Add licensed SoulX native models and corresponding dereverb source

8a2cac8 verified 16 days ago

2.03 kB

# Corresponding source for the GPL dereverb component

`audio_encoders/soulx_singer_dereverb.safetensors` is a converted derivative of anvuew’s model, licensed GPL-3.0. The SoulX preprocessing mirror’s general Apache tag does not replace that original license.

Original publisher: https://huggingface.co/anvuew/dereverb_mel_band_roformer/tree/cef05ad2b5b3145ea5c149d3ad5d1f8439b34d06
Original model SHA-256: `9262877b87e9ebb0fb808a456b0a411fa677f5df31c8383c1254af531c078970`
The original editable PyTorch checkpoint and YAML are supplied in this directory, alongside the complete corresponding ComfyUI integration/converter source snapshot (`comfyui-source-9d9053c5e.tar.gz`). These source files are for inspection/reproduction; load the converted safetensors for inference. Download `source/` only if you need the corresponding source.
The same exact software source is publicly available at https://github.com/jackspooner/ComfyUI/tree/9d9053c5e7a272f1769886f4d89b6f8a7c79c785. Conversion is implemented in `scripts/convert_soulx_singer.py`, which folds checkpoint weight-normalization parameters and saves the native layout. The native dereverb architecture is in `comfy/soulx_singer/separation.py`. Original checkpoint/config entries under `preprocess/dereverb_mel_band_roformer` are the inputs to that converter. No new training is performed.
The raw `.ckpt` is the upstream preferred checkpoint form and is not loaded by the native inference path. It is a PyTorch pickle-format file; inspect/use it only through trusted tooling. License text: `../LICENSES/ANVUEW-DEREVERB-GPL-3.0.txt`.
The optional becruily karaoke checkpoint is NOT distributed. Its original repository at https://huggingface.co/becruily/mel-band-roformer-karaoke/tree/0c149975cfaa261c7d87baf54330a9da85bcf888 declares no license, so redistribution permission could not be verified. Main SVS/SVC and pitch/dereverb paths remain available with already isolated vocals. The full mixed-song karaoke workflow is consequently not included in this release.

	# Corresponding source for the GPL dereverb component

	`audio_encoders/soulx_singer_dereverb.safetensors` is a converted derivative of anvuew’s model, licensed GPL-3.0. The SoulX preprocessing mirror’s general Apache tag does not replace that original license.

	Original publisher: https://huggingface.co/anvuew/dereverb_mel_band_roformer/tree/cef05ad2b5b3145ea5c149d3ad5d1f8439b34d06
	Original model SHA-256: `9262877b87e9ebb0fb808a456b0a411fa677f5df31c8383c1254af531c078970`

	The original editable PyTorch checkpoint and YAML are supplied in this directory, alongside the complete corresponding ComfyUI integration/converter source snapshot (`comfyui-source-9d9053c5e.tar.gz`). These source files are for inspection/reproduction; load the converted safetensors for inference. Download `source/` only if you need the corresponding source.

	The same exact software source is publicly available at https://github.com/jackspooner/ComfyUI/tree/9d9053c5e7a272f1769886f4d89b6f8a7c79c785. Conversion is implemented in `scripts/convert_soulx_singer.py`, which folds checkpoint weight-normalization parameters and saves the native layout. The native dereverb architecture is in `comfy/soulx_singer/separation.py`. Original checkpoint/config entries under `preprocess/dereverb_mel_band_roformer` are the inputs to that converter. No new training is performed.

	The raw `.ckpt` is the upstream preferred checkpoint form and is not loaded by the native inference path. It is a PyTorch pickle-format file; inspect/use it only through trusted tooling. License text: `../LICENSES/ANVUEW-DEREVERB-GPL-3.0.txt`.

	The optional becruily karaoke checkpoint is NOT distributed. Its original repository at https://huggingface.co/becruily/mel-band-roformer-karaoke/tree/0c149975cfaa261c7d87baf54330a9da85bcf888 declares no license, so redistribution permission could not be verified. Main SVS/SVC and pitch/dereverb paths remain available with already isolated vocals. The full mixed-song karaoke workflow is consequently not included in this release.