mvsep is the best imo ) just built the app with Nativefier )

 

UVR5 is the answer, and details have been already said in this thread before
also don't hesitate to experiment with models that split track into stems, not just vocals+instrumental, you can stitch them back easily, but results may be better

 

If you feel like paying/using credits for anything over 16bit wav output, it's possible. Using Nativefier has no impact on the output. For free? No 24bit wav output, which running UVR5 locally does. MVsep is way faster for me, but the 16bit wav output vs UVR5 24bit wav output makes the slower separation worthwhile. UVR5 allows you to set values for segment/chunk size, overlap, number of shifts, etc. This makes things a lot slower for me, but I just do it on a different Mac and don't really care how long it runs.

 

Your client has shit taste, just saying.

 

This happens because of the way this music is generated. Since you're compiling a bunch of code into the task of performing a human feeling, the best it can do is to copy and paste real performances of real musicians in a unnatural way to mimic those feelings. The results are strange instrument portamento, legatos and quantizations. Guitars that blend into polyphonic synths. All sort of weird stuff. When you split, all that flawed things start getting more attention, so it's not that RX can't split it, but the amount of crap under the hood without all the masking is tremendous.

 

thanks everyone for all the suggestions. I completed the job with MVSEP (like a week ago I think) but the result is still bad IMO, but bearable somehow acceptable and I got paid for it. Thanks. I think I need to spend more time to learn about everything that has to do with stem separation and shitty AI Music and how it works.

I never expect anyone ever in this whole world to have better taste than me. So, it's just another day at the office and I got no problem with it.

 

INSTRUMENTAL:
https://voca.ro/1nRJfPZtrdpj

PROCESSED v1:
https://voca.ro/17kkP07jvifi

PROCESSED v2:
https://voca.ro/1cqRRm8p7a3S

Sadly the original AI song was generated on a older suno model (maybe 3.5 or 4) where there is more "shimmer" to the sound at the time so its hard to tame that. But the stems might help you hone it in.


STEM GROUPS 4 MIXING:
https://pixeldrain.com/u/1CE19K6V

another option is to run the instrumental back through suno and have v5.5 generate a better quality version.

 

Last edited: May 11, 2026

both are garbage, use scnet or rofo

im happy that devs of real programs arent using open models - the more gatekept good things are the better

 

can you run through those algos so we can compare?

 

Anything not open source just gets cracked anyway. If expensive commercial programs aren't cracked, you'll just be one of the people not using them.

 

If you have a capable GPU and time to tweak, UVR is your best bet.

Otherwise, go with MVSEP. I wouldn't bother with RX Rebalance much less LALAL.AI. It's honestly kid's toys compared to the more robust alternatives.