Musicgazer_HiResDemux

KO | EN

gitlite — search

#css #redis #react #automation #mysql #devops #containers #cross-platform #json #python

Musicgazer_HiResDemux

(★ 9)

Losslessly separate instruments from any Hi-Res audio, without being limited by the sampling rate of the separation model.

File Explorer

ZIP Download

- - config_karaoke_frazer_becruily.yaml
  - config_mdx23c.yaml
  - config_orchestra_mdx23c.yaml
  - logic_pro_config.yaml
  - model_bs_roformer_ep_368_sdr_12.9628 96khz.yaml
- unofficial_msst_model.json
- - config_musdb18_bs_mamba2.yaml
  - config_musdb18_bs_roformer.yaml
  - config_musdb18_demucs3_mmi.yaml
  - config_musdb18_htdemucs.yaml
  - config_musdb18_mdx23c.yaml
  - config_musdb18_mel_band_roformer.yaml
  - config_musdb18_mel_band_roformer_all_stems.yaml
  - config_musdb18_scnet.yaml
  - config_musdb18_scnet_large.yaml
  - config_musdb18_segm_models.yaml
  - config_musdb18_torchseg.yaml
- - aufr33-jarredou_DrumSep_model_mdx23c_ep_141_sdr_10.8059.yaml
  - config_bs_roformer_4stems_ft.yaml
  - config_dnr_bandit_bsrnn_multi_mus64.yaml
  - config_drumsep.yaml
  - config_htdemucs_6stems.yaml
  - config_musdb18_htdemucs.yaml
  - config_musdb18_mdx23c.yaml
  - config_musdb18_scnet.yaml
  - config_scnet_large.yaml
- - config_apollo.yaml
  - config_apollo_uni.yaml
  - config_aspiration_mel_band_roformer.yaml
  - config_dereverb_echo_mbr_v2.yaml
  - config_dereverb_mdx23c.yaml
  - config_mdx23c_similarity.yaml
  - config_musdb18_htdemucs_FT.yaml
  - dereverb_mel_band_roformer_anvuew.yaml
  - deverb_bs_roformer_8_256dim_8depth.yaml
  - deverb_bs_roformer_8_384dim_10depth.yaml
  - deverb_mel_band_roformer.yaml
  - model_bs_roformer_ep_937_sdr_10.5309.yaml
  - model_mel_band_roformer_crowd.yaml
  - model_mel_band_roformer_denoise.yaml
- - big_beta5e.yaml
  - config_bsrofoL.yaml
  - config_chorus_male_female_bs_roformer.yaml
  - config_instrumental_becruily.yaml
  - config_Kim_MelBandRoformer.yaml
  - config_kimmel_unwa_ft.yaml
  - config_mel_band_roformer_karaoke.yaml
  - config_melbandroformer_inst.yaml
  - config_melbandroformer_inst_v2.yaml
  - config_melbandroformer_instvoc_duality.yaml
  - config_vocals_becruily.yaml
  - config_vocals_htdemucs.yaml
  - config_vocals_mdx23c.yaml
  - config_vocals_mel_band_roformer.yaml
  - config_vocals_segm_models.yaml
  - config_vocals_swin_upernet.yaml
  - model_bs_roformer_ep_317_sdr_12.9755.yaml
  - model_bs_roformer_ep_368_sdr_12.9628.yaml
  - model_mel_band_roformer_ep_3005_sdr_11.4360.yaml
- - 1band_sr16000_hl512.json
  - 1band_sr32000_hl512.json
  - 1band_sr33075_hl384.json
  - 1band_sr44100_hl1024.json
  - 1band_sr44100_hl256.json
  - 1band_sr44100_hl512.json
  - 1band_sr44100_hl512_cut.json
  - 1band_sr44100_hl512_nf1024.json
  - 2band_32000.json
  - 2band_44100_lofi.json
  - 2band_48000.json
  - 3band_44100.json
  - 3band_44100_mid.json
  - 3band_44100_msb2.json
  - 4band_44100.json
  - 4band_44100_mid.json
  - 4band_44100_msb.json
  - 4band_44100_msb2.json
  - 4band_44100_reverse.json
  - 4band_44100_sw.json
  - 4band_v2.json
  - 4band_v2_sn.json
  - 4band_v3.json
  - 4band_v3_sn.json
  - 4band_v4_ms_fullband.json
- augmentations_template.yaml
- config_dnr_bandit_v2_mus64.yaml
- config_some.yaml
- config_vocals_bandit_bsrnn_multi_mus64.yaml
- config_vocals_bs_mamba2.yaml
- config_vocals_bs_roformer.yaml
- config_vocals_scnet.yaml
- config_vocals_scnet_large.yaml
- config_vocals_scnet_unofficial.yaml
- config_vocals_torchseg.yaml
- language.json
- models_info.json
- msst_model_map.json
- vr_model_map.json
- webui_config.json
- augmentations.md
- bs_roformer_info.md
- dataset_types.md
- enhance_demux.md
- inference.md
- logo.png
- mel_roformer_experiments.md
- pretrained_models.md
- README_ja.md
- README_zh-TW.md
- README_zh.md
- training.md
- webui.md
- __init__.py
- filter_utils.py
- hires_separator.py
- pipeline.py
- __init__.py
- comfy_infer.py
- msst_infer.py
- - model_from_config.py
- - bandit.py
  - bandsplit.py
  - film.py
  - maskestim.py
  - tfmodel.py
  - utils.py
- - __init__.py
  - attend.py
  - attend_sage.py
  - bs_roformer.py
  - mel_band_roformer.py
- - __init__.py
  - apollo.py
  - base_model.py
- - __init__.py
  - scnet.py
  - separation.py
- - __init__.py
  - scnet.py
  - utils.py
- - common_separator.py
  - separator.py
  - vr_separator.py
- __init__.py
- demucs4ht.py
- ex_bi_mamba2.py
- mdx23c_tfc_tdf_v3.py
- segm_models.py
- torchseg_models.py
- ts_bs_mamba2.py
- upernet_swin_transformers.py
- 1.png
- 2.png
- - __init__.py
  - i18n.py
  - locale_diff.py
- - __init__.py
  - infer.py
- - bule_miku.json
  - theme_blue.json
  - theme_pink.json
- - download_models.py
  - librosa-0.9.2-py3-none-any.whl
- __init__.py
- train.py
- train_accelerate.py
- valid.py
- __init__.py
- constant.py
- dataset.py
- ensemble.py
- logger.py
- utils.py
- - __init__.py
- __init__.py
- ensemble.py
- init.py
- models.py
- msst.py
- preset.py
- settings.py
- setup.py
- tools.py
- train.py
- utils.py
- vr.py
.gitignore
go-webui.bat
LICENSE
README.md
webUI.py
webUI_for_colab.ipynb

