RVC

ComfyUI-RVC

dependency-free Retrieval-based Voice Conversion (RVC) for ComfyUI. Inference only — no training, no index building.

Models

Download the base assets from Aero-Ex/ComfyUI-RVC on Hugging Face and place them in ComfyUI/models/rvc/. The three base assets are Comfy-format safetensors; voices are used directly from upstream .pth.

| File | What it is | | --- | --- | | hubert_base.safetensors | ContentVec-500 content encoder | | rmvpe.safetensors | RMVPE pitch estimator | | fcpe.safetensors | FCPE pitch estimator (fast) | | <voice>.pth | An RVC voice model |

hubert_base.safetensors, rmvpe.safetensors and fcpe.safetensors are converted from the upstream .pt files with the pack's offline converter. Voice .pth files are RVC checkpoints with the keys weight, config, info, sr, f0, version — put them in the same folder, no conversion needed.

The RVC voice file is the one thing you must supply yourself. Any RVC v1 or v2 model from a voice-model sharing site works. Do not use a raw training checkpoint (G_*.pth / D_*.pth with only model) — the loader requires the RVC wrapper format above and will reject it.

Workflow

Load RVC Hubert ─────────┐
Load RVC Pitch (RMVPE) ──┼─→ RVC Convert ─→ Save Audio
Load RVC Voice ──────────┘        ↑
                              Load Audio

1. Load RVC Hubert — pick hubert_base.safetensors.
2. Load RVC Pitch (RMVPE/FCPE) — pick rmvpe.safetensors or fcpe.safetensors.
3. Load RVC Voice — pick your voice .pth.
4. Load Audio — any source audio. Stereo and arbitrary sample rates are handled for you (resampled to 16 kHz mono internally).
5. RVC Convert — set f0_method to match the pitch model you loaded.
6. Feed the output into Preview Audio or Save Audio.

Node inputs

Loaders

Load RVC Hubert, Load RVC Pitch, each take a filename from models/rvc/.

RVC Convert

| Input | Default | Notes | | --- | --- | --- | | f0_method | rmvpe | Must match the loaded pitch model, except hybrid (below). | | f0_up_key | 0 | Pitch shift in semitones, −12…12. | | speaker_id | 0 | Speaker slot for multi-speaker voices. Range auto-clamps to the voice. | | index_rate | 0.75 | Retrieval blend strength. Only used when an index is connected. | | protect | 0.33 | Protects unvoiced/consonant sounds from the f0 path. 0.33 is the upstream default; 0.5 disables the blend entirely. | | rms_mix_rate | 1.0 | Volume-envelope matching against the input. 1.0 = off. | | resample_sr | 0 | Output sample rate. 0 keeps the voice's own rate. |

pitch_model_b is only needed for f0_method="hybrid", which averages the primary and secondary pitch estimators. Leave it unconnected otherwise — connecting only pitch_model_a with hybrid selected raises a clear error.

Notes

• Output is stochastic. RVC samples from a posterior at inference, so two runs on identical input differ by roughly 0.4–0.7 in waveform amplitude. This matches upstream, which seeds once at GUI startup rather than per conversion. It is not a bug and not fixable without changing the model's math.
• VRAM is dynamic. All three base models plus a voice run comfortably on a 6 GB card at fp16.
• Sample rate of the output is the voice model's native rate unless resample_sr is set. Upstream uses librosa.resample; this pack uses comfy.audio.resample so it matches the rest of ComfyUI's audio handling.

ComfyUI-RVC