Higgs v3 Load Model

Every other node in the Higgs v3 TTS pack is dead on arrival until this one runs. Higgs v3 Load Model is the thing that drags the 4B-parameter Higgs Audio v3 checkpoint into ComfyUI, wires up its audio codec, and hands you a higgs_model that Generate, Voice Clone, and Multi-Speaker all take as their required input. It's the heavy one - expect a 9.31 GB download on first run and roughly 11 GB of VRAM while it's live. You only build it once per workflow, then feed its output into everything downstream.

The name doesn't lie about much: this is genuinely local, in-process inference, not a call to some Boson API. The loader builds the Qwen3 text backbone plus Higgs's audio-token embedding/head straight from the model.safetensors, and it normalizes the bundled Higgs Audio V2 tokenizer config so it works across Transformers 5.3.0–5.5.0 (5.5.0 recommended). Weights load CPU-first and get registered with ComfyUI's model management, so if you run AIMDO DynamicVRAM the castable weights get paged instead of parked permanently in VRAM. There's no keep-loaded toggle and no unload node - switching model, dtype, device, or attention hard-unloads the previous bundle first. That's deliberate, so don't go hunting for a knob that isn't there.

The inputs that matter

The model dropdown is managed: it lists whatever lives in ComfyUI/models/higgsv3tts/. On a fresh install that's just the auto-download entry, which pulls the whole checkpoint for you if download_if_missing is on.

• dtype - auto (bf16 on CUDA, fp32 elsewhere) or bf16. Note there's no fp16 option; the author hides it because it can produce non-finite audio. Don't force it.
• device - auto follows ComfyUI's current torch device, cuda is fastest, cpu is a fallback that the tooltip politely calls "very slow."
• attention - auto/sdpa are the fast defaults. flash_attention needs flash_attn installed; sageattention patches in SageAttention, which the README warns can be slower on this token-by-token path. Benchmark before you bother.

Output is a single higgs_model (HIGGSV3TTS_MODEL) wire that feeds every generation node in the pack.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/Higgs_v3-TTS-ComfyUI.git
cd Higgs_v3-TTS-ComfyUI
python install.py

Restart ComfyUI after. Or skip all that and search "Higgs v3 TTS" in ComfyUI Manager - same result. The pack's install.py and requirements.txt are refreshingly polite: they won't touch your torch, torchaudio, or transformers, which matters because the pack needs Transformers 5.3.0–5.5.0 and an over-eager pip install can wreck the rest of your environment.

Where people get burned

First, VRAM. 11 GB+ is the real number with bf16 on CUDA; AIMDO helps live pressure but isn't a magic shrink. Second, the download: the model lands at ComfyUI/models/higgsv3tts/higgs-audio-v3-tts-4b/model.safetensors, and if your logs mention hf-mirror.com or metadata/HEAD failures, update the pack and retry - downloads are forced through huggingface.co. Third, attention options: pick sageattention or flash_attention without the package installed and you get a clear error, not silent weirdness.

One caveat worth knowing before you commit: Boson releases Higgs for research and non-commercial use, and the license language explicitly says no voice cloning without consent. Fine for tinkering; read it before you build a product on it.