Turning one photo into a K-pop style dance clip - notes from building KDA Firestarter Dance AI

Hi all! I’ve been working on a small project, KDA Firestarter Dance AI, that takes a single portrait photo and generates a short dance video in the style of KDA’s “Firestarter” choreography.

A few things I learned along the way:

  • Pose-driven image-to-video works far better when the source photo is front-facing with the full upper body visible; cropped heads drift a lot.
  • Short templates (8-12s) keep identity much more stable than longer clips.
  • Fast arm movements are where most artifacts show up, so we slowed some template segments slightly before generation.

Would love feedback from anyone working on motion transfer or character animation - especially on keeping faces consistent across frames.