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">Musicgazer</h1>
<div align="center">

English | [简体中文](docs/README_zh.md) | [繁體中文](docs/README_zh-TW.md) | [日本語](docs/README_ja.md)<br>

</div>

## Introduction

Musicgazer is an enhanced fork of [MSST-WebUI](https://github.com/SUC-DriverOld/MSST-WebUI) that adds hi-res audio source separation — output always at the original sample rate regardless of the model's training rate (except when the input rate is lower). MSST-WebUI is a WebUI for [Music-Source-Separation-Training (MSST)](https://github.com/ZFTurbo/Music-Source-Separation-Training), supporting model training and inference, customizable preset processing pipelines, model installation, advanced ensemble mode, and [SOME](https://github.com/openvpi/SOME/) integration.

> [!IMPORTANT]
> [MSST-WebUI](https://github.com/SUC-DriverOld/MSST-WebUI) is ceasing maintenance. Check the original repository for the latest status and migration options (e.g., [pymss](https://github.com/pymss-project/pymss)), or port to UVR5 or other hosts. This project only maintains the hi-res audio separation feature.

## Key Features

- **Lossless Varispeed**: For the common case of separating 48 kHz audio with a 44.1 kHz model, this strategy linearly compresses all frequencies into the model bandwidth, runs inference at a slightly lower pitch, then restores the outputs to the original speed and pitch. Zero quality loss.
- **Band-split**: Splits audio at ~20 kHz. The low band goes through the model normally; the high band is separated using masks derived from each stem's behavior around 20 kHz. The HF distribution is essentially inferred from the low-band separation. This differs from a model genuinely trained on hi-res audio, but all spectral content comes from the original recording — no generative restoration.
- **Model-agnostic** — works with any MSST architecture (BS-RoFormer, Mel-Band RoFormer, MDX23C, HTDemucs, SCNet, Bandit, Apollo, Mamba2, etc.). The strategies only depend on the `(audio) → {stem: audio}` interface.
- **Auto strategy selection** — picks varispeed when `orig_sr / model_sr ≤ 1.15`, band-split otherwise. Manual override available.

See [docs/enhance_demux.md](docs/enhance_demux.md) for the full technical breakdown.

## Quick Start

1. Install the WebUI following [Run from source](#run-from-source) below.
2. Download models via the "Install Models" page, or place `.ckpt` files in `pretrain/` manually.
3. Upload a 48/96/192 kHz audio file. The console will show:
   ```
   Hi-res separation enabled | orig_sr: 96000 Hz → model_sr: 44100 Hz | strategy: band_split
   Separation complete: your_audio.flac → output_sr: 96000 Hz
   ```
4. Verify that the output files retain the original sample rate.

