Separation & stem splitGPU workload

Mega 53 Stems

Q: Who is Mega 53 Stems for?

Separation & stem split Exports detected FLAC stems from MVSep Mega 53 Stems for detailed instrument exploration.

Q: What tasks can Mega 53 Stems help with?

Separation & stem split; Exports detected FLAC stems from MVSep Mega 53 Stems for detailed instrument exploration.; Exploring and identifying more instruments than four-stem or six-stem separation exposes; Full-band songs or dense mixes for experimental extraction of detected instrument FLAC stems; Inspect whether a mix contains specific instruments such as saxophone, brass, mandolin, synth, or percussion; Export many candidate FLAC stems for offline editing, sampling, or arrangement analysis

Q: Which model and keyword entities are connected with Mega 53 Stems?

MVSep Mega 53 Stems (BS-RoFormer), MVSep Mega 53 Stems, MVSep Mega 53 page, HF open weights, ZFTurbo release v1.0.21, ZFTurbo MSST repository, BS-RoFormer, accordion

Exploring and identifying more instruments than four-stem or six-stem separation exposes

Exports detected FLAC stems from MVSep Mega 53 Stems for detailed instrument exploration.

10 credits/useAudio file inputMax 50 MBEstimated 600-2400 seconds

Input

Full-band songs or dense mixes for experimental extraction of detected instrument FLAC stems

Audio formats

mp3wavflacoggm4aaac

Output

Exports detected FLAC stems from MVSep Mega 53 Stems for detailed instrument exploration.

Use case

Exploring and identifying more instruments than four-stem or six-stem separation exposes

Models, parameters, and sources

MVSep Mega 53 Stems is an experimental open-weight BS-RoFormer model for extracting many instrument stems. It is locally deployable from the public checkpoint, but the upstream release warns that dedicated instrument models can be cleaner and that at least 16 GB VRAM is recommended.

MVSep Mega 53 Stems (BS-RoFormer)MVSep Mega 53 StemsMVSep Mega 53 pageHF open weightsZFTurbo release v1.0.21ZFTurbo MSST repositoryBS-RoFormeraccordionacoustic-guitarback-vocalbanjobass

MVSep Mega 53 Stems

Open-weight BS-RoFormer checkpoint for MVSep Mega 53 Stems. It targets many instrument classes and filters detected outputs rather than replacing the six-stem quality default.

MVSep Mega 53 page HF open weights ZFTurbo release v1.0.21 ZFTurbo MSST repository BS-RoFormer

Targets: 53 instrument classes including vocals, drums, bass, guitars, piano, brass, woodwinds, strings, percussion, synth, and many specialist instruments.
Runtime: local MSST inference with model_type bs_roformer, the official Mega 53 YAML config, PCM_24 FLAC output, and dual-GPU segment routing when two CUDA devices are visible.
Quality: useful for discovery and more detailed instrument extraction; upstream notes that dedicated single-instrument or smaller stem models can be cleaner for a specific target.
Output handling: detected FLAC stems from the 53-stem instrument set, ready for offline editing, sampling, or arrangement analysis.

Benchmark chart: Official MVSep Mega 53 per-stem scores

Official MVSep Mega 53 per-stem scores

MVSep Mega 53 page

These are the official per-stem SDR values published on the MVSep Mega 53 page. They are displayed as scores, not as a product ranking.

