If there is a paired dataset, I guess this could be "easy". The stylegan "input" is essentially used to control parameters at various stages within the network, so you could adjust them one at a time, or on varying schedules or something, to get the sort of gradual effect.
I know that randomly instantiated neural networks can produce some pretty trippy image transformations as well, so maybe there is a way to bootstrap without paired data.
Lastly, I doubt this is on the right track but it would be cool if you could produce appropriately styled images with a "compression" approach. ie, trying to fit the audio information into some small visually meaningful latent space, and then using that to generate images.
edit: ok just watched the first example, back to square 1 for me. Its literally pulling stuff from paintings.