F5-TTS Load model

This node won't make a sound. It's the loader half of the two-node F5-TTS pipeline in niknah/ComfyUI-F5-TTS: it downloads the F5-TTS checkpoint, builds the vocoder, and hands both to F5-TTS Audio from model, which is the node that actually synthesizes speech. If voice cloning is the show, this is the roadie that brings the amp.

The pack added the Load model / Audio from model split in v1.0.27. Before that, everything happened inside a single node. The point of splitting it, per the author, is to let ComfyUI manage the model's memory like it does your checkpoints - load it once, reuse it across generations, and ComfyUI can unload it when the queue is done. Fair warning: these checkpoints aren't big enough that this saves you from VRAM hell. It's more about clean workflows and not rebuilding the model every time you change a word. "Doesn't take up much memory anyways," as the changelog puts it.

How it works

On first load, the node pulls the checkpoint and vocab file from HuggingFace (cached to disk, so it's a one-time download of roughly a gigabyte per model), builds the flow-matching TTS model, and wraps the vocoder. It returns two outputs, and you must wire both into the audio node:

• MODEL - the TTS model itself.
• F5TTSMODELINFO - a sidecar with the vocoder, mel-spec type, and model type. The audio node pulls what it needs out of this.

The inputs that matter

• model - the checkpoint. F5v1 is the default and what you want for English; it's SWivid's updated checkpoint, and the pack switched to it as default in v1.0.23. Then there's plain F5 (the original), a batch of community-trained language models (F5-DE, F5-ES, F5-FR, F5-HI, F5-IT, F5-JP, F5-TH), and E2 - that's E2-TTS, the earlier, lighter sibling architecture.
• vocoder - auto (resolves to vocos), vocos, or bigvgan. Most models expect vocos; the tooltip says it plainly: "Most models are usally vocos." BigVGAN is the higher-fidelity option but it needs a one-line manual patch to the bundled code to work under ComfyUI.
• model_type - the architecture the checkpoint uses: F5TTS_Base, F5TTS_Small, F5TTS_v1_Base, or the E2 variants. You mostly don't touch this - it's there so the node can read custom checkpoints you drop in.

That last bit is the useful trick: the pack auto-discovers custom models. Drop a .safetensors file and a same-named .txt vocab into models/checkpoints/F5-TTS, press refresh, and it appears in the model dropdown. That's exactly how the community language models work - they're all HuggingFace downloads keyed to the same naming convention.

Install

ComfyUI Manager is the smooth path - search "ComfyUI F5-TTS" and install. By hand it's:

cd ComfyUI/custom_nodes
git clone https://github.com/niknah/ComfyUI-F5-TTS
cd ComfyUI-F5-TTS
git submodule update --init --recursive
pip install -r requirements.txt

Restart, and note that the pack bundles the entire F5-TTS repo as a submodule - that's where the heavy dependencies come from. Windows users also need the ffmpeg shared library (winget "ffmpeg (Shared)").

Gotchas

The classic failure is "No module named f5_tts" - the submodule didn't check out because your git build is bad with submodules. The fix from the README: rm -rf the F5-TTS folder inside the pack and git clone https://github.com/SWivid/F5-TTS.git F5-TTS in its place. It's common enough that the author built an auto-install fallback into the pack, but it can still happen.

One last thing: this node is only useful if the pack's audio node is in your graph. Old workflows that use the single-node F5-TTS Audio don't need it. And if you're here from an old tutorial, be aware the pack has iterated a lot - it's at v1.0.27+, and the node lineup has grown an advanced variant and this model-management split along the way.