Evaluation context	Model	New model SDR	Detailed metrics
Accordion	MVSep Mega 53 StemsTelkNet	6.6498	Open-source SDR6.2494Delta SDR0.4004
Acoustic guitar	MVSep Mega 53 StemsTelkNet	5.0797	Open-source SDR5.0024Delta SDR0.0773
Back vocal	MVSep Mega 53 StemsTelkNet	6.5262	Open-source SDR6.4179Delta SDR0.1083
Banjo	MVSep Mega 53 StemsTelkNet	3.7532	Open-source SDR3.1593Delta SDR0.5939
Bass	MVSep Mega 53 StemsTelkNet	11.2886	Open-source SDR11.1680Delta SDR0.1206
Bassoon	MVSep Mega 53 StemsTelkNet	5.1663	Open-source SDR4.6595Delta SDR0.5068
Bells	MVSep Mega 53 StemsTelkNet	4.8040	Open-source SDR1.1190Delta SDR3.6850
Bowed strings	MVSep Mega 53 StemsTelkNet	12.4486	Open-source SDR12.4486Delta SDR0.0000
Brass	MVSep Mega 53 StemsTelkNet	6.8487	Open-source SDR6.7042Delta SDR0.1445
Cello	MVSep Mega 53 StemsTelkNet	5.2257	Open-source SDR5.0364Delta SDR0.1893
Clarinet	MVSep Mega 53 StemsTelkNet	5.8690	Open-source SDR5.0505Delta SDR0.8185
Congas	MVSep Mega 53 StemsTelkNet	9.5946	Open-source SDR9.1747Delta SDR0.4199
Digital piano	MVSep Mega 53 StemsTelkNet	9.0179	Open-source SDR7.9634Delta SDR1.0545
Dobro	MVSep Mega 53 StemsTelkNet	8.2290	Open-source SDR7.6562Delta SDR0.5728
Double bass	MVSep Mega 53 StemsTelkNet	15.6032	Open-source SDR14.1731Delta SDR1.4301
Drums	MVSep Mega 53 StemsTelkNet	11.4520	Open-source SDR9.2502Delta SDR2.2018
Electric guitar	MVSep Mega 53 StemsTelkNet	8.1856	Open-source SDR8.1543Delta SDR0.0313
Flute	MVSep Mega 53 StemsTelkNet	6.8557	Open-source SDR6.2134Delta SDR0.6423
French horn	MVSep Mega 53 StemsTelkNet	5.5136	Open-source SDR5.2635Delta SDR0.2501
Glockenspiel	MVSep Mega 53 StemsTelkNet	8.1523	Open-source SDR3.6621Delta SDR4.4902
Guitar	MVSep Mega 53 StemsTelkNet	2.6164	Open-source SDR2.5661Delta SDR0.0503
Harmonica	MVSep Mega 53 StemsTelkNet	11.8575	Open-source SDR10.9265Delta SDR0.9310
Harp	MVSep Mega 53 StemsTelkNet	7.6523	Open-source SDR6.3767Delta SDR1.2756
Harpsichord	MVSep Mega 53 StemsTelkNet	1.9524	Open-source SDR1.6090Delta SDR0.3434
Hi-hat	MVSep Mega 53 StemsTelkNet	3.2904	Open-source SDR2.4681Delta SDR0.8223
Keys	MVSep Mega 53 StemsTelkNet	9.3106	Open-source SDR9.3032Delta SDR0.0074
Kick	MVSep Mega 53 StemsTelkNet	11.5485	Open-source SDR11.5485Delta SDR0.0000
Lead vocal	MVSep Mega 53 StemsTelkNet	5.4663	Open-source SDR5.4663Delta SDR0.0000
Mandolin	MVSep Mega 53 StemsTelkNet	4.7735	Open-source SDR4.4256Delta SDR0.3479
Marimba	MVSep Mega 53 StemsTelkNet	4.8830	Open-source SDR4.4821Delta SDR0.4009
Oboe	MVSep Mega 53 StemsTelkNet	4.5555	Open-source SDR3.3616Delta SDR1.1939
Organ	MVSep Mega 53 StemsTelkNet	10.8244	Open-source SDR10.3684Delta SDR0.4560
Percussion	MVSep Mega 53 StemsTelkNet	2.8897	Open-source SDR2.5008Delta SDR0.3889
Piano	MVSep Mega 53 StemsTelkNet	6.8080	Open-source SDR6.7787Delta SDR0.0293
Saxophone	MVSep Mega 53 StemsTelkNet	9.4985	Open-source SDR8.8875Delta SDR0.6110
Sitar	MVSep Mega 53 StemsTelkNet	5.0434	Open-source SDR4.6529Delta SDR0.3905
Snare	MVSep Mega 53 StemsTelkNet	6.7784	Open-source SDR6.1338Delta SDR0.6446
Strings	MVSep Mega 53 StemsTelkNet	8.8151	Open-source SDR8.8151Delta SDR0.0000
Synth	MVSep Mega 53 StemsTelkNet	2.0539	Open-source SDR2.0539Delta SDR0.0000
Tambourine	MVSep Mega 53 StemsTelkNet	3.5544	Open-source SDR3.0589Delta SDR0.4955
Timpani	MVSep Mega 53 StemsTelkNet	4.9779	Open-source SDR4.6423Delta SDR0.3356
Toms	MVSep Mega 53 StemsTelkNet	-1.0708	Open-source SDR-2.0607Delta SDR0.9899
Triangle	MVSep Mega 53 StemsTelkNet	6.0197	Open-source SDR5.9274Delta SDR0.0923
Trombone	MVSep Mega 53 StemsTelkNet	3.0751	Open-source SDR2.6949Delta SDR0.3802
Trumpet	MVSep Mega 53 StemsTelkNet	5.8668	Open-source SDR4.9658Delta SDR0.9010
Tuba	MVSep Mega 53 StemsTelkNet	7.5229	Open-source SDR7.1957Delta SDR0.3272
Ukulele	MVSep Mega 53 StemsTelkNet	6.9721	Open-source SDR6.7869Delta SDR0.1852
Viola	MVSep Mega 53 StemsTelkNet	1.8581	Open-source SDR1.8581Delta SDR0.0000
Violin	MVSep Mega 53 StemsTelkNet	3.3285	Open-source SDR3.1018Delta SDR0.2267
Vocal	MVSep Mega 53 StemsTelkNet	11.6590	Open-source SDR11.6590Delta SDR0.0000
Wind	MVSep Mega 53 StemsTelkNet	8.6632	Open-source SDR8.6317Delta SDR0.0315
Wind chimes	MVSep Mega 53 StemsTelkNet	6.7529	Open-source SDR4.6440Delta SDR2.1089
Woodwind	MVSep Mega 53 StemsTelkNet	3.3224	Open-source SDR3.3123Delta SDR0.0101

