Supertonic Model Loader 🎤

Most ComfyUI nodes demand a lot of you: a checkpoint, a config, a VAE, a prayer. Supertonic Model Loader has literally no inputs. You drag it in, run the workflow, and it does the one job that would otherwise be a whole tutorial - getting Supertonic-3's model onto your machine and handing a ready-to-synthesize engine to the rest of the graph.

What it is

This is the Supertonic pack's setup node, and its entire interface is a single output: model, of type SUPERTONIC_MODEL. There are no knobs, because there's only one model it can load. That output wires into exactly one place: the Supertonic Text-to-Speech node's model input. Loader → TTS → audio out is the whole pipeline, and this node is step one.

What happens when you run it

On the first run it discovers the ~400MB ONNX model isn't there and downloads it. Where it goes matters: into <your node dir>/models/supertonic-3/, inside the custom node folder rather than ~/.cache where most audio packs quietly stash weights. The author also built a genuinely nice touch into the download - instead of a frozen console you get an animated 0–100% progress bar with byte-accurate counting, with all the HuggingFace Hub logging noise suppressed. Small thing, but it's the difference between "is this hung?" and "oh, three minutes, fine."

It's also a singleton. The first Loader run initializes the engine once per ComfyUI session, and every TTS node in your workflow shares that one instance. Run it once, forget it.

The engine underneath

That engine is Supertone's Supertonic-3 - a roughly 66M-parameter flow-matching TTS running through ONNX Runtime. The selling point is the compute floor: it's designed for CPU, needs no GPU, and community benchmarks put it comfortably faster than real time even on plain CPU cores. That's the whole reason this pack exists - voice generation that doesn't fight your image model for VRAM.

Install and first-run gotchas

Install via ComfyUI Manager (search "Supertonic"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/Anonymzx/ComfyUI-Supertonic3TTS.git
pip install -r ComfyUI-Supertonic3TTS/requirements.txt

restart, and the node appears under audio/Supertonic. Then the three things that bite people:

• Internet on first run is mandatory. The download can't be skipped. If it dies partway, the SDK stages into a temp folder and only moves it into place on success, so a clean retry works - but let the progress bar finish before you kill ComfyUI.
• The supertonic pip package is the hard dependency. If the import fails you'll see supertonic package not found. Install: pip install supertonic. The rest of requirements.txt (numpy, soundfile, librosa) is common ComfyUI fare, and torch/torchaudio come from ComfyUI itself.
• The model license is not the code license. The pack is MIT, but the Supertonic-3 weights ship under OpenRAIL-M (Supertone). That's permissive-with-conditions - fine for personal and most projects, but read it before you ship a product on top of it.

Should you even notice it?

Honestly, this is the pack's quiet node, and that's the point. You'll wire it once per workflow, run it, and never look at it again - but skip it and SupertonicTTS has nothing to synthesize with. It's the boring node that makes the interesting one work, and there's a lot to be said for that.