So there are models for Pose estimation like [PoseNet] or Qualcomm's [MediaPipe] that give you a graph of the body position. This metadata can be converted to a

So there are models for Pose estimation like [PoseNet] or Qualcomm's [MediaPipe] that give you a graph of the body position. This metadata can be converted to a string representation and be fed into a generative model. If I record myself in different exercises, extract the poses, and label them with the appropriate steps, I might be able to perform few-shot inference or RAG over a set of labeled positions, such as the model can infer new body positions from exercise steps. If that'd possible, then it is fairly easy:tm: to generate illustrations (e.g. like stickman figures. See [Kawamoto] for an example) Of course, none of this is necessary if I can record (or let users record) and upload videos of each exercise, having enough of them to not feel they are repetitive. [PoseNet](https://github.com/tensorflow/tfjs-models/tree/master/posenet) [MediaPipe](https://huggingface.co/qualcomm/MediaPipe-Pose-Estimation) [Kawamoto](https://github.com/kenkawakenkenke/stickfigure-recorder)