Paper and benchmark notes

Music Source Separation with Band-Split RoPE Transformer

SDX23 music separation track - 2023

BS-RoFormer uses band-split spectrogram features plus RoPE Transformers; the SDX23 system ranked first in the music source separation track and a smaller version reported 9.80 dB average SDR on MUSDB18HQ.

arXiv:2309.02612

Official / repositories

MVSep Mega 53 page HF open weights ZFTurbo release v1.0.21 ZFTurbo MSST repository

Papers / technical notes

BS-RoFormer

Parameter guide

Model: MVSep Mega 53 Stems, open-weight checkpoint mvsep_mega_model_bs_roformer_53_stems_v1.ckpt with the official YAML config.
Outputs: detected non-empty FLAC stems from the 53-stem instrument set for detailed offline editing, sampling, or arrangement analysis.
Quality note: this is an advanced exploration tool. For overall separation quality and predictable stems, use six-stem separation.
Runtime note: the release recommends at least 16 GB VRAM. Dual-GPU segment routing can accelerate long audio, but each GPU still needs enough VRAM to load the model.

Use cases

Inspect whether a mix contains specific instruments such as saxophone, brass, mandolin, synth, or percussion
Export many candidate FLAC stems for offline editing, sampling, or arrangement analysis
Use as an experimental discovery tool when six fixed stems are not detailed enough

Workflow

1Upload a song and keep the fixed official Mega 53 settings
2The worker runs the open-weight BS-RoFormer checkpoint with dual-GPU segment routing when two CUDA cards are visible
3Review detected stems from the task output list before offline editing

Pre-flight checklist

For best overall separation quality, use six-stem separation first
Mega 53 is experimental and each individual stem may be weaker than a dedicated specialist model
The model is memory intensive; each participating GPU still needs enough VRAM to load the checkpoint

Separation & stem splitGPU workload

Mega 53 Stems

Exploring and identifying more instruments than four-stem or six-stem separation exposes

Exports detected FLAC stems from MVSep Mega 53 Stems for detailed instrument exploration.

10 credits/useAudio file inputMax 50 MBEstimated 600-2400 seconds

Input

Full-band songs or dense mixes for experimental extraction of detected instrument FLAC stems

Audio formats

mp3wavflacoggm4aaac

Output

Exports detected FLAC stems from MVSep Mega 53 Stems for detailed instrument exploration.

Use case

Exploring and identifying more instruments than four-stem or six-stem separation exposes

Models, parameters, and sources

MVSep Mega 53 Stems (BS-RoFormer)MVSep Mega 53 StemsMVSep Mega 53 pageHF open weightsZFTurbo release v1.0.21ZFTurbo MSST repositoryBS-RoFormeraccordionacoustic-guitarback-vocalbanjobass

MVSep Mega 53 Stems

Open-weight BS-RoFormer checkpoint for MVSep Mega 53 Stems. It targets many instrument classes and filters detected outputs rather than replacing the six-stem quality default.

MVSep Mega 53 page HF open weights ZFTurbo release v1.0.21 ZFTurbo MSST repository BS-RoFormer

Targets: 53 instrument classes including vocals, drums, bass, guitars, piano, brass, woodwinds, strings, percussion, synth, and many specialist instruments.
Runtime: local MSST inference with model_type bs_roformer, the official Mega 53 YAML config, PCM_24 FLAC output, and dual-GPU segment routing when two CUDA devices are visible.
Quality: useful for discovery and more detailed instrument extraction; upstream notes that dedicated single-instrument or smaller stem models can be cleaner for a specific target.
Output handling: detected FLAC stems from the 53-stem instrument set, ready for offline editing, sampling, or arrangement analysis.

