MVSEP models from UVR5 GUI explained

 

- Ensemble option - further developed custom code of the original MDX23 (not available in UVR) - custom tech, consisting of various models from UVR and in-house, non-public models unavailable in UVR

 

- Demucs 4 ft - (settings might be shifts 1 and overlap 0.75 as he tested once) - same as in Colabs or UVR

 

MDX B (not sure about whether min, avg, max is set) - the option has MDX arch models:

- Newest MDX models added - Kimberley - Kim inst (ft other), Kim Vocal 1 & 2, HQ_2, 3

- 8.62 2022.01.01 - is NET 1 (9.7) with Demucs 2, this one has a new name now. It had slightly bigger SDR in the same multisong dataset as the newer model below - discrepancy vs UVR5 SDR results might be on the server side (e.g. different chunks), so it might be still the same. The dates can only relate to date of adding the model to the site and nothing more, not sure here, but it might be it - NET 1 is older model than below indeed. Looks like that the model is used with Demucs 2 enabled (at least he said it was configured like this at some point)

- 8.51 2022.07.25 - might be vocal 423 a.k.a main, not sure if with Demucs 2 (judging how instrumental from inversion in 423 looked like - cannot be any inst model yet, since they were released in the end of 2022 - epoch 418 in September to be precise) - it was tested in multisong dataset on page 2 as MDX B (UVR 2022.07.250 - the date is the same as before, so nothing new here), can't say now if Demucs 2 is used here. In times of 9.7/NET 1 it was decreasing SDR a bit on I don't know which dataset, but instrumentals usually sounded kinda richer with this enabled. Now it's better to use other models to ensemble.

The change in MDX-B models scheme was probably to unify SDR metrics to multisong dataset.

- Demucs 3 Model B - mdx_extra (and rather not mdx_extra_q as ZFTurbo said it's "original" and used mdx_extra name on the channel referring to this model; in most cases the one below should be better)

 

- Ultimate Vocal Remover HQ - the option has VR architecture models

Window size used - 512

Here are all VR arch model names

- UVRv5 Demucs - rather the same names

- MVSEp models - unavailable in UVR5

- MDX B Karaoke - Possibly MDX-UVR Karokee or MDX-UVR Karokee 2 (karokee_4band_v2_sn irc), maybe the latter

 

The rest below on the MVSEP’s list is outdated and not really recommended to use anymore

 

Issues using MVSEP

 

- NaN error during upload is usually cause by unstable internet connection, and it usually happens on mobile connections when you already upload more than one file elsewhere.

If you have NaN error, just retry uploading your file.

 

- Rarely it can happen after upload that error about not uploaded file occurs - you need to upload your file again.

 

- If you finished separation and click back, model list can disappear till you won’t click on other algorithm and pick yours again. But if you click separate instead, it will process with the first model which was previously on the list (at least if it was also your previous choice).

 

- Slow download issues. Separation was complete, and I was listening to the preview when playback on the preview page simply stopped, and couldn't be started. Main page didn't load (other site worked).

Also, I couldn't download anything. It showed 0b/s during attempt of downloading.

Two solutions:

- close all MVSEP tabs completely and reopen

- Connect to VPN, preview some track, but after a short time, the same can happen and nothing is playing or buffering. Then fire up Free Download Manager, and simply copy the download link there, and it will start downloading. Later, the browser can also start downloading something you clicked a moment a go. Crazy.

 

—

 

Comparing to MDX with 14.7kHz cutoff, depending on a track, VR models only (not MDX/Demucs) might leave or cut more instruments or leave more constant vocal residues, but in general VR is trained model at 20kHz with possible mirroring covering 20-22kHz, generally less aggressive vocal removing (with exceptions) but most importantly, comparing to MDX, VR tends to leave some specific noise in a form of leftover artifacts of vocal bleeding, but from the other hand MDX, especially models with cutoff, can be more muddy and recall original track mastering less.

______