No idea how it technically works, but from listening to a few samples it sounds like they've setup an environment with different instruments and generate notes to play those instruments.
The environment in all of the samples I've listened to so far sound exactly the same, as if they all share the same mixing configuration and mastering.
OpenAI's jukebox for example (check youtube for samples) is actually generating waveforms and sounds completely different.
The image equivalent to this is that I feel soundraw.io has a database of transparent stock photo images that it uses to compose a collage image that adheres to some strict rules depending on the genre.
Whereas OpenAI Jukebox is more like Dalle2 and Stable Diffusion.