Benchmark chart: Official MVSep Mega 53 per-stem scores

Official MVSep Mega 53 per-stem scores

MVSep Mega 53 page

These are the official per-stem SDR values published on the MVSep Mega 53 page. They are displayed as scores, not as a product ranking.

Evaluation context	Model	New model SDR	Detailed metrics
Accordion	MVSep Mega 53 StemsTelkNet	6.6498	Open-source SDR6.2494Delta SDR0.4004
Acoustic guitar	MVSep Mega 53 StemsTelkNet	5.0797	Open-source SDR5.0024Delta SDR0.0773
Back vocal	MVSep Mega 53 StemsTelkNet	6.5262	Open-source SDR6.4179Delta SDR0.1083
Banjo	MVSep Mega 53 StemsTelkNet	3.7532	Open-source SDR3.1593Delta SDR0.5939
Bass	MVSep Mega 53 StemsTelkNet	11.2886	Open-source SDR11.1680Delta SDR0.1206
Bassoon	MVSep Mega 53 StemsTelkNet	5.1663	Open-source SDR4.6595Delta SDR0.5068
Bells	MVSep Mega 53 StemsTelkNet	4.8040	Open-source SDR1.1190Delta SDR3.6850
Bowed strings	MVSep Mega 53 StemsTelkNet	12.4486	Open-source SDR12.4486Delta SDR0.0000
Brass	MVSep Mega 53 StemsTelkNet	6.8487	Open-source SDR6.7042Delta SDR0.1445
Cello	MVSep Mega 53 StemsTelkNet	5.2257	Open-source SDR5.0364Delta SDR0.1893
Clarinet	MVSep Mega 53 StemsTelkNet	5.8690	Open-source SDR5.0505Delta SDR0.8185
Congas	MVSep Mega 53 StemsTelkNet	9.5946	Open-source SDR9.1747Delta SDR0.4199
Digital piano	MVSep Mega 53 StemsTelkNet	9.0179	Open-source SDR7.9634Delta SDR1.0545
Dobro	MVSep Mega 53 StemsTelkNet	8.2290	Open-source SDR7.6562Delta SDR0.5728
Double bass	MVSep Mega 53 StemsTelkNet	15.6032	Open-source SDR14.1731Delta SDR1.4301
Drums	MVSep Mega 53 StemsTelkNet	11.4520	Open-source SDR9.2502Delta SDR2.2018
Electric guitar	MVSep Mega 53 StemsTelkNet	8.1856	Open-source SDR8.1543Delta SDR0.0313
Flute	MVSep Mega 53 StemsTelkNet	6.8557	Open-source SDR6.2134Delta SDR0.6423
French horn	MVSep Mega 53 StemsTelkNet	5.5136	Open-source SDR5.2635Delta SDR0.2501
Glockenspiel	MVSep Mega 53 StemsTelkNet	8.1523	Open-source SDR3.6621Delta SDR4.4902
Guitar	MVSep Mega 53 StemsTelkNet	2.6164	Open-source SDR2.5661Delta SDR0.0503
Harmonica	MVSep Mega 53 StemsTelkNet	11.8575	Open-source SDR10.9265Delta SDR0.9310
Harp	MVSep Mega 53 StemsTelkNet	7.6523	Open-source SDR6.3767Delta SDR1.2756
Harpsichord	MVSep Mega 53 StemsTelkNet	1.9524	Open-source SDR1.6090Delta SDR0.3434
Hi-hat	MVSep Mega 53 StemsTelkNet	3.2904	Open-source SDR2.4681Delta SDR0.8223
Keys	MVSep Mega 53 StemsTelkNet	9.3106	Open-source SDR9.3032Delta SDR0.0074
Kick	MVSep Mega 53 StemsTelkNet	11.5485	Open-source SDR11.5485Delta SDR0.0000
Lead vocal	MVSep Mega 53 StemsTelkNet	5.4663	Open-source SDR5.4663Delta SDR0.0000
Mandolin	MVSep Mega 53 StemsTelkNet	4.7735	Open-source SDR4.4256Delta SDR0.3479
Marimba	MVSep Mega 53 StemsTelkNet	4.8830	Open-source SDR4.4821Delta SDR0.4009
Oboe	MVSep Mega 53 StemsTelkNet	4.5555	Open-source SDR3.3616Delta SDR1.1939
Organ	MVSep Mega 53 StemsTelkNet	10.8244	Open-source SDR10.3684Delta SDR0.4560
Percussion	MVSep Mega 53 StemsTelkNet	2.8897	Open-source SDR2.5008Delta SDR0.3889
Piano	MVSep Mega 53 StemsTelkNet	6.8080	Open-source SDR6.7787Delta SDR0.0293
Saxophone	MVSep Mega 53 StemsTelkNet	9.4985	Open-source SDR8.8875Delta SDR0.6110
Sitar	MVSep Mega 53 StemsTelkNet	5.0434	Open-source SDR4.6529Delta SDR0.3905
Snare	MVSep Mega 53 StemsTelkNet	6.7784	Open-source SDR6.1338Delta SDR0.6446
Strings	MVSep Mega 53 StemsTelkNet	8.8151	Open-source SDR8.8151Delta SDR0.0000
Synth	MVSep Mega 53 StemsTelkNet	2.0539	Open-source SDR2.0539Delta SDR0.0000
Tambourine	MVSep Mega 53 StemsTelkNet	3.5544	Open-source SDR3.0589Delta SDR0.4955
Timpani	MVSep Mega 53 StemsTelkNet	4.9779	Open-source SDR4.6423Delta SDR0.3356
Toms	MVSep Mega 53 StemsTelkNet	-1.0708	Open-source SDR-2.0607Delta SDR0.9899
Triangle	MVSep Mega 53 StemsTelkNet	6.0197	Open-source SDR5.9274Delta SDR0.0923
Trombone	MVSep Mega 53 StemsTelkNet	3.0751	Open-source SDR2.6949Delta SDR0.3802
Trumpet	MVSep Mega 53 StemsTelkNet	5.8668	Open-source SDR4.9658Delta SDR0.9010
Tuba	MVSep Mega 53 StemsTelkNet	7.5229	Open-source SDR7.1957Delta SDR0.3272
Ukulele	MVSep Mega 53 StemsTelkNet	6.9721	Open-source SDR6.7869Delta SDR0.1852
Viola	MVSep Mega 53 StemsTelkNet	1.8581	Open-source SDR1.8581Delta SDR0.0000
Violin	MVSep Mega 53 StemsTelkNet	3.3285	Open-source SDR3.1018Delta SDR0.2267
Vocal	MVSep Mega 53 StemsTelkNet	11.6590	Open-source SDR11.6590Delta SDR0.0000
Wind	MVSep Mega 53 StemsTelkNet	8.6632	Open-source SDR8.6317Delta SDR0.0315
Wind chimes	MVSep Mega 53 StemsTelkNet	6.7529	Open-source SDR4.6440Delta SDR2.1089
Woodwind	MVSep Mega 53 StemsTelkNet	3.3224	Open-source SDR3.3123Delta SDR0.0101

