[SUGGESTION] More demucs (models), more fun !

[MarcoRavich]

I've recently tested @CarlGao4's Demucs GUI that allows to inference - even on CPU (but without OpenVINO afaik) many interesting separation models/versions.

Models details from @adefossez's Demucs-fork-git:

• htdemucs: first version of Hybrid Transformer Demucs. Trained on MusDB + 800 songs. Default model.
• htdemucs_ft: fine-tuned version of htdemucs, separation will take 4 times more time but might be a bit better. Same training set as htdemucs.
• htdemucs_6s: 6 sources version of htdemucs, with piano and guitar being added as sources. Note that the piano source is not working great at the moment.
• hdemucs_mmi: Hybrid Demucs v3, retrained on MusDB + 800 songs.
• mdx: trained only on MusDB HQ, winning model on track A at the [MDX][mdx] challenge.
• mdx_extra: trained with extra training data (including MusDB test set), ranked 2nd on the track B of the [MDX][mdx] challenge.
• mdx_q, mdx_extra_q: quantized version of the previous models. Smaller download and storage but quality can be slightly worse.
• SIG: where SIG is a single model from the model zoo.

I know that the upcoming version will allow their use, anyway I believe that some conversions is needed to.

[RyanMetcalfeInt8]

Thanks for the share -- next version will include htdemucs (same one we've always had), htdemucs_ft, and htdemucs_6s (mainly for completeness ... as all sources point out, it's not very good).

It also includes mdx23c variant of drum separation.

I haven't yet tried htdemucs_mmi, and other variants of mdx* -- how do they stack up against htdemucs?

[MarcoRavich]

I haven't yet tried htdemucs_mmi, and other variants of mdx* -- how do they stack up against htdemucs?

Well, there are detailed per-model Quality leaderboards @ MVSEP:
https://mvsep.com/en/quality_checker

(note: it would be really helpful to have metrics - example: https://mvsep.com/quality_checker/entry/8196 - inside the model info tab to understand which to use)

Anyway it would be cool to have all models - or a simple method to use publicly released weights - from @ZFTurbo's Music Source Separation Universal Training Code repo, IMHO.