games for training my model. So I am considering using community levels. Communities of rhythm games usually have a lot of self-made levels, and their levels are freely available. After recognizing this, I collected nearly 100 homemade levels from OSU!, extracted their rhythms, and put them into model training.
The training of the first module was successful, and the model is now able to recognize a track’s rhythm score applicable to rhythmic music games quite accurately. The second module is relatively more complex and is still under development. However, I believe I will be able to complete it several months later.
Personally, it takes me about 8 hours to create a level, whereas with the generative model, it only takes a few minutes. Then, it’s just a matter of modifying and optimizing it manually, and it’s ready to go. The improvement was significant. In addition to that, if I can complete this project, I will be able to keep my rhythm game running in a sustainable way.
Finally, it made me realize that while most AI-generated content is not accepted by the general public (e.g. AI-generated images and music). But certain areas of generated content have been able to be widely used. This provides a new starting point for the future of generative AI. For many projects a more sustainable way of creating content has also been explored.