dancenet

Why we included this project

DanceNet packs three generative ideas into a small Keras codebase that is easy to read end to end. A variational autoencoder squeezes dance frames into a compact latent space, the LSTM models how those latent states evolve over time, and a mixture density layer predicts a distribution over possible next poses instead of committing to a single guess, which is what gives the generated motion variety. The whole pipeline runs from a notebook: download the trained weights, open dancegen.ipynb, and generate video, or retrain from scratch on your own image sequences using separate scripts for encoding frames and reconstructing motion. Researchers building pose or motion generators can borrow the VAE-plus-MDN pattern as a template, and learners get a concrete, visual demonstration of how recurrent generative models behave. It is a study reference rather than a production service, so treat it as something to learn from before reaching for a larger framework.