50 models sorted by SDR
(from the public ones - so only those available to download and offline use)
(last updated 07.10.2024)
These are basically the top single models for now
(conventionally after these, additional vocal residues kick in, especially if not a vocal model)
Based on Multisong dataset evaluation on MVSEP chart with similar or the same parameters and inference if applicable.
Kim’s Mel Roformer
model_bs_roformer_ep_317_sdr_12.9755
model_bs_roformer_ep_368_sdr_12.9628 (viperx/UVR beta)
BS-Roformer_LargeV1 (unwa’s ft)
Unwa's Mel-Roformer Beta 3 (although for this day, SDR wasn’t tested with the same parameters vs above, so it’s based on assumption that unwa used the same parameters in synth dataset measurement)
Unwa's Mel-Roformer Beta 4
Unwa's Mel-Roformer Beta 5e
0) InstVoc MDX23C HQ (fullband a.k.a. 1648, 8K FFT)
0b) InstVoc MDX23C HQ 2 (fullband)
1) voc_ft
1b) UVR-MDX-NET HQ_4 (inst)
2) MDX23C_D1581 (a.k.a. narrowband)
3) Kim Vocal 2
4) Kim Vocal 1
5) UVR-MDX-NET_Main_427 (voc)
6) UVR-MDX-NET_Main_406 (voc)
7) UVR-MDX-NET HQ_5 (inst)
7) UVR-MDX-NET HQ_3 (inst)
8) UVR-MDX-NET_Main_438 (voc)
9) UVR-MDX-NET_Main_390 (voc)
10) Kim inst (a.k.a. other)
11) UVR-MDX-NET_Main_340 (voc)
12) Inst 3 (a.k.a. 464)
13) UVR-MDX-NET HQ_2 (inst)
(for vocal models, here start those with more vocal residues in instrumentals - can be still handy for specific songs)
+4 pos.
9) Inst Main (496)
10) Inst 2
11) UVR-MDX-NET HQ1
12) UVR-MDX-NET HQ 337 >382>338 epoch
13) Inst 1
14) HQ 386>403>292 epoch
15) UVR-MDX-NET2>NET3>NET1>9482 (NET3 a.k.a. 9.7)
16) htdemucs_ft (4 stem) (S 10/O 0.95)
17) hdemucs_mmi (4 stem)
18) htdemucs_6s (6 stem)
19) UVR-MDX-NET_Inst_82_beta
20) Demucs3 Model B (4 stem)
21) UVR-MDX-NET_Inst_187_beta
(dango.ai, Audioshake, Bandlab not evaluated)
Somewhere here, trash begins (excluding GSEP)
22) Moises.ai (probably before transition to newer Roformer models)
23) DeMIX Pro 4.1.0
24) Myxt (AudioShake 128kbps)
25) UVR-MDX-NET_Inst_90_beta
26) RipX DeepRemix 6.0.3
27) kuielab_b (4 stem) (MDX Model B from 2021 MDX Challenge)
28) kuielab_a (4 stem)
29) LALAL.AI
30) GSEP (6 stem) (although it sometimes gives much better results than its SDR)
VR arch
31) 7_HP2-UVR (a.k.a. HP2-MAIN-MSB2-3BAND-3090_arch-500m)
32) 3_HP-Vocal-UVR
33) 2_HP-UVR (HP-4BAND-V2_arch-124m)
34) 9_HP2-UVR (HP2-4BAND-3090_4band_arch-500m_1)
35) 1_HP-UVR (HP_4BAND_3090_arch-124m)
36) 8_HP2-UVR (HP2-4BAND-3090_4band_arch-500m_2)
37) 14_SP-UVR-4B-44100-2 (4 band beta 2)
38) 4_HP-Vocal-UVR
39) 13_SP-UVR-4B-44100-1 (4 band beta 1)
39) 15_SP-UVR-MID-44100-1
40) 16_SP-UVR-MID-44100-2
41) 14_HP-Vocal-UVR
42) VR | MGM_LOWEND_A_v4
43) 12_SP-UVR-3B-44100
44) Demucs 2 (4 stem)
(6 other old VR models proceeds)
50) Spleeter 4 stems
51) Spleeter 2 stems
52) GSEP after mixdown from 4 stems separation
Only instrumental models listed (outdated)
(4 stem and MDX23C models lies in all categories):
Tier 1
MDX-Net models (trained by UVR team)
0) MDX23C HQ 1648 fullband
1) MDX23C HQ 2 fullband
1b) UVR-MDX-NET HQ_4 (inst)
2) MDX23C_D1581 narrowband
7) HQ3
10) Kim inst (other)
12) Inst 3
13) HQ2
Tier 2
+4 pos.
9) Inst Main (496)
10) Inst 2
11) HQ1
12) HQ 337 >382>338 epoch
13) Inst 1
14) HQ 386>403>292 epoch
Demucs 4
16) htdemucs_ft (S 10/O 0.95)
17) hdemucs_mmi
18) htdemucs_6s
20) Demucs 3 Model B (mdx_extra)
Tier 3
(somewhere between place 9-20 might be dango.ai, Audioshake, later maybe Bandlab)
22) Moises.ai
23) DeMIX Pro 4.1.0
24) Myxt (AudioShake 128kbps)
26) RipX DeepRemix 6.0.3
27) MDX-Net Model B from 2021 MDX Challenge (kuielab_b)
28) kuielab_a
29) LALAL.AI
30) GSEP (although it sometimes gives much better results than its SDR)
Tier 4
VR arch
31) 7_HP2-UVR (a.k.a. HP2-MAIN-MSB2-3BAND-3090_arch-500m)
33) 2_HP-UVR (HP-4BAND-V2_arch-124m)
34) 9_HP2-UVR (HP2-4BAND-3090_4band_arch-500m_1)
35) 1_HP-UVR (HP_4BAND_3090_arch-124m)
36) 8_HP2-UVR (HP2-4BAND-3090_4band_arch-500m_2)
Tier 5
37) 14_SP-UVR-4B-44100-2 (4 band beta 2)
38) 13_SP-UVR-4B-44100-1 (4 band beta 1)
Tier 6
39) 15_SP-UVR-MID-44100-1
40) 16_SP-UVR-MID-44100-2
42) VR | MGM_LOWEND_A_v4
43) 12_SP-UVR-3B-44100
44) Demucs 2
(6 other old VR models proceeds)
Tier 7
50) Spleeter 4 stems
51) Spleeter 2 stems
52) GSEP after mixdown from 4 stems separation
Differences by SDR divided for vocals and instrumentals are important to divide I think only in ensembles. In all other cases, if SDR is bigger for instrumentals in some model, it will be bigger for vocals vs the same model. At least only for ensembles there were so little differences that we had two top ensembles for both vocals and instrumentals.
__________________________________