This is my ever-expanding experimentation into generative tools for music visualisation. The end result will be an end-to-end system for live visualisation of a DJ set, but the intermediate steps seem to continually breed their own cans of worms. So far I've handled most of the audio feature engineering, and I'm now playing with mechanisms for using these audio features to control GAN outputs.
- Extract controller features (from screen recording of Rekordbox) at each time stamp
- Extract tracklist
- For each track in the tracklist:
- extract audio features
- use extracted audio features to feed StyleGAN to generate frame-by-frame visualisation (this will constitute the default representation of a track, & will be later modified in the mixing phase)
- For each time step in the mix, use controller features + default track visualisations of all tracks currently playing to create 'mixed' visualisation