Manipulating individual latent dimensions of a RAVE model morphs continuously between audio-effect and synthesizer behavior

Between the encode and decode stages, each latent dimension can be intercepted and modified before it reaches the decoder. This spans a full spectrum: when all latents come from the audio encoder the model is an audio effect (timbre transformation); when all latents are set by the user it is a synthesizer; mixing the two gives a hybrid. The tutorial’s example takes the first four latent dimensions from a live audio encoder (audio-reactive), sets dimensions 5 and 7 from sliders (direct control), and drives dimensions 6 and 8 with parametrized LFOs (automation). Latent dimensions are sampled from an isotropic normal distribution during training, so most meaningful variation lies between -3 and 3 (model-dependent). RAVE sorts dimensions by their impact on the output sound, so the low dimensions stay reactive to input while higher ones remain controllable.

Examples

Encode audio → 8 latents. Replace latent 5 with a 0-1 slider so it morphs one dimension; replace latent 6 with an LFO for automatic timbral evolution; leave latents 1-4 from the encoder for audio-reactivity.

Assessment

In the hybrid patch, how do latent dimensions 1-4 differ functionally from 5-8? What value range covers most meaningful variation in a RAVE latent space, and why?