Feature matching loss in RAVE aligns intermediate discriminator activations between real and generated audio

In addition to the adversarial loss, RAVE computes a feature matching loss: the distance between the discriminator’s intermediate feature maps for real and generated audio. This loss stabilizes GAN training and encourages the generator to produce audio whose deep spectral/temporal statistics match the real data, not just its surface waveform. In the default loss weights, feature matching is weighted at 20 versus 1 for the adversarial term, making it the dominant term in phase 2.

Examples

The default loss weights include 'adversarial': 1. and 'feature_matching' : 20, — feature matching dominates the phase-2 objective.

Assessment

Why does RAVE weight feature matching 20× higher than the adversarial loss? What would happen to training stability if feature matching were removed entirely?