spoonn/SoulX-Singer-Native-ComfyUI

SoulX-Singer-ComfyUI

Native singing synthesis (SVS), singing voice conversion (SVC), pitch extraction and dereverberation.

This is a converted model package for Jack Spooner’s ComfyUI fork, derived from Soul-AILab/SoulX-Singer. It is not a new fine-tune.

Required code

The integration modifies ComfyUI core; this is not a drop-in custom_nodes package.

• Model integration commit
• Implementation · Offline converter
• Development branch · Pinned compatible reference snapshot 9d9053c5e
• Installation, node usage and acceptance details

For an existing ComfyUI installation, use the linked core changes and their prerequisites against a compatible base. The integration commit alone is not asserted to be an independently cherry-pickable patch. The pinned reference snapshot includes the required shared loaders, attention and memory changes. Follow the fork’s installation README and dependencies when reproducing that snapshot; update the application only while its queue is idle. Upstream ComfyUI is Comfy-Org/ComfyUI; its README describes the underlying node engine and model management.

Files and installation

Copy the model category folders into ComfyUI/models/, preserving their names and subfolders. Supporting pronunciation/configuration assets in this repository are part of the package.

File	Size
audio_encoders/soulx_singer_dereverb.safetensors	0.913 GB
audio_encoders/soulx_singer_rmvpe.safetensors	0.181 GB
audio_encoders/soulx_singer_whisper_base.safetensors	0.082 GB
diffusion_models/soulx_singer_svc.safetensors	1.774 GB
diffusion_models/soulx_singer_svs.safetensors	1.798 GB
text_encoders/soulx_singer/pronunciation.safetensors	0.003 GB
vae/soulx_singer_svc_vocos.safetensors	1.020 GB
vae/soulx_singer_svs_vocos.safetensors	1.020 GB

Manifest/provenance files record the conversion and source revisions. SHA256SUMS and FILES.json describe this release. Filenames and model folders matter. No original source snapshot needs to be converted again.

Usage

Copy diffusion_models, vae, audio_encoders and text_encoders into ComfyUI/models with their subfolders intact. Install the documented SoulX prerequisites (NumPy, SciPy, librosa, soundfile; pypinyin optional for automatic Mandarin pronunciation). Start with the supplied Singing Voice Conversion or Melody/Score Synthesis workflow and choose your own isolated singer/source clips. English pronunciation assets are included. The mixed-song karaoke/lyric-replacement example is withheld because its optional becruily karaoke weights have no license declared by their original publisher. Use already isolated vocals with the supplied SVS/SVC workflows. The linked local-installation docs also describe optional components that are not all included in this public release.

Example graphs are in workflows.

Verification and limits

Default-step real-weight CPU comparisons and short-clip RTX 5090 queue/lifecycle checks are documented. Song-quality acceptance is incomplete: singer identity, correct replacement lyrics, alignment and polyphonic pitch tracking are not guaranteed. SVC preserves source lyrics; SVS requires inspected target timing/metadata. Automatic alignment is an estimate. Read the upstream usage disclaimer and use voices/recordings with appropriate rights.

These statements describe the linked retained tests, not a new cross-platform acceptance suite. File size is not a minimum RAM/VRAM requirement. No inference-time downloads or external inference worker are required by these native model paths.

Conversion and provenance

Splits SVS/SVC and their Vocos decoders, extracts Whisper-base encoder-only content weights, folds preprocessing weight normalization and packages RMVPE/karaoke/dereverb. Converts original G2P assets into a data-only pronunciation package; retains CMU dictionary/license. Checkpoint tensors retain FP32 and documented exact round trips.

Source model revision: 40493ad90286056c7a9095035164434a79daa8c9. See PROVENANCE.json, the conversion manifest and NOTICE. No training data, private inputs or generated user recordings are included.

License and attribution

Main SoulX SVS/SVC/vocoder weights: Apache-2.0. The optional anvuew dereverb component is GPL-3.0, with original checkpoint, YAML and full corresponding integration/converter source supplied in source/. Whisper, G2P and CMU dictionary retain their notices. The becruily karaoke component is withheld because its original publisher declares no redistribution license. No component is relicensed by the mirror’s general Apache metadata. See corresponding source and exclusions.

See LICENSE, NOTICE, and the retained upstream model card/source notices. Model authors and publishers retain credit; this is a third-party ComfyUI repackaging and does not imply endorsement. The weights’ terms and the separately linked ComfyUI code license are distinct.

The example workflow files are copied/adapted from the linked ComfyUI fork; its GPL-3.0 license is retained in LICENSES/COMFYUI-GPL-3.0.txt. This does not relicense the model weights.

Source and withheld component

source/ provides the original GPL dereverb checkpoint, YAML and full matching converter/inference source archive. Those are optional reproduction sources, not additional inference-model dependencies. Original-author license declarations take precedence over a mirror’s general metadata. The becruily karaoke weights are absent from the current release because their original publisher provides no declared license.