Model guide · htdemucs, htdemucs_ft and htdemucs_6s

Demucs Models Compared: htdemucs vs htdemucs_ft vs htdemucs_6s

Demucs v4 ships three pretrained Hybrid Transformer models, and Demucs.net runs all of them on cloud GPUs. Compare their stems, quality, speed and cost.

Try It Free Hear htdemucs

Model

StandardFast, right for most songsBest qualityCleanest stems6 stemsAdds guitar and piano

Stems

4 stemsVocals, drums, bass, other2 stemsVocals and instrumental

Output

MP3320 kbpsWAV16-bitStudioFLAC24-bitStudio

I own this audio or have permission to process it. Licensing

15 free credits every month · No card needed · Uploads auto-deleted

The Three Demucs Models at a Glance

htdemucs, htdemucs_ft and htdemucs_6s are all Demucs v4 Hybrid Transformers trained on MUSDB18-HQ plus 800 songs. Time and credits are for a 4-minute song here.

`htdemucs`

Standard on this site

Default

Vocals
Drums
Bass
Other

Or vocals and instrumental

SDR: 8.80 dB
Time: 20.6 s
Credits: 4

2 stems: 13.1 s, 3 credits

Pick it for: Most songs, karaoke, first listens

`htdemucs_ft`

Best quality on this site

Vocals
Drums
Bass
Other

Or two; four fine-tuned models inside

SDR: 9.00 dB
Time: 32.0 s
Credits: 7

2 stems: 5 credits

Pick it for: Acapellas and bass lines you will publish

`htdemucs_6s`

6 stems on this site

Vocals
Drums
Bass
Other
Guitar
Piano

Also splits four or two

SDR: None
Time: 25.1 s
Credits: 6

No SDR published

Pick it for: Guitar parts; piano with care

SDR: signal-to-distortion ratio on the MUSDB HQ test set, from Table 1 ofHybrid Transformers for Music Source Separation(Rouard, Massa and Défossez, 2022) and theDemucs README. Time: GPU time on an Nvidia A100, September 2026, 320 kbps MP3 output.

Hear htdemucs, the Default Model, on a Real Song

A Creative Commons song split into four stems with htdemucs on this site. Solo each stem before you pay extra for htdemucs_ft or htdemucs_6s.

0:000:00

Vocals
Drums
Bass
Other

Track credits

htdemucs: The Default Demucs Model

htdemucs is the first Hybrid Transformer Demucs model and the default of the demucs command, trained on MUSDB plus 800 songs.

It splits a mix into vocals, drums, bass and other, or into vocals and instrumental, which the project's documentation calls karaoke mode.

In the paper's Table 1 it scores 8.80 dB, ahead of Hybrid Demucs v3 on the same data at 8.34 dB and Spleeter at 5.91 dB.

Pick it when: you want stems fast, for a karaoke track or a practice mix, or you are not sure which model suits the song. Run it first.

Outputs: Vocals
Drums
Bass
Other
SDR by stem (dB): Drums10.05
Bass9.78
Vocals8.93
Other6.42
Time, 4-minute song: 20.6 s, or 13.1 s for 2 stems
Credits: 4 for 4 stems, 3 for 2

Hybrid Transformer Demucs: a waveform branch and a spectrogram branch joined by a Transformer, splitting a mix into four stems — The "ht" is Hybrid Transformer: one branch reads the waveform, one the spectrogram, and a Transformer links the two.

htdemucs_ft: Fine-Tuned for the Best Quality

htdemucs_ft is htdemucs fine-tuned once per source: four copies of the model, each trained further on drums, bass, other or vocals.

All four run on every song, which is why the README warns it "will take 4 times more time but might be a bit better".

That 4x is model time. End to end here it took 1.56 times as long, because MP3 encoding costs the same for both models.

Pick it when: the vocal or the bass is the stem you will publish, or htdemucs left bleed you can hear. For a party karaoke track, skip it.

Outputs: Vocals
Drums
Bass
Other
SDR change after fine-tuning (dB): Bass+0.61
Vocals+0.27
Drums+0.03
Other−0.10
Time, 4-minute song: 32.0 s, 1.56 times as long
Credits: 7 for 4 stems, 5 for 2

htdemucs_6s: Demucs 6 Stems, With Guitar and Piano

htdemucs_6s adds guitar and piano to vocals, drums, bass and other, the only Demucs model with those two stems. It runs the 6 stems option here.

Demucs 6 stems means htdemucs_6s; ht_demucs6s is the same model spelt differently.

The README says plainly that "the piano source is not working great at the moment", and no SDR is published for it.

Pick it when: you need the guitar on its own, to learn a part or sample a riff. Check the piano stem before you rely on it.

