RunningHub RVC One-Click Train

In the RVC WebUI, training a voice model is a sequence of staged scripts: preprocess, extract F0, extract features, train, build an index, export. Each one run by hand, each with its own knobs. The RunningHub RVC One-Click Train node collapses all of that into a single node you drop into a ComfyUI graph. You point it at a folder of audio, pick a save_name, and when it's done you have a .pth, a .index, and a ready-to-share .zip waiting in ComfyUI/output/RVC/<save_name>/.

How it works

The node materializes your dataset - either from trainset_dir (a folder of wav/mp3/flac, as an absolute path or relative to ComfyUI's input directory) or from the optional audio input, which accepts AUDIO, lists of AUDIO, file paths, or path lists. Then the pipeline runs: audio cleanup, F0/feature extraction with the bundled RVC source, generator/discriminator training, an optional FAISS index, and packaging.

Two cleanup behaviors are worth knowing before they surprise you:

• audio_auto_clean (on by default) - converts to mono, trims silence, skips clips that are too short or too quiet, and peak-normalizes. Keep it on.
• auto_extract_vocals (also on by default) - runs torchaudio's HDEMUCS to pull the vocal out of backing music. For a dry, already-isolated vocal dataset the README says turn it off: you get separation artifacts and slower training for nothing. This is the one default I'd flip before a clean-vocal training run.

The inputs that actually matter

• trainset_dir - your data. More clean, single-speaker audio beats more audio. Multi-speaker, reverb-heavy, or noisy sources should be cleaned upstream; this node can't fix a messy dataset.
• version - v2 (768-dim features, default) vs v1 (256-dim). Unless a model you're matching was trained v1, stay on v2.
• sample_rate - 40k is the compatibility sweet spot; 48k keeps more highs at a compute cost; 32k is v2-only.
• use_f0 - on for singing and anything where you want to keep the melody; you can turn it off for pure speech.
• f0_method - rmvpe is the default and the quality pick, but it needs models/RVC/_assets/rmvpe/rmvpe.pt; harvest is the reliable fallback.
• total_epoch / batch_size - 50 and 4 by default. Small datasets like 20–100 epochs; drop batch size if you OOM, and if batch 1 still OOMs you need a bigger card.
• train_index - on by default. This builds the FAISS index that the conversion node's index_rate retrieval relies on. Turn it off only if you're sure you don't want retrieval.

Also in the optional slot: pretrained_G/pretrained_D paths - leave them empty and the node auto-detects the pretrained weights under _assets/pretrained(_v2), falling back to training from scratch if they're missing. cache_dataset_in_gpu can speed up sub-10-minute datasets but eats VRAM on anything bigger.

Output

One info string: the log summary and the generated file paths. It's an output node, so it's the end of the line - the .zip it writes is also registered as a ComfyUI output file so RunningHub's post-processing can grab it, and it feeds straight back into the ZIP Model Loader. That loop - train here, load the zip, convert - is the whole workflow in two nodes.

Honest costs

Training is real work: runtime and VRAM scale with dataset length, batch_size, total_epoch, and your GPU. A full run on CPU-only will try your patience. HuBERT at models/RVC/_assets/hubert/hubert_base.pt is mandatory even to start, and if you picked rmvpe the RMVPE file has to be in place too.

Install

ComfyUI Manager: search ComfyUI_RH_RVC. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI_RH_RVC
cd ComfyUI_RH_RVC
pip install -r requirements.txt

Then restart ComfyUI. Note the requirements include tensorboard and faiss-cpu specifically for training - the conversion-only path doesn't need them, so if your environment is tight, this node is why they're there.