I’m looking into improving VAEs for diffusion models and it’s surprising how few diverge from the original (no pun intended) of the KLD loss. This paper arxiv.org/pdf/2309.13160 by Mariano Rivera seems much more promising by calculating the mean and variance per batch instead of per sample

Why isn’t the constant Q transform more used when it comes to audio ML? Its properties seem much more advantageous vs the Mel spectrogram when it comes to music or audio generation in general where there’s both low frequency and high frequency signals