UVR Guide

Assuming you have UVR already installed, go to settings, download center, click the key, enter 02aeb35c203ed0a9 as the code, and click confirm.

You will now have access to VIP models for MDX-Net. Some of these sound significantly better than others.

General Settings

Enable Help Hints in the settings. When hovering over settings in the GUI, you'll see helpful hints explaining the setting.

To reduce the time it takes to generate outputs (on my system, it went from 1 minute to less than 20 seconds on average), select GPU Conversion on the main screen.

If you don't care about the vocal stem, select Instrumental Only. Alternatively, you can also select Vocals Only

Output should always be FLAC or WAV. You can always rencode the audio to MP3 later. It's not worth the quality loss selecting MP3.

Good Music

UVR generally works best on songs with less proccessed vocals. For example, UVR works very well on Ruthless by Tristam, but may perform badly on Unfold by Porter Robinson, where the main vocals are pitched up and distorted.

When creating stems, the more prominent and present the instrument is, the better the end result will be. For example, wait for me by jives will produce a good result when separating the "Drums" stem.

Good Models

Models that contain the word 'karaoke' will not remove background vocals from the instrumental. This is helpful for songs like The City by Madeon, where "I got tomorrow to-" is considered a background vocal, and you may not wish to remove it.

MDX-Net is generally the best model, but by default does not create stems.

MDX-Net

You can reduce artifacting across all MDX models by visiting settings -> choose option -> advanced MDX-Net options -> denoise output. Note that this will make the output generation step take significantly longer.


UVR-MDX-NET Inst HQ 1 - The default model. Sounds fine, works well enough.

VIP Main_471 - A slightly better HQ 1. May take longer to process, but has less artifacting in general.

Main - Slower than HQ 1. May produce better results

Inst Main - Faster than HQ 1, but slower than main. If you're looking for speed, this is best.

Reverb HQ by FoxJoy - Removes all reverb. Makes music sound really, really weird, but may produce cleaner vocals when paired with another model.

Karaoke - Best if you don't wish to include background vocals in your final product.

Demucs

V4 models are reccommended, and produce the best results. V3 kinda sucks. These are the best models if you're looking to separate stems. It also seems to produce very good instrumental tracks with little artifacting. In my testing, the vocals are still sometimes barely audible. Ensamble with MDX-Net for best vocal/instrumental results.


htdemucs_ft - Finely tuned model. Best for quality, but takes slightly longer than other models.

htdemucs - Default model. Fast and works well for stems.

Ensemble mode

This mode allows you to combine models. I recommend only combining models in the same family to reduce artifacting, but feel free to explore!

My personal settings for ensemble mode are as follows:

Max Spec/Max Spec - This setting depends on your music. Having a max spec vocal stem, while sounding cleaner, may produce unwanted artifacts. The default Max Spec/Min Spec may be best.

Selected Models:

  • MDX-NET Inst HQ 1
  • Main 438
  • Inst_full 292
  • htdemucus_ft - May leave some background vocals

Generally the more models you select, the better the final audio will sound, but this isn't always the case, and you may instead get unwanted artifacts.

Edit Report
Pub: 14 May 2023 16:40 UTC
Edit: 14 May 2023 16:48 UTC
Views: 558