Dance to the music

This article reveals the planning and strategy behind one of HavefunwithAIch's YouTube releases—from thumbnails to audience targeting.

We used prompts during image generation to specify dancing in sync with the music, and in Vidu or Ray2, we also specified the music itself to see how well the AI could understand it.

First of all, most keywords related to music and dance are quite strong and do elicit a reaction. However, the results tend to be difficult to convey clearly to viewers.

To prioritize clarity and ease of understanding, we structured the process as follows:

1. Use Recraft to generate an outfit that clearly matches the specific dance style.
2. Use Recraft to generate a pose that symbolizes that particular dance.
3. Since video generation prioritizes movement based on the initial pose, the resulting motion tends to resemble the intended dance.

Additionally, since AI’s understanding of video is still underdeveloped, it often learns from iconic scenes from famous films or videos (not to replicate them exactly, but in a way that resembles them). The task, then, becomes finding those references and triggering them through keywords.

If there is no training data available, or if the AI has learned from incorrect references, the results will inevitably be broken or inconsistent.

To put it simply: in the case of samurai action, the AI tends to prioritize inaccurate movements learned from anime, so no matter how hard you try, it will only return incorrect motions.

When it comes to music and dance, however, the AI seems to have reached a reasonably acceptable level of understanding.

Not just a YouTube channel — a channel where tools are tested and compared with extensive knowledge and practice, pushing the limits of AI.