The RAVE VST loads a .ts model in any DAW as an audio effect that re-timbres incoming audio

The RAVE VST (available as beta for Windows/Mac/Linux from the IRCAM Forum) loads a streaming-exported RAVE .ts model file. In the DAW, it acts as an audio effect: audio on the track is encoded into latent variables and decoded through the neural model in realtime. Because RAVE is highly nonlinear and sensitive to input level, input gain/compression is a key control — over-hot input produces harsh distortion. The VST exposes editable latent dimensions and a dry/wet mix, and can also run the prior method when a prior was exported with the model.

Examples

In Ableton: insert the RAVE VST on a guitar track, load a streaming .ts model, then adjust input gain until the output sounds natural before automating a latent dimension for timbral modulation.

Assessment

A producer inserts the RAVE VST on a synth bass and gets harsh distortion. What is the most likely cause and what knob should they adjust first?