They really need to bump up to 48 KHz so all the music doesn't sound like it's being played over a telephone. A factor of two in the cost shouldn't be prohibitive. So much of the audio generation stuff I've seen has fatal flaws like this baked into the dataset and/or training process that ensure the output can't sound good even in theory, it's kinda frustrating.
It's also frustrating that nobody AFAIK has trained a music model on an actually large dataset. We're training large language models on a significant fraction of the text of the whole internet! Where are the audio models trained on a significant fraction of all recorded music? This one was trained (in part) on a dataset of 280k hours of music, if I read the paper correctly. I don't think that comes anywhere close to being a significant fraction of all recorded music.