3 comments

[ 3.1 ms ] story [ 13.5 ms ] thread
It's obviously not at the quality level of human producers, but I found it surprisingly listenable. The setup is basically just streaming MusicGen songs to AzuraCast with the Multi-Band Diffusion decoder.

This was buried a few days ago with no discussion so I'm posting it again. Lmk if that's discouraged.

Not bad! It's about as coherent as background music needs to be, at least in the typical Lo-Fi beats context.

I don't have access to Sonauto. I wish I could ask them what the compute requirements are for this. How much energy does it take to produce 1 hour?

Thanks! It's actually all generated by a single RTX 3090 (so constant ~250 watts). It runs at about 0.8X realtime. The Sonauto web app uses serverless A10Gs, though.