🎭 Load RVC Character Model

If the RVC Engine is the machine, this is where you insert the voice. 🎭 Load RVC Character Model loads a trained .pth RVC model - one file that encodes a single target speaker - and hands it to the 🔄 Voice Changer as the identity to convert into. One model, one voice: an RVC .pth learns exactly one speaker, so you'll have a different file for every character you want to convert to. This node also handles the matching FAISS .index file, which is the not-so-secret ingredient for making conversions actually sound like the target.

How it works

An RVC .pth holds the trained weights that map generic voice content onto a specific speaker's timbre. The companion .index (a FAISS index) stores the target speaker's feature vectors, and during conversion it's used to retrieve the closest-matching target features for each moment of source audio. That retrieval step is why RVC sounds like a real person and not a smeared approximation - so pairing the right index with the model matters. This node lets you load both and validates/caches them so repeated runs are fast.

The inputs and outputs that matter

• model - the .pth to load, from a dropdown of what's in your RVC models folder (the samples show names like Claire.pth, Monika.pth, Sayano.pth). This is the voice you're converting to.
• index_mode (auto, none, custom, default auto) - how to find the FAISS index. auto matches an index to your model by name automatically and is what you want most of the time. none skips the index (faster, lower similarity). custom lets you pick one by hand.
• index_file - only relevant when index_mode is custom; pick the specific .index (e.g. Monika_v2_40k.index).
• auto_download (default true) - pulls missing models from official sources automatically, so a referenced model you don't have locally can still resolve.
• training_artifacts (optional) - accepts a TRAINING_ARTIFACTS bundle, so a model you just trained in the suite can flow straight in without saving and re-picking it.

Two outputs: rvc_model (the loaded model - connect it to 🔄 Voice Changer's narrator_target) and model_info (a text summary of what loaded, including whether an index attached).

Installing it

Comes with TTS Audio Suite. Easiest: ComfyUI Manager → search TTS Audio Suite → install → restart. Manager runs the pack's install.py, which sorts the RVC dependencies plus the usual conflicts (NumPy, librosa, s3tokenizer) and Python 3.13. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/diodiogod/TTS-Audio-Suite.git, then python install.py from inside the folder with your ComfyUI venv active - run the script, not just a requirements install. Linux: portaudio19-dev libsamplerate0-dev first. Drop your character .pth files in ComfyUI/models/TTS/RVC/ and their .index files where the model layout doc specifies; restart so they show in the dropdown.

Common issues

• Model isn't in the dropdown. It's not in the RVC models folder, or you didn't restart/refresh after adding it. Check ComfyUI/models/TTS/RVC/.
• Conversion sounds generic / low similarity. Usually a missing index. Keep index_mode on auto and make sure the .index sits alongside the .pth; without it, RVC loses the retrieval step that carries most of the resemblance. index_mode: none will do this on purpose.
• Index picks the wrong file. If auto-matching grabs the wrong index (mismatched names), switch to custom and select it yourself.
• Output isn't wired. rvc_model goes to the Voice Changer's narrator_target, and the actual conversion runs there - this node only loads, it doesn't convert.
• Quality is capped by the model. A weakly-trained .pth (little data) converts poorly no matter what you do here. That's a training problem, not a loading one.