Higgs Audio V3 Model Loader

Higgs Audio v3 is Boson AI's 4-billion-parameter multilingual TTS - the heavyweight corner of local text-to-speech, up there with Chatterbox for cloning quality but with 100+ language support and inline emotion/SFX tags. The catch: it normally runs through an SGLang-Omni server, which means a second process you have to start, watch, and keep alive. This node is the one that hides all of that from you. Wire it in, give it a model path, and it boots the pipeline from inside ComfyUI - no separate sgl-omni serve to babysit, no API key, nothing.

It's the first half of a two-node pack (Higgs Audio V3 Model Loader → Higgs Audio V3 Local TTS → SaveAudio). What the loader actually does is three things: find the model, start the runtime, and hand the result to your TTS node as a reusable HIGGS_AUDIO_V3_MODEL object.

How it works

The loader first scans ComfyUI's model folders - models/higgs_audio, models/LLM, and anything mapped in extra_model_paths.yaml - for a complete model snapshot. Find one and it uses it locally, which is what you want with a 4B model; otherwise it falls back to the Hugging Face id bosonai/higgs-audio-v3-tts-4b and lets HF handle download. Then it starts the pipeline. By default that means spawning a Python worker subprocess that runs the SGLang-Omni runtime (python_worker mode), and if SGLang-Omni is installed in the same environment as ComfyUI you can flip to in_process instead. Either way the node caches the running pipeline keyed on its config, so a second loader - or a second TTS node feeding off it - reuses the same instance instead of reloading 4B of weights. That's why the first load is slow and everything after is fast.

Inputs and outputs that matter

Only two inputs, and you'll barely touch either:

• model_path (default bosonai/higgs-audio-v3-tts-4b) - the dropdown is populated from the local model folders it found. See the trap below about what a dropdown containing only the HF id means.
• device - cuda or cpu. The model wants a CUDA card; CPU works but you'll be waiting.

Outputs are model, which you plug into Higgs Audio V3 Local TTS, and model_info, a JSON string showing the resolved path, where it came from (local vs. HF), and the runtime mode - handy for the inevitable "why is this loading from the internet?" question.

Installing it

ComfyUI Manager, search "Higgs Audio V3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/endman100/ComfyUI-HiggsAudioV3TTS

Then activate your ComfyUI environment, install requirements, and restart:

pip install -r ComfyUI/custom_nodes/ComfyUI-HiggsAudioV3TTS/requirements.txt

Here's the trap the README glosses over: requirements.txt is two lines - soundfile and numpy. The actual inference engine, SGLang-Omni, is not in there. It has to already exist in your Python environment, or you set HIGGS_AUDIO_V3_PYTHON_EXECUTABLE to point at a Python that has it. If you install this pack fresh and get an import error on first run, that's why.

Model placement, the part beginners actually stub their toe on:

ComfyUI/models/higgs_audio/bosonai/higgs-audio-v3-tts-4b

It must be a complete Hugging Face snapshot (weights, config.json, the works). Drop the folder there, restart ComfyUI, and it appears in the dropdown.

Gotchas worth knowing

• Dropdown only shows bosonai/higgs-audio-v3-tts-4b? That's the loader telling you it found no local copy. It'll still run through HF's download cache, but if you wanted offline/once-only loading, the folder isn't there or isn't complete.
• CUDA toolkit / nvcc errors are the classic WSL2 problem. Set HIGGS_AUDIO_V3_ATTENTION_BACKEND=triton and HIGGS_AUDIO_V3_DISABLE_CUDA_GRAPH=true before launching ComfyUI. This is the documented default for WSL2 boxes without a local toolkit.
• The license. Higgs v2 was Apache 2.0; v3 moved to Boson's Research and Non-Commercial license. For a hobby pipeline that's fine; for anything paid, re-read the license before you build on it.

One more thing to keep your expectations honest: at 4B and fresh off a June 2026 release, this is not a zero-setup weekend toy. Expect a chunky model download and a long first startup - the default HIGGS_AUDIO_V3_STARTUP_TIMEOUT_SECONDS is 600, and that's a clue.