FL ClearVoice Model Loader

FL ClearVoice Model Loader looks like the boring half of the filliptm/ComfyUI_FL-ClearVoice pack: one dropdown, one toggle, one output. But that dropdown is the whole point - it picks which audio model you're about to run, and the pack wraps four different backends behind a single clean list. The loader downloads the model on first use, caches it in memory, and hands the Process node a CLEARVOICE_MODEL handle to work with. Simple, but it's where every meaningful choice in this pack lives.

If you recognize the author from ComfyUI_Fill-Nodes or the FL TTS packs, the pattern is familiar: tiny node surface, big behavior behind it. This is the node you set once and mostly leave alone.

The model dropdown - the actual content

The one required input, model, defaults to MossFormer2_SE_48K and offers eight choices. Here's what they actually mean, because the names do not explain themselves:

• MossFormer2_SE_48K (default) - best-quality speech enhancement, 48 kHz output. Start here.
• FRCRN_SE_16K - faster 16 kHz enhancement. Good for quick previews.
• MossFormerGAN_SE_16K - GAN-based enhancement, 16 kHz output.
• MossFormer2_SR_48K - super-resolution: upscales lower-rate audio to 48 kHz.
• NovaSR - 16 kHz to 48 kHz super-resolution, but it runs in fp16 on CUDA. On CPU or Mac this model will not cooperate - skip it unless you're on an NVIDIA GPU.
• Resemble_Enhance - full restoration from Resemble AI: denoise + enhance + vocoder, 44.1 kHz out. Needs pip install resemble-enhance since it's optional in the requirements.
• Resemble_Denoise - denoise only, 44.1 kHz. The fast option when noise is your only problem.
• VoiceFixer - all-in-one restoration (noise, reverb, clipping, bandwidth) from the VoiceFixer project. Also an optional dependency: pip install voicefixer.

That's three ClearVoice models, two Resemble ones, VoiceFixer, and NovaSR - genuinely different tools behind one menu. The loader doesn't care which you pick; it loads the right backend, and the Process node figures out the rest from the handle.

The one optional input

force_reload (boolean, default false). When false, the loader reuses a model that's already cached in memory - instant. Flip it to true and it reloads from disk, which is your debugging lever when a model misbehaves or you want to force a clean load after a partial download. You'll rarely touch it.

The output is a single model wire (CLEARVOICE_MODEL) that plugs directly into FL ClearVoice Process's model input. There's nothing else to do with it - it's a private type, so ComfyUI won't let you jam it anywhere else.

How the downloading works

Models are centralized in ComfyUI/models/clear_voice/, one subfolder per backend. First time you select a model, it downloads from the source (ClearVoice models come from Alibaba's Hugging Face repos, VoiceFixer's checkpoint from Zenodo, NovaSR from its own HF repo) - and the pack monkey-patches ClearVoice's own downloader so you get real per-file progress bars with speed and ETA instead of the stock "Fetching X files" wall. Long enough to grab a coffee on first run; instant after that.

Install and gotchas

Install is the same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_FL-ClearVoice.git
cd ComfyUI_FL-ClearVoice
pip install -r requirements.txt

Or search "FL ClearVoice" in ComfyUI Manager. Two things to know: requirements.txt pins transformers<4.46.0, so installing can downgrade a newer transformers that other nodes rely on - the usual custom-node dependency friction. And if you pick Resemble or VoiceFixer models without installing their optional pip packages, you'll get an import error at load time, not a friendly warning. On the bright side, the default MossFormer2_SE_48K works on NVIDIA, CPU, and Mac (MPS) out of the box, so you can be up and running without touching any of that.