These tools are awesome, I have iZotope RX which also does a version of this, but when I'm in the music flow grabbing something off of youtube vs firing up another piece of software can be more conducive creatively for me.

Yeah makes sense! I’m sure we will start to see this integrated into the next generations of a lot of music making software with it being open source. FWIW Demucs sounds a lot better than the RX separation (which I think uses Spleeter, could be wrong) to my ears - it doesn’t have any of that mushy “low bitrate” kind of sound, it’s much crisper

I wonder how hard it would be to write a Max4Live device that wraps this? That would be awesome, have the original track in my project for reference, and have a Max4Live device generate isolated stems of the different parts.

I thought about the exact same thing when I first saw Spleeter and it turned out someone had already done it: https://github.com/diracdeltas/spleeter4max

Looks like the meat of this is using a JS script to invoke Spleeter using execSync so it might be quite simple to adapt to Demucs, maybe I’ll give it a go!

Songs are also composed of each instrument on a single audio track and they are sometimes distributed

Oh yeah of course, but in my experience it’s quite rare for these multi-tracks to end up online. Demucs can in many cases extract an acappella which would be good enough to make a remix with (of course using ML trickery, so it’s not going to be the actual real acappella and might have some leakage from other sounds in the track etc.)