Ensembles

(for vocals)

 

Note: Max FFT and Spec can be used interchangeably

Ensemble instructions

MVSep-exclusive (paid):

- #1 - BS becruily 124 fullness + PolarFormer fullness lv 1 (Avg)

 

- #2 - BS MVSep 124 fullness lv 1 + PolarFormer fullness lv 1 (Avg)

--->The first one should be better (dca100fb8)

 

Open/free ensembles:
It’s still better to use Deux on its own, but at the cost of more bleeding, you can get more BVs.
No particular order anymore.

- BS anvuew Mag v2 + SW (Max Spec) - bleedy with instruments bleed and very noisy, turns out too much imo (at least for batch separation)

- BS Mag + SW (Max Spec)

- BS anvuew ft1 + SW (Max Spec) - noisy at silent passages

- Mel deux + BS SW (Max Spec)

- BS HyperACE v2 Voc + BS SW (Max Spec)
- Mel voc fv6 + BS SW (Max Spec)

- Mel vocfv7beta1 + BS SW (Max Spec)
- (dca100fb8)

 

Note. E.g Mag v2 and voc fv6 ensembled with the SW can be also used on nextgen.uvronline.app.

“Both free users and paid users can use it.” - squiddagod (more)

- MVSep presets editor now can be used to create a node for the ensemble of your choice (“If 2+ algorithm nodes, it is premium only”), avg used by default. Some presets are already shared by community

- SESA Colab

- MVSepless Colab

 

- medicdoesstuff’s ensemble

(for extracting FULLNESS acapellas)

 

Steps:

1. Process the song using SCNet very high fullness (MVSEP) + BS PolarFormer (124 bands) separately, then download their respective vocal tracks.

2. Do a Max-FFT ensemble of the two audio files together using UVR or this Colab notebook.

3. Process the ensemble vocal track via UVR Online using the BS-RoFormer Mag v2 (anvuew) model.

4. Denoise that track the using Mel-RoFormer De-noise (aufr33)bmodel with "average" setting.

5. Reprocess that denoised track again with the 'UVR De-noise' in "minimum" setting, that's all.

 

“From what I've tested so far, these leads to acapellas that sound almost official (some tracks even somehow get their headphone bleed isolated along with the vocals) and much less muddier than the standard "only process thru becruily deux" way.

(also if available, do an inversion before all of these since it obviously will sound better)” - medicdoesstuff’

 

 

Older ensembles:

- Mel vocfv7beta3 + BS SW (Max Spec)

 

- Mel deux + BS 2025.07 (Max FFT/Spec) (former “best Ensemble for vocals”) (-||-)

 

- MVSep Polarformer, Deux and MVSep 2025.07 (Max FFT) (“The result is great.

Has all the vocals from Polarformer and 2025.07 but the fullness of Deux, along with a bit of instrumental bleed. But the result is better than any of the separate models”) (Kashi/omaroka)

 

- deux + big beta 7 (Max FFT) (“another good ensemble”) (neoculture)

 

- BS Roformer HyperACE Voc v2 + 2025.07 (Max FFT) (formerly best) (dca)

 

- BS Revive 3e + BS 2025.07 (Max FFT) (former “best vocal ensemble”) (-||-)

 

- Mel Becruily Vocal + MVSEP’s BS 2025.07 (Max FFT) (-||-) (-||-)

 

- unwa’s bigbeta5 + becruily vocal (Max FFT) (midol)

 

- voc_gaboxFv2 + becruily vocal (heauxdontlast/gilliaan)

 

- big beta 7 + bs_roformer_mag (Max FFT) (good for RVC) (neoculture)

- Unwa “Big Beta 4 + Big Beta 5e (Average Spec) (“really good to reduce the noise while keeping the fullness”) (gilliaan)

 

- unwa beta6 + voc_fv4 (good for anime and RVC)

 

- unwa beta6x + voc_fv4 (“some songs I can use big beta 6x, and it's enough, others I need to ensemble it with voc_fv4”) (Rainboom Dash)

 

- unwa beta6x + voc_fv6 (“would make a good ensemble, but the amount of noise is horrific

and I heard that phase swapper would fix it”)

 

- unwa beta 6x + beta5e (avg/avg) (good for RVC) (gilliaan)

 

- BSRoformer-Viperx1297, BSRoformer-LargeV1 by Unwa, unwa_ft2_bleedless, mel_band_roformer_vocals_becruily, Gabox voc_fv4 - Average/Average Spec (good for cleaning inverts) (AG89)

 

- Models ensembled (inst, voc) available for premium users on mvsep.com

(SDR 10.44-11.93 and “High Vocal Fullness” variants)