Paper and benchmark notes

Music Source Separation with Band-Split RoPE Transformer

SDX23 music separation track - 2023

arXiv:2309.02612

Official / repositories

MVSep Mega 53 page HF open weights ZFTurbo release v1.0.21 ZFTurbo MSST repository

Papers / technical notes

BS-RoFormer

Parameter guide

Model: MVSep Mega 53 Stems, open-weight checkpoint mvsep_mega_model_bs_roformer_53_stems_v1.ckpt with the official YAML config.
Outputs: detected non-empty FLAC stems from the 53-stem instrument set for detailed offline editing, sampling, or arrangement analysis.
Quality note: this is an advanced exploration tool. For overall separation quality and predictable stems, use six-stem separation.
Runtime note: the release recommends at least 16 GB VRAM. Dual-GPU segment routing can accelerate long audio, but each GPU still needs enough VRAM to load the model.

Use cases

Inspect whether a mix contains specific instruments such as saxophone, brass, mandolin, synth, or percussion
Export many candidate FLAC stems for offline editing, sampling, or arrangement analysis
Use as an experimental discovery tool when six fixed stems are not detailed enough

Workflow

1Upload a song and keep the fixed official Mega 53 settings
2The worker runs the open-weight BS-RoFormer checkpoint with dual-GPU segment routing when two CUDA cards are visible
3Review detected stems from the task output list before offline editing

Pre-flight checklist

For best overall separation quality, use six-stem separation first
Mega 53 is experimental and each individual stem may be weaker than a dedicated specialist model
The model is memory intensive; each participating GPU still needs enough VRAM to load the checkpoint

Login required to use this tool

Browsing tool info is always free - no account needed.