Our test: the rock song in the demo, which has no piano

Guitar

−5 dB peak, solid

Piano

−45 dB peak, in effect silent

Outputs: Vocals
Drums
Bass
Other
Guitar
Piano
SDR: None published
Time, 4-minute song: 25.1 s, 1.22 times htdemucs
Credits: 6 for 6 stems

htdemucs vs htdemucs_ft: Is Best Quality Worth the Credits?

Same architecture, training data and outputs. The trade is 0.2 dB overall, mostly in vocals and bass, for 1.56 times the time and 1.6 times the credits.

SDR by stem, MUSDB HQ (dB)htdemucshtdemucs_ft

Bass+0.61: 9.78
10.39
Vocals+0.27: 8.93
9.20
Drums+0.03: 10.05
10.08
Other−0.10: 6.42
6.32
Overall+0.20: 8.80
9.00

The cost, for a 4-minute song

	htdemucs	htdemucs_ft
Time, 4 stems	20.6 s	32.0 s
Credits, 4 stems	4	7
Credits, 2 stems	3	5

Our rule: run htdemucs first. If the vocal stem carries traces of the drums or the bass wobbles, rerun the song on htdemucs_ft. Most songs never need the second pass.

Which Is the Best Demucs Model for Your Job?

A karaoke track or backing track

htdemucs, 2 stems

3 credits and about 13 seconds. Our karaoke maker opens in that mode.

An acapella for a remix

htdemucs_ft, 2 stems

Its largest gains are on vocals and bass.

Drums, bass and vocals as stems

htdemucs, 4 stems

Or htdemucs_ft if the bass stem will be published. The stem splitter opens with four stems selected.

A guitar or piano stem

htdemucs_6s

The only model with those sources. Listen to the piano stem before you trust it.

Learning a part by ear

htdemucshtdemucs_6s

htdemucs_6s for a guitar line. htdemucs_ft adds nothing you will hear at practice volume.

Not sure

htdemucs

The default for a reason, and the cheapest. Move up only when you hear why.

Other Demucs Models You May Come Across

All score lower than the Hybrid Transformers on the same benchmark, 7.7 dB for Hybrid Demucs v3 in the README's table, so Demucs.net runs only htdemucs, htdemucs_ft and htdemucs_6s.

hdemucs_mmi: Hybrid Demucs v3, retrained on MUSDB plus 800 songs
mdxmdx_extra: The v3 models from the 2021 Music Demixing Challenge
mdx_qmdx_extra_q: Quantised versions of those two

Demucs Models FAQ

Quick answers on htdemucs, htdemucs_ft and htdemucs_6s: their names and what they need to run. Anything else? Get in touch .

What does the "ht" in htdemucs stand for?

Hybrid Transformer. Demucs v4 replaced the innermost layers of Hybrid Demucs (v3) with a cross-domain Transformer encoder that reads the waveform and the spectrogram at once; with 800 extra training songs, the paper puts it 0.45 dB of SDR ahead of Hybrid Demucs.

Is htdemucs_ft really 4 times slower than htdemucs?

In model time, roughly, because four fine-tuned models run instead of one. End to end here htdemucs_ft is 1.56 times slower, 32.0 against 20.6 seconds for a 4-minute song, since encoding the MP3 stems takes the same time either way.

What is the difference between htdemucs_6s and ht_demucs6s?

None; they are two spellings of the same six-stem model. Its name in the Demucs repository is htdemucs_6s, and it is the model behind the 6 stems option here.

Which Demucs model is best for vocals?

htdemucs_ft: 9.20 dB on vocals in the paper’s Table 1 against 8.93 dB for htdemucs. For karaoke, where the instrumental matters more than the vocal, the standard model is usually enough; htdemucs_6s is not the tool for this, its extra sources are guitar and piano.

Does htdemucs_6s separate piano well?

Not reliably. The authors say the piano source "is not working great", and in our own test on a song without piano the piano stem came back empty while the guitar stem was solid. Use htdemucs_6s for guitar, audition the piano stem before you rely on it, and use htdemucs or htdemucs_ft when you need neither.

Can I use all three Demucs models on a free account?

Yes. Every signed-in account can pick htdemucs, htdemucs_ft or htdemucs_6s; they differ only in credits per minute of audio, at 1, 1.6 and 1.3 times the standard rate.

How much GPU memory do the Demucs models need?

Running them yourself, the README asks for at least 3 GB of GPU memory, about 7 GB with default settings; on a CPU expect roughly 1.5 times the length of the track. Here htdemucs, htdemucs_ft and htdemucs_6s run on cloud GPUs, so none of that applies.

Is Demucs still maintained?

The original repository says it is not maintained anymore, since its author left Meta; a fork carries important bug fixes. The pretrained models are unaffected: htdemucs, htdemucs_ft and htdemucs_6s are finished weights, and they are what this site runs.

Try All Three Demucs Models Free

Upload a song, pick htdemucs, htdemucs_ft or htdemucs_6s, and hear the difference yourself. 15 free credits every month, no card needed.

Split a Song Free