## Usage

- **Windows**: Download the installer from [Releases](https://github.com/SUC-DriverOld/MSST-WebUI/releases) or clone this repository and run from source.
- **Linux/macOS**: Clone this repository and run from source.
- **Google Colab**: [Click here](https://colab.research.google.com/github/SUC-DriverOld/MSST-WebUI/blob/main/webUI_for_colab.ipynb) to run the WebUI on Google Colab.

### Download links

For installers, models, and mirror download options, see the upstream [MSST-WebUI](https://github.com/SUC-DriverOld/MSST-WebUI) repository.

### Documents

Chinese documentation is available [here](https://r1kc63iz15l.feishu.cn/wiki/JSp3wk7zuinvIXkIqSUcCXY1nKc). Additional docs can be found in the `docs/` folder. See also [deton24](https://github.com/deton24)'s [Instrumental and vocal & stems separation & mastering guide](https://docs.google.com/document/d/17fjNvJzj8ZGSer7c7OFe_CNfUKbAxEh_OBv94ZdRG5c).

## Run from source

- Clone this repository.

```bash
  git clone https://github.com/LDDROG/Musicgazer_HiResDemux.git
  cd Musicgazer_HiResDemux
  ```

- Create a Python environment and install dependencies. Python 3.10 is recommended.

```bash
  conda create -n msst python=3.10 -y
  conda activate msst
  pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
  pip install -r requirements.txt --only-binary=samplerate
  ```

- Fix a librosa compatibility issue: in `site-packages/librosa/util/utils.py` line 2185, change `np.dtype(complex): np.dtype(np.float).type,` to `np.dtype(complex): np.dtype(float).type,`. Alternatively:

```bash
  pip uninstall librosa -y
  pip install tools/webUI_for_clouds/librosa-0.9.2-py3-none-any.whl
  ```

- Start the WebUI:

```bash
  python webUI.py
  ```

- For advanced launch arguments or cloud deployment, see [docs/webui.md](docs/webui.md).

- If you encounter `ValueError: Make sure that the channel dimension of the pixel values match with the one set in the configuration` when using `swin_upernet`, refer to [this issue](https://github.com/SUC-DriverOld/MSST-WebUI/issues/24).

## CLI & Python API

See [docs/inference.md](docs/inference.md) for details.

### Integration example

The `enhance_demux` module is designed to drop into any MSST-compatible pipeline:

```python
import librosa
from enhance_demux import separate_hires

mix, orig_sr = librosa.load("audio_96k.wav", sr=None, mono=False)

waveforms = separate_hires(
    mix, model, config, device, model_type,
    orig_sr=orig_sr,    # e.g. 96000
    target_sr=44100,
)
# waveforms["vocals"] → (channels, time) at orig_sr
```

### Porting to UVR5

UVR5 follows the same `load → demix → save` pattern. To add native sample rate support:

1. Copy `enhance_demux/` into UVR5's `gui/` directory.
2. Load audio at its native rate:
   ```python
   mix, sr = librosa.load(path, sr=None, mono=False)
   ```
3. Add the hires branch after the model call:
   ```python
   if sr > 44100:
       from enhance_demux import separate_hires
       results = separate_hires(mix, model, config, device, model_type,
                                orig_sr=sr)
   else:
       results = demix(config, model, mix, device, model_type=model_type)
   ```
4. Write output at `sr` instead of hardcoding `44100`.

The strategies are model-agnostic — compatible with MDX-Net, VR Arch, Demucs, and any architecture exposing a `(audio) → {stem: audio}` interface.

## Training & Validation

See [docs/training.md](docs/training.md).

## Reference

### Upstream projects

- [MSST-WebUI](https://github.com/SUC-DriverOld/MSST-WebUI) — the WebUI this project is forked from
- [Music-Source-Separation-Training](https://github.com/ZFTurbo/Music-Source-Separation-Training) — MSST training framework by ZFTurbo
- [Ultimate Vocal Remover](https://github.com/Anjok07/ultimatevocalremovergui) — VR model inference

### Tools & libraries

- [librosa](https://github.com/librosa/librosa) — audio analysis (STFT, resampling, filters)
- [scipy](https://github.com/scipy/scipy) — signal processing (zero-phase Butterworth filters)
- [PyQt-Fluent-Widgets](https://github.com/zhiyiYo/PyQt-Fluent-Widgets) — UI components
- [PySide6](https://doc.qt.io/qtforpython-6) — Qt for Python
- [Singing-Oriented MIDI Extractor (SOME)](https://github.com/openvpi/SOME/) — MIDI extraction tool

### Community

- [deton24's Instrumental and vocal & stems separation & mastering guide](https://docs.google.com/document/d/17fjNvJzj8ZGSer7c7OFe_CNfUKbAxEh_OBv94ZdRG5c)
- [ComfyMSS](https://github.com/KitsuneX07/ComfyMSS) — ComfyUI nodes for music source separation
- [python-audio-separator](https://github.com/nomadkaraoke/python-audio-separator) — Python audio separation library

Musicgazer_HiResDemux

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README